Why I am all in on AI
A personal note on what this technology is really for, and what it is not.
Let me say the quiet part out loud. The fear in every room I walk into is that AI is here to take the job. I want to put that fear where it belongs, because I think it is pointed in exactly the wrong direction. AI is not here to remove people. It is here to remove the parts of the work that were never worth a person's time in the first place.
I have spent my career inside operations, finance, and the field, where the day gets eaten by the same small tasks over and over. Rekeying data between systems. Chasing a status nobody updated. Rebuilding the same report every Monday. None of that is the reason anyone was hired. It is friction, and it sits between good people and the work that actually moves the company. When I bring AI into a workflow, my goal is never to subtract a person. It is to give that person their hours back and point those hours at something that compounds.
Here is the part that gets lost in the headlines. The numbers, read honestly, tell a story of churn, not collapse. The World Economic Forum projects that by 2030, around 92 million roles will be displaced and 170 million new ones created, a net gain of roughly 78 million jobs, with about 22 percent of the global workforce changing shape in the next five years.10 The International Monetary Fund estimates that about 40 percent of jobs worldwide are exposed to AI, rising to roughly 60 percent in advanced economies, and crucially, that about half of those exposed roles are positioned to be made better by AI rather than replaced by it.11 Even the often quoted Goldman Sachs figure, that AI could affect the equivalent of 300 million full-time jobs, is a statement about tasks shifting, not people vanishing.12
The risk to your job was never the machine. It is the person beside you who learned to use it first.
That is the outside the box view I want you to sit with. We keep framing this as humans versus AI. The real contest is between people who reinvest their freed time into higher value work and people who do not. A technician who lets AI write the service report and uses the saved twenty minutes to close one more call is not being replaced. He is becoming the most valuable person on the truck. An analyst who hands the data pull to a model and spends that time on the recommendation is not obsolete. She is finally doing the job she was hired to do. AI does not justify cutting a position. Used well, it is the strongest argument for keeping one, because it turns a cost center into a force multiplier.
So my ask is simple. Do not wait to be told. Pick one task you dread, the one that drains an hour you will never get back, and put a tool on it this week. Then take that hour and aim it at something only a human can do: the judgment call, the relationship, the idea nobody else had. That is how you advance your own position, and it is how you advance the company at the same time. Embracing AI is not a surrender of your value. It is the clearest way I know to prove it.
I am betting my work on that belief. I hope you will too.
Rocky Hill
AI to Enterprise Production
A department by department field guide for putting AI to work across the enterprise. Real tools, real costs, real ROI signals, run through the Edge Framework.
Most companies have adopted AI. Almost none have captured the value. Industry data shows a large majority of AI projects stall before they ever reach production, and the cause is rarely the technology. It is poor strategy, weak data, and no one tracking the return.1 This book is built to fix all three, laying out a practical, department by department path from first pilot to full production.
The promise here is simple. Every department is treated as a set of workflows to redesign, not a license to buy. For each one you get where AI moves the needle, the step by step integration sequence, a table of real tools with what they do and what they cost, the ROI signal to watch, and a "build it yourself" panel for teams that want to own the capability instead of renting it.
What this book covers
Six kinds of work separate the companies that get value from AI from the ones that stall: a strategy and roadmap tied to the business model, automation that redesigns workflows with people kept in the loop, custom tools built where they create a real edge, governance from day one, ROI measured against a locked baseline, and data the AI can actually trust. Each of those threads runs through the chapters that follow.
Start with the Edge Framework, the six phase method this book runs on. Then visit any department in the sidebar. The framework is the spine; the departments are where it gets applied. Pricing reflects publicly reported figures at the time of writing and should be confirmed with each vendor before you budget.
What AI really is
Ask most people what AI is and they will name a chatbot: ChatGPT, Claude, Copilot. That answer is not wrong, but it is a thumbnail of a far larger picture. Chat is one young branch of a field that is almost seventy years old and runs underneath things you already trust every day.
The iceberg under the chatbot
The chat assistant everyone has tried is part of generative AI, which is itself one slice of machine learning, which is one part of artificial intelligence. Below the waterline sit the workhorses that have been quietly running for years: computer vision that reads an X-ray or a defect on a production line, recommendation systems that decide what you see next, forecasting that positions inventory before demand arrives, anomaly detection that flags a fraudulent charge in milliseconds, speech recognition, optimization that routes a fleet, and the control systems inside robotics. Look ahead at the tools in this book. Augury listening to a failing motor, Viz.ai triaging a scan, Feedzai catching fraud, Samsara coaching a driver. Almost none of them are chatbots. That is the point. If your mental model of AI stops at a text box, you are missing the half that already moves money.
It is much older than the hype
The idea is not new. In 1950, Alan Turing asked whether machines could think and proposed a test for it.15 The field got its name in 1956 at a summer workshop at Dartmouth College, where John McCarthy, Marvin Minsky, Claude Shannon, and a small group set out to make machines do things that would require intelligence in a human.16 What followed was not a straight line. Decades of optimism ran into the limits of the hardware and the data, and the field twice went through long, cold stretches that researchers still call the AI winters. The technology you are watching go vertical today is the product of seventy years of people who kept working through the quiet decades.
The shift that changed everything
The thaw came in 2012. A neural network now known as AlexNet, built by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, won a major image recognition contest by a margin that stunned the field and proved that deep learning, fed enough data and enough computing power, worked. Hinton, along with Yann LeCun and Yoshua Bengio, would later share the Turing Award, computing's highest honor, for that line of work, which is why the press calls them the godfathers of modern AI.17 The second hinge moment came in 2017, when a team at Google published a design called the transformer, the architecture that made it possible to train models on enormous amounts of text efficiently.18 Almost every large language model in use today descends from that one paper. Then, in late 2022, a chat interface put all of it in front of the public at once, and the world decided AI had just been invented. It had not. It had just become visible.
Where it is heading
The near future is less about a smarter chatbot and more about AI that acts. The current frontier is agentic systems that do not just answer but take steps: look something up, draft the document, file the ticket, and hand the consequential decision back to a person. Expect AI to keep moving from a destination you visit to a layer embedded inside the tools you already use, and to keep widening beyond language into vision, sound, and the physical world. Treat confident predictions about exact dates with caution, including mine. The honest forecast is directional: the capability is compounding, the cost is falling, and the question for any business is not whether AI arrives in your industry, but whether you are using the right kind of it.
AI is not a single product and it is not just chat. It is a seventy year old field that finally has the data and the computing power to deliver, across vision, prediction, language, and action. The teams that win will not be the ones with the cleverest chatbot. They will be the ones who pick the right kind of AI for each job, which is exactly what the rest of this book is for.
What happens to the companies that sit it out
Most leaders picture the risk of AI as something that might go wrong if they adopt it. The larger risk is quieter, and it belongs to the companies that do nothing. Here is the part nobody says out loud: waiting is not neutral, and this is the first technology where standing still actually moves you backward.
Every previous technology shift was forgiving to the latecomer. The factory could be retooled over years. The website could be built after your competitors had one. The migration to the cloud could wait until the playbook was proven and the price had dropped. In each case the laggard lost some ground, then bought the same tools and mostly caught up. The gap was fixed. You could close it.
AI is the first shift where the gap widens while you decide. The leader's team gets better every week. Yours waits with you.
That is the difference almost everyone misses. The model itself does not get smarter on its own, but the company using it does. Every quote, every ticket, every transaction becomes data that sharpens their prompts, their knowledge base, and their workflows, and the people using it get a little more fluent every week. A competitor who started a year ago is not one year ahead of you. They are one year of compounding ahead of you, and the distance is growing faster than you can sprint. For the first time, the cost of deliberation is not zero. It is paid, quietly, every single day, to the companies that moved.
When you choose not to adopt, you are not standing still. You are subsidizing your competitor's learning curve. Every customer they serve, every order they fill, feeds a process that gets harder to beat, while your spreadsheets learn nothing from the exact same work. Nonadoption is a transfer of advantage, and you are the one paying it.
What the slow decline actually looks like
It does not arrive as a single bad quarter. It arrives as a hundred small losses that each look survivable, which is exactly why it is so dangerous.
- Margin gets quietly compressed. A competitor who answers faster, quotes faster, and serves at a lower cost to serve can underprice you and still make more per job. You hold price and lose deals, or match price and lose margin. Either way the spread that funded your business thins.
- Your best people leave first. The strongest performers will not spend their careers rekeying data and rebuilding the same report. They go where their time is spent on judgment, and they go to the firms that freed them. You are left with the work nobody wanted to automate and the people who did not mind doing it.
- The competitor you lose to is one you never saw. It is rarely the rival across the street. It is a smaller, AI native operator who runs on a fraction of your overhead, enters your market from the side, and takes your most price sensitive customers before they show up in any report you read.
- Your data turns to exhaust. Every company sits on years of orders, tickets, and outcomes. The AI native firm treats that as fuel and compounds it into a moat. The firm that waits lets it evaporate, unlabeled and unused, which means even when you finally start, you start from behind on the one asset that was uniquely yours.
- Knowledge walks out the door. When your veterans retire, decades of judgment leave with them. Competitors are busy capturing that kind of expertise into systems that keep working after the expert is gone. Wait too long and you lose the chance to bottle yours at all.
The math that makes it irreversible
None of these losses are dramatic on their own. That is the trap. A five percent edge per decision sounds ignorable, until you remember it applies to every decision, every day, and that the lead it produces is reinvested into a system that widens the edge again next quarter. Small advantages do not add. They compound, and compounding is unkind to whoever started late.
The corporate graveyard is not filled with companies that bet wrong on a technology. Kodak invented the digital camera. Blockbuster was offered Netflix. Borders, Sears, and a long line of others had the resources, the brand, and the time. None of them were killed by a better version of themselves. Each was killed by someone who used a tool they had dismissed as a toy, and by the time the threat looked serious, the gap had already closed behind the leader and locked them out. The lesson is not that change is dangerous. It is that the safest looking choice, waiting until it is obvious, is the one that quietly runs out of road.
You do not have to bet the company. You have to refuse to bet against it. Start one thing, measure it honestly, and get on the side of the curve that compounds in your favor. The window does not stay open because you are not ready. It closes because someone else was.
Why AI fails in 60% of businesses
The money is going in. The value is not coming out. Before you start your own rollout, it is worth understanding exactly where everyone else is getting stuck.
Why does AI fail in 60% of businesses? That number is not a guess. In a 2025 survey of 1,250 companies, BCG found that 60 percent were achieving no material value from AI at all, reporting minimal revenue and cost gains despite substantial investment. Only 5 percent were getting value at scale.19
Other studies land in the same place, or worse, depending on what they measure. IBM surveyed 2,000 CEOs and found that only 25 percent of AI initiatives had delivered the expected return, and only 16 percent had scaled across the enterprise.20 Gartner predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027.22 The most quoted figure of all, MIT's claim that 95 percent of generative AI pilots show no measurable impact on profit, is also the most disputed, because it looks at individual pilots rather than whole companies.21
| Source | What was measured | Failure rate |
|---|---|---|
| BCG, 2025 | Companies getting no material value from AI | 60% |
| IBM, 2025 | AI initiatives that missed their expected ROI | 75% |
| IBM, 2025 | AI initiatives never scaled enterprise wide | 84% |
| Gartner, 2025 | Agentic AI projects predicted to be canceled by 2027 | 40%+ |
| MIT NANDA, 2025 | Generative AI pilots with no measurable profit impact | 95% (disputed) |
Each study measures something different, so the numbers do not conflict. I lead with 60 percent because it counts whole businesses rather than single pilots, and it comes from one of the largest recent surveys.
1. Chasing the tool instead of the problem
Too many rollouts start with "we need AI" instead of "this workflow costs us too much." DXC's 2025 global study found that 65 percent of organizations cannot build a clear business case for AI, and that adoption is often driven by board pressure rather than a defined outcome.23 Without a specific problem and a number attached to it, there is nothing to measure success against.
2. Laying AI on top of the old process
The factor most tied to real financial impact is redesigning the workflow around the AI, yet McKinsey found that only about 21 percent of adopters have redesigned any workflow at all.2 Dropping a smart tool onto a broken process just makes the broken process run faster.
3. Starving the people side
BCG's rule of thumb is that AI value comes 10 percent from algorithms, 20 percent from technology and data, and 70 percent from people and process. Most companies invert that ratio, spending heavily on software while underinvesting in training, change management, and new ways of working.19
4. Data that is not ready
AI is only as good as what it can see. Half of the CEOs IBM surveyed said rapid AI investment had left them with disconnected technology, and 68 percent called an integrated, enterprise wide data architecture critical to getting value.20 Scattered, duplicated, or untrusted data quietly caps every project built on top of it.
5. No baseline, so no proof
If you do not measure the workflow before AI touches it, you cannot prove what AI changed. Projects without a locked baseline drift into vague claims of "productivity," lose executive support at budget time, and get cut even when they were working.
6. Stuck in pilot
A demo that impresses in a meeting is not a system that survives Monday morning. MIT's research traces much of the failure to a learning gap: tools that cannot retain feedback, adapt to context, or improve over time.21 That is why so few initiatives ever scale beyond the first team.
7. Governance as an afterthought
Gartner names escalating costs, unclear business value, and weak risk controls as the reasons agentic projects get canceled.22 When security, compliance, and human oversight are bolted on late, projects stall in review, or launch and then get pulled after the first incident.
AI rarely fails because the model was wrong. It fails because the business around it never changed.
What the successful few do differently
The 5 percent that BCG calls future built are not using secret technology. Nearly all of them have senior leaders deeply engaged in AI, compared with only 8 percent of laggards, and they have deployed 62 percent of their AI initiatives versus just 12 percent for the companies falling behind. They redesign workflows, invest in their people, and treat AI as a business program rather than an IT experiment.19 Deloitte's 2026 research points the same way: the skills gap is now the top barrier to integration, and only about one in five companies has mature governance for agentic AI.24
Every one of these seven failure points is a choice, not a fate. Each one maps to a phase of the Edge Framework later in this book: pick a real problem, lock the baseline, redesign the workflow, govern from day one, and scale only what proves itself. The rest of this book is built to keep you out of the 60 percent.
My take: two things that make the difference
If I had to boil all of this down, there are two things a business has to do to make AI work. First, start with one real problem that has a number attached to it. Not "use AI," but "cut invoice processing from four days to one" or "answer every customer email within an hour." Measure where you are today, put AI to work on that one workflow, and prove the gain before you spend another dollar. Second, invest in your people as seriously as you invest in the software. The tool is the easy part. The value shows up only when the people doing the work are trained, trusted to use it, and given a say in how the workflow changes around it. Get those two right, and the rest of this book becomes a repeatable playbook instead of another stalled pilot.
The beliefs holding teams back
Most of the resistance to AI is not about the technology. It is about a handful of beliefs that sound reasonable and are quietly wrong. Here are the ones worth retiring, and what to put in their place.
Every one of these myths assumes AI is something done to you. Flip it. AI is a tool you pick up, point at the work you do not want, and use to spend more of your day on the work only you can do. That mindset is the whole game.
The Edge Framework
Six structured phases that take an AI initiative from discovery to production. Run well, the first workflow goes live in weeks, and a full rollout follows over months. This is the methodology behind this book, and it is the spine of every department chapter.
Discovery
Stakeholder interviews, a full data inventory, and a documented financial baseline. No assumptions, everything written down before a single design decision.
Use Case Selection
Score every opportunity by impact and feasibility, then get executive sign-off on the top two or three before proceeding. Focus beats a long wish list.
ROI Modeling
A three year financial model with conservative, moderate, and optimistic scenarios. CFO review is required. Year one is cost heavy by design; the return shows in years two and three.
Architecture & Build
A system design review before any code is written, then two week sprints with working demos throughout so there are no surprises at delivery.
Governance & Testing
Bias audit, regulatory review, load testing, and integration testing before anything reaches production. Governance is a gate, not an afterthought.
Deploy & Monitor
A shadow mode rollout, then a pilot on ten to twenty percent of traffic, then full deployment only after the pilot metrics come back green.
How the framework maps to each department
Every department chapter follows a condensed, in the field version of these six phases, shown as five working steps:
- Discover mirrors Phase 01. Audit the current workflow and capture the baseline before you change anything.
- Select mirrors Phase 02. Pick one bottleneck with a number attached to it, not a vague "use AI here."
- Build mirrors Phases 03 and 04. Stand up the tool or assistant, with the ROI case behind it.
- Govern mirrors Phase 05. Add the fact check, bias, and approval gates and redesign the workflow around them.
- Scale mirrors Phase 06. Measure output and outcome, then expand what works and retire what does not.
The single factor most correlated with real financial impact is workflow redesign, not the model or the vendor. Rebuilding how work flows around the AI is what separates the few high performers from everyone stuck in pilots.2 The Edge Framework is engineered to force that redesign rather than skip it.
Speed comes from sequencing, not shortcuts. Because Discovery, Use Case Selection, and ROI Modeling are done before any building starts, the first build is narrow and well defined, so one workflow can be live in weeks. Scaling it across the department, with the testing, pilot, and governance gates that come after, is measured in months. Most delay in AI projects is rework caused by skipping those first three phases.
Governance, built in from day one
Governance is continuous, not a one time checkbox. You map the rules before you design, audit for bias in five phases, document everything, and monitor for drift after launch, before the AI can ever create liability.
1. Regulatory mapping, before design
Identify every regulation that applies to the AI system before any design decisions are made. Depending on the use case that can mean HIPAA, GDPR, ECOA, and the EU AI Act, among others. Hiring and lending uses carry the heaviest obligations and are treated as high risk from the start. Mapping first means you never have to rebuild a finished system to make it compliant.
2. A full five phase bias audit
Bias is checked at every stage, not once at the end. The five phases are: examine the training data for representation gaps, run a feature analysis to catch proxies for protected attributes, conduct predeployment testing against held out groups, make a deliberate fairness metric selection appropriate to the decision, and stand up a monitoring plan that keeps watching after launch. No shortcuts on any of the five.
3. Model cards and documentation
Every model ships with a complete audit trail: its purpose, the training data behind it, performance metrics, known limitations, and the results of its bias tests. This documentation is what lets a regulator, an auditor, or a new team member understand and trust the system without reverse engineering it.
4. Ongoing monitoring
After deployment, performance drift, fairness, and data drift are tracked with automated alerts, backed by quarterly governance reviews. A model that was fair and accurate at launch can quietly degrade as the world changes; continuous monitoring is what catches that before it becomes a problem.
Governance also means integrity in how AI and its data are used. Never use AI to get around a policy, a platform rule, or an internal control, and never let it generate records or identities that are not real. If a workaround only works because someone is being misled, it is a finding, not a feature. Govern the workflow honestly or do not ship it.
The standards this maps to
This approach lines up cleanly with the recognized public frameworks, which are worth knowing by name. The NIST AI Risk Management Framework organizes the work into Govern, Map, Measure, and Manage.3 ISO/IEC 42001 is the first certifiable management system standard for AI.4 The EU AI Act sets risk tiers, with employment uses classed as high risk.5 And in the United States, rules like New York City's Local Law 144 already require bias audits for automated hiring tools.6
Securing AI, end to end
AI does not just add a tool to your stack. It adds a new attack surface, a sharper set of attackers, and, used well, a stronger defender. This page is the practical playbook: how AI changes your security picture, a step by step program to lock it down, the specific risks to know, and the watchouts that catch most teams.
Three ways AI changes your security picture
A new attack surface
Every model, prompt, data feed, vector store, and agent is a new place to be attacked. Sensitive data can leak, instructions can be hijacked, and an over permissioned agent can act on a bad command at machine speed.
A sharper attacker
Attackers have AI too. Phishing is now flawless and personalized, voice and video can be cloned from seconds of audio, and malware and reconnaissance move faster. The con is more convincing than it has ever been.
A stronger defender
The same technology defends. AI triages alerts, spots anomalies in logs, summarizes incidents, and flags the phishing your filters miss, giving a small security team the reach of a much larger one.
The step by step: securing AI adoption
Run every AI initiative through these ten steps. They map onto Phase 05 of the Edge Framework, where governance and testing are a gate before anything reaches production.
Find every AI tool and data feed in use. Include the shadow AI nobody approved. You cannot secure what you cannot see, and the unsanctioned tool is usually the leak.
Write an acceptable use policy. Spell out which tools are approved and what data may and may not be entered. Make it short enough that people actually read it.
Choose enterprise grade tools. Confirm in writing that your data is not used for training, check retention and deletion, and require SSO, data residency, SOC 2, and breach notification.
Tag your sensitive data. Know where PII, PHI, financials, and trade secrets live, and decide what is allowed anywhere near a model. Add DLP or redaction on the way in.
Enforce least privilege. Give people and especially agents the minimum access needed. Scope each tool an agent can call, and isolate tenants in any shared vector store.
Harden the build. Protect the RAG and vector layer, keep secrets and credentials out of prompts, and never pass model output straight into code, a query, or a browser without validation.
Red team before launch. Try to break it with prompt injection and jailbreaks, test for data leakage, and fix what you find. Assume an attacker will try the same things.
Log and watch. Keep an audit trail of prompts and outputs, alert on anomalies and unusual data access, and set rate and cost limits so usage cannot run away.
Plan the incident. Write an AI specific response plan: how to revoke keys, disable an agent, roll back a model, and notify the right people. Decide the kill switch before you need it.
Teach the people. The strongest control is a workforce that knows the risks, recognizes a deepfake or a phishing attempt, and follows the policy without being chased.
The patterns below cause most real world AI incidents. If you see any of them, stop and fix it before you scale.
- Public chatbots holding private data. Staff pasting customer records, source code, or contracts into a consumer tool that may store or train on it.
- Vague vendor terms. A tool that trains on your data by default, or will not commit to retention, deletion, and data residency in writing.
- Agents with write access too soon. An agent given permission to send, pay, or change records before it earned trust read only.
- Secrets in prompts. API keys, passwords, or connection strings sitting inside a system prompt or a shared notebook. If the model knows it, an attacker can pull it.
- No logs. No record of what was asked or answered, so an incident cannot be investigated after the fact.
- Shadow AI. Tools adopted team by team with no review, no policy, and no owner.
- Output trusted blindly. Model output run as code, used in a database query, or shown to a customer without a human or a validation step in between.
- The system prompt as a vault. Treating hidden instructions as a security boundary. They are not; enforce real controls in code.
The top LLM security risks to know
The OWASP Top 10 for LLM Applications is the most widely used map of these risks. Here it is in plain language, with the practical guard for each.13
| Risk | What it is | How to guard against it |
|---|---|---|
| Prompt injection | Crafted input the model treats as a new instruction, typed directly or hidden inside content it reads | Separate instructions from untrusted content, constrain what tools the model can call, never auto trigger actions from output |
| Sensitive info disclosure | The model reveals data from the prompt, retrieved context, memory, or its training | Minimize what the model can see, redact and apply DLP on inputs, set retention, monitor outputs |
| Supply chain | A compromised base model, fine tune, dataset, library, plugin, or vector database | Vet sources, pin and scan dependencies, prefer trusted providers, review third-party adapters |
| Data and model poisoning | Tampered training or fine tuning data that plants a hidden backdoor passing normal tests | Control data provenance, use trusted adapters only, test against known triggers and odd behavior |
| Improper output handling | Treating model output as safe and passing it straight into code, a query, or a browser | Validate and sanitize every output, never execute it blindly, keep a human in consequential paths |
| Excessive agency | An agent with too many permissions or tools taking harmful actions on its own | Least privilege, tightly scoped tools, human approval on anything that spends money or sends a message |
| System prompt leakage | Secrets or rules hidden in the system prompt being extracted by a user | Keep credentials and access logic out of prompts, enforce controls in deterministic code instead |
| Vector and embedding weaknesses | A poisoned RAG layer or weak access control leaking data across tenants | Access control the vector store, validate ingested content, isolate tenants, monitor retrieval |
| Misinformation | Confident, wrong output with invented facts or citations that people act on | Ground answers in cited sources, keep a human in the loop on decisions that matter |
| Unbounded consumption | Uncontrolled usage driving runaway cost or denial of service | Rate limits, budgets, quotas, and usage monitoring on every endpoint |
Adapted in plain language from the OWASP Top 10 for LLM Applications (2025). For adversary techniques against AI systems, the MITRE ATLAS knowledge base is the companion reference.14
When the attacker has AI too
The fastest moving threat is social engineering. Generative tools write flawless, personalized phishing at scale, and a voice can be cloned convincingly from a few seconds of audio, which makes the classic "urgent call from the boss" or a fake vendor asking to change banking details far harder to spot. The defense is mostly human and procedural: verify money movements and credential changes through a second channel, train people to expect convincing fakes, and never let a phone call or email alone authorize a payment or a password reset.
Turn AI into a defender
The same capability that worries you is a force multiplier for defense. AI assistants triage and summarize security alerts, surface anomalies buried in logs, draft incident timelines, and flag the suspicious message your filters let through. A small security team gains the reach of a much larger one, as long as a human still owns the final call. This is the practical version of the rule that runs through this whole book: automate the toil, keep the judgment with a person.
Govern the agents, not just the models
As agents spread across a company, the new risk is not one bad answer. It is dozens of agents that nobody remembers approving, each with its own access and its own credentials. Treat every agent like a new hire: give it a named owner, a written job, the least access that job needs, and a way to switch it off. Agent management is now its own product category. Microsoft's Agent 365, for example, gives IT a single place to inventory and control agents, and other platforms are adding the same. Whatever you use, you should be able to answer three questions at any moment: which agents are running, what they can touch, and who is accountable for each one.
You do not have to invent this. Anchor your program to recognized frameworks: the OWASP Top 10 for LLM Applications for application risks, MITRE ATLAS for adversary techniques, the NIST AI Risk Management Framework for the overall program, and ISO/IEC 42001 for a certifiable management system. Layer your data protection duties (GDPR, HIPAA, and the like) on top, and you have a defensible posture rather than a pile of point fixes.
Buy, build, or assemble
Before any department rollout, decide how you will get the capability. Most enterprises need all three approaches in different places. Choosing well per workflow is what keeps cost and control in balance.
Buy
A governed, off-the-shelf assistant or platform. Fastest path, lowest build cost, least control. Ideal for building literacy and for common workflows where a vendor already does it well.
Assemble
Low code automation plus model APIs wired into your systems. Middle ground: fast to stand up, cheaper than custom code, flexible enough for the long tail of workflows.
Build
A custom platform on your own data and stack. Highest cost and control. Reserve it for the few workflows where AI creates durable, defensible advantage.
The honest sequence
The order that works for most teams: buy a governed assistant to build literacy, assemble two or three painful workflows on a low code platform to prove value quickly, then build one custom assistant on your own data where it creates lasting advantage. Add agents that can take actions last, read only first, with a human approval gate on anything that touches money or customers.
Every department chapter that follows ends with a "build it yourself" panel describing exactly what data, stack, guardrails, and skills the in-house version needs. Read those alongside this page to decide where each workflow belongs.
The build option is where teams most often overinvest or underinvest. The rule of thumb: do not build what you can buy, and do not buy what defines your edge.
How to read the department playbooks
Why the next several chapters go one department at a time, and what to expect from each.
AI does not transform a company all at once. It transforms one workflow at a time, inside one team at a time. That is why the rest of this book is organized by department rather than by technology. Each group has its own work, its own data, its own risks, and its own definition of a good day. A tool that is a breakthrough in support can be a liability in legal. The only way to get real return is to meet each department where it actually works.
The chapters that follow walk through thirteen functions, from Marketing and Sales to Field Service, Legal, IT, and Order to Cash. They are built to be read in order or jumped into as needed. You do not have to adopt all of them at once. Pick the one or two where the pain is sharpest and the numbers are clearest, prove the value there, then carry the same pattern to the next team.
Every department chapter follows the same shape, so you always know where to look:
- Where it earns its keep. The handful of workflows where AI moves a real number, not a demo.
- The integration sequence. A step by step path from first audit to scaled rollout, run through the Edge Framework: Discover, Select, Build, Govern, Scale.
- Real tools, real costs. A table of named tools, what each one does, where to get it, rough pricing, and the ROI signal to watch.
- An illustrative example. A short before and after showing the method applied to one workflow. These are composites built to illustrate the approach, not reported results from a specific company.
- Start here if you are small. The lowest cost way in for a small team, usually a tool you already own.
- Build it yourself. What it takes to own the capability in-house instead of renting it, for teams ready to go deeper.
Start where it hurts, prove the number, then repeat. That is the whole strategy.
Read each chapter with one question in mind: which single workflow on my team is costing the most hours for the least judgment? That is almost always where AI should go first. The departments are different, but the discipline is the same. Redesign the workflow, govern it from day one, and hold it to a number you would be willing to show your CFO. The pages ahead give you the map for each function; the decision about where to start is yours.
Marketing & Content
Marketing is where most enterprises feel AI first, because the work is language- and asset heavy and the feedback loop is fast. The risk is volume without quality. The win is a team that ships more, tests more, and personalizes at a scale humans cannot match.
Audit the content supply chain. Where does a blog post, ad, or email actually get stuck: ideation, drafting, design, or approvals? Pull your last quarter of campaigns and time each stage.
Pick one bottleneck with a number on it. Example: "cut first draft time for paid social variants from 3 days to same day." A clear before/after beats a vague "use AI for content."
Stand up a brand trained assistant. Load your voice guide, top performing examples, and product facts into an enterprise tool (Jasper, Writer) or a custom RAG assistant. Generate variants; a human edits and approves.
Rebuild the workflow, not just the draft. Make the assistant the first step in every brief, add a fact check and brand check gate, and connect outputs to your CMS or ad platform so approved copy ships without rekeying.
Measure output and outcome. Track assets shipped per week, time to publish, and the metric that matters (CTR, conversion, pipeline). Retire prompts that produce low acceptance drafts.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| Jasper | Marketing specific copy generation with brand voice and campaign workflows | jasper.ai | ~$39 to $69/seat/mo; Business tier custom | First draft time; assets per FTE per week |
| Writer | Enterprise platform for on brand content with governance and guardrails | writer.com | Team from ~$18/user/mo; Enterprise custom | Brand compliance rate; review cycles avoided |
| HubSpot Breeze | AI agents and copilots embedded in CRM/marketing for content, email, and social | hubspot.com | Bundled into Marketing Hub tiers; usage credits | Campaign throughput; lead to MQL lift |
| Surfer SEO | Data driven content briefs and on page optimization scoring | surferseo.com | ~$99 to $219/mo by tier | Organic rankings; content score vs. SERP |
| Adobe Firefly | Commercially safe image and video generation inside Creative Cloud | adobe.com/firefly | Credit plans from ~$9.99/mo; CC bundles | Creative production cost; time to asset |
| Persado / Mutiny | Message optimization (Persado) and website personalization (Mutiny) for conversion lift | persado.com · mutinyhq.com | Enterprise custom (annual) | Conversion rate lift vs. control |
| Canva Magic Studio | AI design, image, and resizing tools for on brand social, ads, and presentations | canva.com | Free plan; paid per seat | Design turnaround; assets per week |
| HeyGen / Synthesia | AI video with avatars and voice for training, product, and social video | heygen.com · synthesia.io | Free or starter tiers; paid by video minutes | Cost per video; time to publish |
Pricing varies by seat count and contract; treat the figures above as starting points and confirm with each vendor.
A five person marketing team spends three days producing each round of paid social ads. They load their brand guide and best performing ads into a brand trained assistant and make it the first step of every brief. Draft time falls from three days to the same afternoon, the team tests four times as many variants, and the editor's job shifts from writing to choosing.
Skip the enterprise platforms at first. A business plan from Microsoft 365 Copilot, ChatGPT, or Claude loaded with your brand guide, plus Canva for design, covers most of what a small team needs. Add Jasper or Writer when brand control across many writers becomes the bottleneck.
What you need to build the in-house version
Brand voice guide, top performing assets, product facts, and approved claims, cleaned and stored as a knowledge base.
A model API (OpenAI / Anthropic), a RAG layer over your brand docs, and a simple internal UI or a Slack/Teams bot.
A brand check and fact check pass, a banned claims list, and a human approval gate before anything publishes.
One prompt engineer / marketing ops owner; light dev help to wire the RAG and CMS connection.
Sales
Sales AI pays off in two places: giving reps back the hours they lose to research and admin, and giving managers a clear read on what is actually happening in deals. The trap is letting AI spray generic outreach; the win is sharper targeting and faster, better coached reps.
Time study a rep's week. Quantify hours in research, manual CRM entry, follow-up writing, and call notes. These are the recoverable hours.
Start with admin, not autonomy. Auto call summaries and CRM updates are low risk, high trust wins that earn rep buy-in before you touch outbound messaging.
Pilot on one team. Roll a conversation intelligence or enrichment tool to a single pod, baseline their selling time and win rate, and let results recruit the rest of the org.
Wire it into the motion. Connect the tool to the CRM and dialer so notes, next steps, and risk flags appear where reps already work. Coach managers to use the call insights in 1:1s.
Track selling time and win rate. Watch ramp time for new hires, forecast accuracy, and outreach reply rates. Kill any "personalization" that tanks reply quality.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| Gong | Records and analyzes calls/emails for coaching, deal risk, and forecasting | gong.io | ~$1,600/user/yr + platform fee; custom | Win rate; ramp time; forecast accuracy |
| Clay | Data enrichment and AI research that builds and personalizes prospect lists | clay.com | ~$149 to $800+/mo by tier + credits | Cost per qualified contact; reply rate |
| Apollo.io | Prospecting database plus AI sequencing and outreach | apollo.io | Free tier; paid ~$49 to $99/seat/mo | Meetings booked per rep; data accuracy |
| Salesforce Agentforce | Autonomous and assistive agents for SDR, service, and CRM tasks | salesforce.com/agentforce | Consumption based (per action/credit); custom | Tasks handled per agent; pipeline created |
| ZoomInfo Copilot | Buyer intent signals and account research surfaced to reps | zoominfo.com | Enterprise custom (annual) | Intent sourced pipeline; account coverage |
| Clari | Revenue platform for forecasting, pipeline inspection, and deal risk | clari.com | Enterprise custom (annual) | Forecast accuracy; slipped deals |
| Otter / Fireflies / Granola | AI meeting notes, summaries, and follow ups that sync to the CRM | otter.ai · fireflies.ai · granola.ai | Free tiers; paid per seat | Admin hours saved per rep |
Most sales platforms quote annually by seat and module; verify current pricing and minimums directly.
A ten rep sales team loses about six hours per rep each week to call notes and CRM updates. They roll AI meeting notes and automatic CRM updates to one pod first. Admin time drops by half, managers coach from real call data instead of memory, and the other pods ask for the tool within a month.
Start with an AI note taker such as Otter, Fireflies, or Granola connected to your CRM, and use Apollo's free tier for prospecting. Gong and Clari earn their price once the team is large enough to coach and forecast across.
What you need to build the in-house version
CRM access, call transcripts (from your dialer/meeting tool), and your ICP and playbook docs.
A model API for summarization, a transcription source, and an automation layer (Zapier/Make) to push summaries and fields into the CRM.
Rep review before any auto sent message; never let an agent email a prospect without a human gate early on.
A RevOps owner plus light integration work; no ML team required for the assistive version.
Customer Support & Success
Support is the department with the cleanest AI business case: a large volume of repetitive questions, a measurable resolution rate, and per outcome pricing that maps directly to value. Done well, AI deflects routine tickets and lets humans handle the hard, high empathy cases.
Tag your ticket volume. What share of contacts are repetitive, answerable from existing docs, and low risk? That percentage is your deflection ceiling.
Start with a copilot, then an agent. Drafting suggested replies for human agents is the safe first step; full autonomous resolution comes once your knowledge base is clean.
Feed it real knowledge. Connect help center articles and past resolved tickets. Run the AI agent on a slice of traffic with a confidence threshold and a clean human handoff.
Close the content gaps. Every AI miss is a missing or unclear article. Build a loop where unresolved questions become knowledge updates, raising the resolution rate over time.
Watch resolution and CSAT together. Deflection that tanks satisfaction is a false win. Track resolution rate, reopen rate, CSAT on AI handled contacts, and cost per resolution.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| Intercom Fin | AI agent that resolves customer conversations end-to-end across channels | intercom.com/fin | $0.99 per resolution + seat plans from ~$29/seat/mo | Resolution rate; cost per resolution; CSAT |
| Zendesk AI | Automated resolutions, agent copilot, and intelligent routing in the Zendesk suite | zendesk.com | Suite seats + automated resolution pricing | Auto resolution %; handle time |
| Ada | Multichannel AI agent platform with an LLM reasoning engine | ada.cx | Enterprise custom (reported ~$30K/yr and up) | Automated resolution rate; containment |
| Sierra | Enterprise conversational AI agents for complex, branded support | sierra.ai | Enterprise custom (six figure year one typical) | Resolution quality; brand adherence |
| Decagon | AI support agents for high volume B2C operations | decagon.ai | Enterprise custom (reported from ~$95K/yr) | Containment; agent hours saved |
Per resolution models scale with success, so forecast cost at your real ticket volume and resolution rate before committing.
Support is where automation and AI compound fastest. Automation routes, tags, and closes the simple tickets and updates the record, while an AI assistant drafts grounded replies and resolves common tier-1 issues end to end. Agents stop copy pasting macros and spend their time on the harder, higher stakes conversations.
A support team handling 4,000 tickets a month finds that 40 percent are password resets, order status, and the same ten how to questions. They clean up those help articles, then let an AI agent answer that slice with a clear handoff to a person. Automated resolution settles near 30 percent of total volume within a quarter, CSAT holds steady, and agents spend their time on the hard cases.
If you already use Intercom, Zendesk, Freshdesk, or HubSpot, turn on the AI features in the tool you have before buying anything new. Per resolution pricing means you pay only when the AI actually closes a ticket.
What you need to build the in-house version
A clean, current help center and a corpus of resolved tickets to ground answers and set tone.
A model API, a RAG pipeline over your knowledge base, a vector store, and your helpdesk's API for handoff.
A confidence threshold, "I'll connect you to a human" fallbacks, and zero auto actions on billing or account changes without verification.
A support ops owner for content, plus engineering for the RAG and helpdesk integration and ongoing evals.
Operations & Back Office
Operations is where AI quietly compounds. The wins are rarely glamorous (a form that fills itself, an exception that routes itself, an SOP that writes itself), but they remove friction from the processes that run the whole company. This is automation's home turf, now supercharged by language models.
Hunt for swivel chair work. Anywhere a person copies data between two systems, retypes a PDF, or routes a request by hand is a candidate. Inventory these with volume and time per task.
Pick high volume, rule bound tasks first. Invoice intake, order entry, and status updates have clear logic and clear savings, ideal for an automation + AI extraction combo.
Assemble before you build. Most ops wins come from an automation platform (Power Automate, Zapier, n8n) plus an AI extraction or reasoning step, with no full software project required.
Document the new process and own the exceptions. Automate the 80% happy path and route the 20% edge cases to a person. Update the SOP so the team trusts and maintains the flow.
Monitor for silent failure. Automations break quietly when an upstream system changes. Alert on error rates and exception spikes, and review monthly.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| UiPath | Enterprise RPA plus AI agents for document understanding and process automation | uipath.com | Enterprise custom (per bot/per user) | Hours automated; error reduction |
| Microsoft Power Automate | Low code workflow automation with AI Builder and Copilot | microsoft.com/power-automate | ~$15/user/mo; per flow and credit options | Processes automated; cycle time |
| Zapier | Connects 7,000+ apps with AI steps and agents for cross tool automation | zapier.com | Free tier; paid by task volume (~$20+/mo) | Manual tasks eliminated per month |
| n8n | Open source, self hostable workflow automation with native AI/agent nodes | n8n.io | Self host free (OSS); cloud from ~$20+/mo | Cost vs. SaaS; flows owned in-house |
| Automation Anywhere | Cloud native RPA with generative AI process automation | automationanywhere.com | Enterprise custom; usage based tiers | FTE hours returned; straight through rate |
| Microsoft Copilot Studio | Build and publish AI agents and automations on Microsoft 365 and Power Platform | microsoft.com | Usage based capacity; confirm with Microsoft | Requests handled per agent |
| Make | Visual automation builder with AI modules for multistep workflows | make.com | Free tier; paid by operations volume | Manual steps removed |
For ops, the cheapest viable path is often an automation platform you already license (e.g., Power Automate inside Microsoft 365) plus a model API for the AI step.
Let rule based automation move data between systems, kick off the standard workflows, and close the routine requests, then let AI handle the judgment: reading messy documents, flagging the items that need attention, and routing exceptions to a person. The result is fewer manual handoffs and a team focused on the calls that matter.
An operations team keys 1,200 vendor invoices a month by hand from PDFs. They build a flow that reads each invoice with AI, checks it against the purchase order, and drops clean ones straight into the ERP while routing the rest to a person. About 70 percent go through untouched, and the two people who spent their weeks keying data move to vendor management and exceptions.
If you have Microsoft 365, Power Automate is likely already in your license. Otherwise, Zapier or Make with an AI step will handle most small team automations for a modest monthly fee.
What you need to build the in-house version
Sample documents (invoices, orders, forms) and the schema of the destination system you are filling.
An automation platform as the spine, a model API or vision model for extraction, and API/webhook connections to your ERP or systems of record.
Confidence scoring with human review for low confidence extractions, and validation rules before anything writes to a system of record.
An ops/automation owner; light scripting. RPA developers only if you are automating legacy desktop apps without APIs.
Field Service Operations
Field service runs on trucks, technicians, and time. Every repeat visit, every missing part, and every report written at the end of a long day is margin walking out the door. AI earns its keep here by putting the right history, the right part, and the likely fix in front of the technician before the truck rolls, and by turning the paperwork after the job into minutes instead of an hour.
Follow the work order. Trace a typical call from request to invoice and time each handoff. Pull first time fix rate, repeat visits within 30 days, the average time from job complete to invoice, and the hours technicians spend on paperwork.
Pick the leak with the biggest dollar sign. Example: "raise first time fix from 75 to 85 percent" or "get every service report written the same day so invoices go out within 24 hours."
Put knowledge in the technician's hands. Load equipment manuals, service history, and past fixes into your field service platform's AI or a grounded assistant technicians can ask from the truck. Add AI drafting that turns a technician's notes or voice into a clean service report.
Keep the technician as the authority. AI suggests the likely fix and drafts the report; the technician confirms what was actually done, and anything safety related or billable gets a human sign off. Put in writing how driver and camera data will be used.
Measure fix rate and cash speed. Track first time fix, repeat visits, report completion time, and days from job complete to invoice. Expand from one branch or service line once the numbers move.
Telematics and AI dashcams can make roads safer, but they also watch people at work. Tell technicians what is collected and why, use the data for coaching rather than surprise discipline, and check state laws on employee monitoring and recording before rollout. AI troubleshooting advice never overrides lockout procedures, refrigerant handling rules, or electrical safety practice.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| ServiceTitan | Field service management with Titan Intelligence for job insights, pricing, and call analysis | servicetitan.com | Custom, priced by technician | Average ticket; booked call rate |
| BuildOps | Field service and project platform for commercial contractors with AI assisted scheduling and reporting | buildops.com | Custom, priced by technician | Report turnaround; days to invoice |
| Salesforce Field Service | Scheduling, mobile work, and Agentforce agents for service organizations | salesforce.com | Per user add-on; custom | Schedule utilization; first time fix |
| Aquant | Service intelligence that guides technicians to the likely fix from service history | aquant.ai | Enterprise custom | First time fix; repeat visits |
| Samsara | Telematics, AI dashcams, and routing for service fleets | samsara.com | Subscription per vehicle | Safety events; drive and idle time |
| Motive | AI dashcams, driver coaching, and fleet operations | gomotive.com | Subscription per vehicle | Accident rate; fuel spend |
Field service platforms are usually priced per technician or per vehicle on annual contracts. Ask vendors to break out the AI features so you know what you are paying for.
Let automation create the work order, confirm the appointment, send the "technician on the way" text, and push the finished job to billing. Let AI rank which calls need a senior technician, suggest the parts to load, and draft the service report from the technician's notes. The dispatcher stops juggling and the technician stops typing.
A 25 technician commercial service team writes reports at the end of the day, so invoices wait an average of four days after the work is done. They pilot voice to report drafting on one crew: the technician talks through what they found and fixed, AI drafts the report, and the technician corrects it before leaving the site. Within a quarter, reports are finished the same day, days to invoice drop from four to one, and each technician gets back about 30 minutes a day.
If you run a handful of trucks, you likely do not need an enterprise platform. Jobber and Housecall Pro both bundle scheduling, invoicing, and AI features for small trades, and a general assistant can turn a technician's notes into a clean service report today.
What you need to build the in-house version
Equipment manuals, service history, past work orders with their resolutions, parts lists, and your price book.
Your field service platform's API, a model API with a RAG layer over manuals and fix history, speech to text for voice notes, and a mobile front end.
The technician confirms every report, no automatic billing changes, safety procedures always override AI suggestions, and monitoring data is used under a written policy.
A service manager who owns the content, plus light integration help to connect the field service platform.
Finance & Accounting
Finance demands accuracy, traceability, and control, which makes it a place to deploy AI deliberately, not loosely. The highest value uses keep a human accountable for every number: AI drafts, codes, and surfaces; people review and approve. The payoff is a close that runs faster and an FP&A team that spends time on analysis instead of assembly.
Map the close and the cash cycle. Where do days disappear: coding invoices, chasing approvals, reconciling accounts, building decks? Time each step of close.
Automate coding and reconciliation first. These are repetitive, rule based, and auditable, the safest place to start, with a clean audit trail.
Keep humans accountable. Let AI propose GL codes, match transactions, or draft a variance narrative; a controller reviews and signs off. Never auto post without review during the pilot.
Build the control into the workflow. Bake AI suggestions into your AP and FP&A tools with approval thresholds, and document the control for auditors.
Track close days and exception rates. Measure days to close, touchless invoice rate, and how often AI codings are overridden. Rising overrides mean the model needs attention.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| Ramp | Spend management with AI that codes expenses, flags policy issues, and finds savings | ramp.com | Core free; paid tiers + interchange model | Time on expense admin; savings found |
| BILL | AP/AR automation with AI invoice capture and approval routing | bill.com | ~$45 to $79+/user/mo by tier | Touchless invoices; approval cycle time |
| Vic.ai | Autonomous accounts payable processing and invoice automation | vic.ai | Enterprise custom (by volume) | Cost per invoice; straight through rate |
| Microsoft Copilot in Excel | Natural language analysis, formula generation, and charting in spreadsheets | microsoft.com/copilot | Copilot Business $21/user/mo (under 300 users) or Enterprise $30, plus M365 base | Modeling time; analyst hours freed |
| Datarails / Cube | FP&A platforms with AI for reporting, variance, and consolidation | datarails.com · cubesoftware.com | Enterprise custom (annual) | Reporting cycle time; forecast accuracy |
| FloQast | Close management with AI for reconciliations, checklists, and flux analysis | floqast.com | Custom, priced by user | Days to close; reconciliation time |
Treat finance AI as decision support with a human signer, not an autonomous bookkeeper. The audit trail is part of the deliverable.
Finance gains the most from pairing automation with AI. Straight through automation captures invoices, runs the three way match, and posts the clean entries, while AI reads the messy documents, drafts the variance narrative, and surfaces the anomalies worth a second look. The close shrinks from a week of keying to a day of reviewing.
A finance team closes the books in eight business days, most of it spent coding invoices and reconciling accounts. They add AI suggested GL coding and transaction matching, with a controller reviewing every exception. Within two quarters the close drops to five days, and the time saved goes into variance analysis the leadership team actually reads.
Ramp's core plan and the AI features already inside QuickBooks or Xero cover expense coding and basic reconciliation for most small companies. Copilot in Excel or a general assistant can draft the monthly variance narrative from your own numbers.
What you need to build the in-house version
Read only access to your GL/ERP, the chart of accounts, prior period actuals, and budget data.
A model API for analysis and narrative, a secure data connector to your ERP, and a spreadsheet or BI surface for output.
No write access to financial systems; numbers come from the source of record, and AI only explains and drafts. Mandatory human review.
A finance systems owner plus a data/integration engineer; strong data governance from day one.
Human Resources
HR runs on documents and conversations (policies, cases, questions, reviews), which is exactly what language models handle well. The opportunity is to give every employee instant, accurate answers and free HR business partners for the human work. The duty of care is high: this is people's data and people's livelihoods.
Find the repetitive asks. Pull your HR ticket and inbox data. The same benefits, PTO, and policy questions repeat constantly. That is your self-service opportunity.
Start with answers, not decisions. A policy Q&A assistant grounded in your handbook is high value and low risk. Keep AI out of hire/fire/pay decisions.
Ground it in your real policies. Load the current handbook and benefits docs into a RAG assistant with strict scoping, so employees only see what applies to them.
Set the privacy and escalation rules. Define what the assistant can answer, when it must hand off to a human, and how sensitive topics (harassment, leave, accommodations) always route to a person.
Audit for accuracy and fairness. Spot check answers, monitor for outdated policy responses, and review any analytics models for disparate impact.
Anything touching hiring, pay, promotion, or termination is treated as high risk under the EU AI Act and is increasingly regulated in the US.5 Keep humans accountable for every employment decision, restrict who can query personal data, and never feed an open consumer tool with employee records. Use enterprise tiers that exclude your data from training.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| Workday AI (Illuminate) | Embedded AI across HCM for self-service, insights, and process automation | workday.com | Enterprise custom (part of HCM) | HR case deflection; admin time saved |
| Visier | People analytics with AI to surface attrition, DEI, and workforce trends | visier.com | Enterprise custom (annual) | Attrition predicted vs. actual; decision speed |
| Lattice | Performance and engagement platform with AI writing assistance | lattice.com | ~$11 to $15/person/mo by module | Review completion; manager time |
| Microsoft 365 Copilot / ChatGPT / Claude | General assistants for drafting policies, comms, and case notes on business plans that exclude your data from training | microsoft.com · openai.com · claude.ai | Copilot Business $21/user/mo (under 300 users), Enterprise $30, plus M365 base; ChatGPT and Claude per seat business plans | Drafting time; HRBP hours freed |
| Moveworks (ServiceNow) | AI assistant that resolves employee HR and IT requests across systems; now part of ServiceNow | moveworks.com | Enterprise custom (annual) | Requests auto resolved; HR ticket volume |
HR runs on repeatable workflows, which is exactly what automation is for. Automate onboarding checklists, PTO approvals, and document generation, and let AI answer routine policy questions and draft the first version of postings and reviews. People team time shifts from paperwork to the human work of hiring, coaching, and retention.
A 600 person company's HR team answers the same benefits, PTO, and policy questions hundreds of times a month. They launch a policy assistant in Teams grounded only in the current handbook, with sensitive topics routed straight to a person. Within a few months, routine questions to the HR inbox fall by more than half, and HR business partners spend that time on coaching and retention.
A business plan from Microsoft 365 Copilot, ChatGPT, or Claude with your handbook loaded as a knowledge source can answer routine policy questions on day one. Keep anything about pay, leave, discipline, or complaints with a person.
What you need to build the in-house version
Current handbook, benefits summaries, and PTO/leave policies, version controlled so answers stay accurate.
An enterprise model API, a RAG layer with per employee scoping, and a Slack/Teams or intranet front end.
Hard escalation rules for sensitive topics, no access to individual records unless authorized, and a "verify with HR" disclaimer on edge cases.
An HR ops owner for content currency; engineering for scoped retrieval and access control.
Recruiting, Talent Acquisition & Interviewing
Recruiting AI can collapse the busywork at the top of the funnel (sourcing, screening, scheduling, and candidate communication) so recruiters spend their time where judgment matters. It is also the function regulators watch most closely, because automated screening can encode and scale bias. Speed is the prize; fairness is the nonnegotiable.
Find the funnel drag. Time study sourcing, screening, and scheduling. Most recruiter hours vanish into search and coordination, not evaluation.
Automate logistics, assist on judgment. Scheduling and candidate FAQ are safe full automation wins; screening must stay assistive with human review.
Pilot with a bias check built in. Run a sourcing or screening tool on one req family, and from day one measure pass through rates across demographic groups against your baseline.
Document the audit and the human role. Recruiters make every advance/reject call. Keep records of how the tool is used, because several jurisdictions now require bias audits and candidate notice.
Track time to fill and adverse impact. Watch time to fill, quality of hire, candidate experience, and selection rates by group. Pause any tool that shows disparate impact.
Automated screening is regulated. New York City's Local Law 144 requires an independent bias audit and candidate notice for automated employment decision tools, and the EU AI Act classes hiring AI as high risk.6 Keep a human accountable, bias audit before and during use, and confirm any vendor can supply audit results.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| Eightfold | Deep learning matching of candidates to roles using skills, not just keywords | eightfold.ai | Enterprise custom (annual) | Time to fill; match quality; diversity of slate |
| SeekOut | AI sourcing across a large talent graph with diversity and skills filters | seekout.com | Enterprise custom (annual) | Qualified candidates surfaced per req |
| Paradox (Olivia) | Conversational AI that screens, schedules, and answers candidates 24/7 | paradox.ai | Enterprise custom (by volume) | Scheduling time saved; apply to interview rate |
| LinkedIn Hiring Assistant | AI agent for sourcing and candidate management inside LinkedIn Recruiter | linkedin.com/talent-solutions | Recruiter seat + add-on (custom) | Recruiter productivity; InMail response |
| Microsoft 365 Copilot / ChatGPT / Claude | Drafting job descriptions, outreach, structured interview guides, and screening rubrics | microsoft.com · openai.com · claude.ai | Copilot Business $21/user/mo (under 300 users), Enterprise $30, plus M365 base; ChatGPT and Claude per seat business plans | JD and guide drafting time |
Interviewing: structure over surveillance
Interviewing is where AI can make hiring both faster and fairer, if you point it at structure and consistency rather than at judging people through a webcam. The strongest, lowest risk use is interview intelligence: generating structured guides, capturing notes, and helping every interviewer evaluate the same things the same way.
Start by auditing consistency: do interviewers ask the same job related questions and score the same way? Then lead with structure and notes. Auto generated interview guides and AI note taking tied to your rubric are high value and low risk, and debriefs should reference the captured evidence, not gut feel. Track interviewer agreement, quality of hire, and selection rates by group to confirm the structure is improving outcomes.
Recording and analyzing interviews triggers consent laws and the same high risk hiring rules as screening.56 Get explicit candidate consent before recording, prefer tools that assess content against job criteria over tools that infer traits from appearance or voice, and keep a human as the decision maker.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| BrightHire | Interview intelligence: records, transcribes, and structures interviews for fairer hiring | brighthire.com | Enterprise custom (annual) | Interviewer consistency; quality of hire |
| Metaview | AI note taker that produces structured interview notes and scorecards | metaview.ai | Per seat tiers; free trial | Note taking time; debrief speed |
| HireVue | Structured video interviews and skills assessments with AI scoring | hirevue.com | Enterprise custom (annual) | Time to interview; completion rate |
Favor tools that improve structure and consistency. Be cautious with any product that claims to judge a candidate's competence from facial expression or tone.
A recruiting team hiring 200 hourly workers a year spends most of its week on scheduling. They automate interview scheduling and candidate FAQs, keep every screening decision with a recruiter, and add structured interview guides with AI notes. Time to fill drops by about a week, interviewers score more consistently, and selection rates by group are checked every month.
Most applicant tracking systems built for smaller companies now include AI drafting and scheduling, so check yours first. A general assistant can write job descriptions and structured interview guides today. Hold off on automated screening until you can run a bias audit.
What you need to build the in-house version
Your structured job requirements and a clear, job related rubric, never proxies for protected characteristics.
A model API to summarize and structure resumes against the rubric, plus your ATS API for read/write.
Score against job related criteria only, surface evidence for every rating, keep a recruiter decision on every candidate, and run a bias audit before launch.
A TA ops owner plus engineering; involve legal/compliance and an audit partner early.
IT & Engineering
Engineering has the most measurable AI ROI in the building, because the unit of work (code, tickets, incidents) is countable. AI coding assistants are now standard, and the frontier has moved to agents that write, review, and ship. IT, meanwhile, uses AI to deflect help desk tickets and triage incidents before a human ever looks.
Baseline developer and IT throughput. Measure cycle time, PR throughput, and help desk ticket mix before you roll anything out, so the gain is provable.
Coding assistants first. They have the clearest payback and fastest adoption. For IT, internal help desk deflection mirrors the customer support playbook.
Pilot with guardrails on data. Roll a coding assistant to a team with policy controls on public code matching and secrets. Connect an enterprise search tool to internal docs for codebase Q&A.
Move from suggestions to agents carefully. Let agents draft PRs and triage incidents with human review and budget caps. Usage based AI billing can spike fast on agentic workloads, so set org limits.
Track flow and quality, not just speed. Watch lead time, change failure rate, and review load. Faster code that breaks more is not a win.
GitHub's own research reported developers completing tasks up to 55% faster with Copilot, and follow-on studies have cited several hours saved per developer per week; a Forrester commissioned study modeled a triple digit ROI over three years.7 Vendor sponsored figures run optimistic, so measure your own baseline, but the direction is consistent across studies.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| GitHub Copilot | AI pair programmer and agent across IDEs and the GitHub workflow | github.com/features/copilot | Business ~$19/user/mo; Enterprise ~$39 (+GHE) | PR throughput; cycle time; acceptance rate |
| Cursor | AI native code editor with agentic multifile editing | cursor.com | Pro ~$20; Business ~$40/user/mo | Developer velocity; satisfaction |
| Claude Code | Agentic coding from terminal, IDE, or app for larger autonomous tasks | anthropic.com/claude-code | Team Premium ~$100/seat/mo, or API usage | Tasks completed per developer; toil reduced |
| Glean | Enterprise search and assistants grounded in your internal knowledge | glean.com | Enterprise custom (per user annual) | Time to answer; ticket deflection |
| Snyk | AI assisted security scanning and fix suggestions in the dev workflow | snyk.io | Free tier; paid per product/seat | Vulns caught preprod; mean time to remediate |
| ServiceNow Now Assist | Generative AI across IT service management for ticket summaries, resolution, and agents | servicenow.com | Enterprise custom (add-on to ITSM) | Mean time to resolve; deflection |
| PagerDuty | Incident response with AIOps for noise reduction, triage, and summaries | pagerduty.com | Tiered per user; AIOps add-on | Alert noise reduced; time to acknowledge |
| OpenAI Codex | Cloud coding agent that writes, tests, and proposes changes in parallel | openai.com/codex | Included in ChatGPT paid plans or by API usage | Tasks completed; review load |
Agentic coding and premium models increasingly bill by token/credit usage. Set per org budget caps before broad rollout to avoid surprise invoices.
A 40 person engineering team rolls a coding assistant to one squad and measures before and after. Pull requests move faster, the internal help desk adds an AI agent that resolves routine access and password requests, and the team watches change failure rate to make sure speed is not costing quality.
GitHub Copilot or Cursor on a business plan covers most small engineering teams. For IT support, the AI features in the help desk tool you already own are the cheapest place to start.
What you need to build the in-house version
Your repos, internal docs, runbooks, and architecture decisions, indexed for retrieval.
A model API (or self hosted open weights for sensitive code), a RAG layer over your knowledge, and IDE/Slack integrations. See the Build Lab.
Secret scanning, public code match controls, scoped repo access, and human review before merge or deploy.
A platform engineering owner; this is the one region where an in-house ML/infra capability genuinely pays off.
Legal & Compliance
Legal work is dense, document heavy, and expensive, a natural fit for AI that reads, drafts, and compares at speed. The catch is that hallucination is not an inconvenience here; it is malpractice risk. The right pattern uses purpose built legal AI with citations and a lawyer reviewing every output, turning hours of review into minutes of verification.
Find the repetitive review work. NDAs, vendor contracts, and standard agreements consume disproportionate hours. Quantify volume and average review time.
Start with review and extraction, not advice. Flagging nonstandard terms against your playbook is high value and verifiable; AI never gives the final legal opinion.
Use legal grade tools with sources. Pilot a tool that cites its basis so a lawyer can verify fast. Avoid open consumer chatbots for anything privileged or confidential.
Encode your playbook. Load your standard positions and fallback clauses so the AI flags deviations consistently, with a lawyer approving every redline that goes out.
Track turnaround and accuracy. Measure contract turnaround time and verify accuracy on a sample. A rising correction rate means the model or playbook needs tuning.
Treat confidentiality and verification as absolute. Keep privileged material out of consumer tools, require citations a lawyer can check, and have an attorney sign off on every output. Courts have already sanctioned lawyers for filing AI fabricated citations. Verification is not optional.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| Harvey | Domain specific legal AI for research, drafting, and analysis | harvey.ai | Enterprise custom (annual) | Hours per matter; review time saved |
| Spellbook | Contract drafting and review inside Microsoft Word | spellbook.legal | Per seat (annual); quote based | Contract turnaround; redline quality |
| Ironclad | Contract lifecycle management with AI for extraction and workflow | ironcladapp.com | Enterprise custom (annual) | Cycle time; self-serve contract rate |
| Legora | Collaborative legal AI for review, research, and drafting across matters | legora.com | Enterprise custom (annual) | Review time; matters per lawyer |
| Thomson Reuters CoCounsel | Professional grade legal research and document review assistant | thomsonreuters.com | Enterprise/professional custom | Research time; citation reliability |
| Luminance | Legal AI for contract review, negotiation, and due diligence | luminance.com | Enterprise custom (annual) | Contract turnaround; issues flagged |
A three person in house legal team reviews about 60 NDAs and vendor contracts a month. They encode their standard positions as a playbook and use an AI review tool to flag every deviation, with a lawyer approving each redline. Turnaround on standard agreements falls from five days to one, and the attorneys spend their time on the deals that are actually negotiable.
Spellbook works inside Word for contract review without a large enterprise contract. A business tier general assistant with your playbook loaded can flag nonstandard terms, as long as privileged material stays in an approved tool and a lawyer checks every output.
What you need to build the in-house version
Your contract playbook, standard and fallback clauses, and a set of past redlines as examples.
An enterprise model API with strong retrieval over your clause library; a Word or CLM integration for delivery.
Citations to your playbook for every flag, no privileged data in nonapproved tools, and mandatory attorney review.
A legal ops owner who can encode the playbook; engineering for secure retrieval and integration.
Executive, Strategy & Business Intelligence
For leadership, AI's value is decision velocity: turning scattered internal data and external signals into answers in minutes. The discipline is to insist on sourced, verifiable outputs. A confident sounding wrong answer is far more dangerous at the board table than at the help desk.
Map your decision inputs. Where do executives wait days for a number, a market read, or a synthesis? Those waits are the targets.
Start with research and BI, sourced. Natural language analytics over governed data and AI market research with citations are the highest trust starting points.
Pilot on a recurring deliverable. Pick the weekly business review or a competitor scan and let AI assemble the first draft from governed sources, with a human owner verifying.
Standardize the trusted sources. Connect AI only to governed data and vetted external feeds, and require every figure to trace back to a source.
Track decision speed and trust. Measure time to insight and how often outputs are accepted without rework. Falling acceptance signals a data or grounding problem.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| AlphaSense | AI market and competitive intelligence across filings, news, and research | alpha-sense.com | Enterprise custom (annual) | Research time; signal to decision speed |
| Hebbia | AI that reasons over large internal document sets for analysis and diligence | hebbia.com | Enterprise custom (annual) | Analysis throughput; diligence speed |
| Power BI Copilot / Tableau Pulse | Natural language analytics and automated insights over governed data | powerbi.microsoft.com · tableau.com | Platform license + AI add-on | Self-serve query volume; analyst load |
| Glean | Cross source enterprise search and assistants for leadership questions | glean.com | Enterprise custom (per user annual) | Time to answer; adoption |
| ChatGPT / Claude (research features) | Sourced research reports, board narrative drafts, and scenario thinking | openai.com · claude.ai | Per seat business plans | Prep time; quality of synthesis |
| ThoughtSpot | Natural language search and AI analyst over governed warehouse data | thoughtspot.com | Tiered; enterprise custom | Self serve questions answered |
| Perplexity Enterprise | Cited web and internal research for fast market and competitor scans | perplexity.ai | Per seat enterprise plans | Research time; sources per brief |
A leadership team waits every Monday for the weekly business review deck, which takes an analyst most of Friday to build. They connect an AI assistant to governed dashboards and let it draft the review with every number linked to its source. The analyst now spends an hour verifying instead of a day assembling, and the meeting starts with the questions instead of the numbers.
The AI features inside the BI tool you already have, such as Power BI Copilot or Tableau, plus a general assistant with research features, cover most leadership needs without an enterprise research subscription.
What you need to build the in-house version
Governed access to your BI/warehouse, key metrics definitions, and approved external feeds.
A model API, a connector to your BI layer (read only), and a template that enforces sourced, structured output.
Every number traces to a source, no fabricated figures, and a named human owner who verifies before it circulates.
A BI/strategy owner plus a data engineer; tight metric governance so the AI cannot misdefine a KPI.
Supply Chain & Procurement
Supply chain is where AI pays back fastest in hard dollars, because the work is forecasting, planning, and exception handling at a scale no spreadsheet survives. The win is a network that sees demand sooner, positions inventory smarter, and flags disruption before it becomes a stockout or a fire drill.
Map where the plan breaks. Pull your last year of forecasts against actuals and find the SKUs, lanes, and suppliers that miss most. Time how long it takes to react to a disruption today.
Pick one driver with a dollar sign. Example: "cut excess safety stock on the top 200 SKUs by 15 percent without dropping fill rate." A measurable target beats "use AI for the supply chain."
Stand up AI assisted planning. Use a demand planning platform or a custom model over your order history, then layer supplier and visibility signals so the plan reacts to reality, not last quarter.
Keep a planner in the loop. Add confidence thresholds, an override log, and an exception queue so people approve the big moves and the system handles the routine ones automatically.
Measure fill rate and working capital. Track forecast accuracy, inventory turns, and cash tied up in stock. Expand to more SKUs and lanes once the first cohort proves out.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| o9 Solutions | AI driven integrated planning across demand, supply, and revenue | o9solutions.com | Enterprise custom (annual) | Forecast accuracy; planner productivity |
| Blue Yonder | Cognitive demand, fulfillment, and warehouse planning | blueyonder.com | Enterprise custom | Fill rate; inventory turns |
| Kinaxis | Concurrent planning with the Maestro AI layer for fast what ifs | kinaxis.com | Enterprise custom | Time to replan; excess inventory |
| project44 | Real-time transportation visibility and ETA prediction | project44.com | Subscription by volume | On time delivery; expedite spend |
| ThroughPut.ai | Supply chain analytics that spot bottlenecks and free up cash | throughput.ai | Subscription (tiered) | Bottleneck cycle time; working capital |
| GEP SMART | AI enabled procurement and spend analysis | gep.com | Enterprise custom | Addressable spend captured; cost avoidance |
| Pactum | AI agents that negotiate routine supplier terms and contracts at scale | pactum.com | Enterprise custom | Savings per negotiation; suppliers covered |
Most supply chain AI is sold as enterprise software; treat figures as directional and scope a pilot on one category or region first.
Supply chain is where automation and AI work best together. Let rule based automation handle the routine replenishment and PO creation, and let AI handle the judgment: forecasting demand, ranking supplier risk, and surfacing the exceptions a planner should actually look at. The planner stops keying orders and starts managing the few decisions that move working capital.
A distributor carries too much safety stock on its top 200 items because the forecast misses every seasonal swing. They layer an AI demand model over two years of order history and route big replenishment changes to a planner for approval. Excess inventory on those items drops about 15 percent with no loss in fill rate, freeing cash that had been sitting on shelves.
Most small companies do not need an enterprise planning suite. Start with the forecasting built into your ERP or inventory system, or a spreadsheet model a general assistant helps you build, and move to o9 or Kinaxis only when complexity demands it.
What you need to build the in-house version
Order and shipment history, lead times, supplier performance, and on hand inventory, cleaned and joined into one planning view.
A forecasting model (statistical or ML), a rules layer for replenishment, and a dashboard plus alerting on exceptions.
Confidence thresholds, an override log, and human approval on any large or unusual order before it releases.
A demand planner who owns the logic, plus light data engineering to wire the feeds and keep them fresh.
Order to Cash
Order to cash is the heartbeat of the business, from the moment an order lands to the day the cash clears. It is also full of manual matching, follow-ups, and reconciliation that quietly delay revenue. AI and automation here mean you get paid faster, with fewer people chasing paper.
Find where cash gets stuck. Walk the cycle from order to cleared payment and time each stage. Pull DSO, the aging report, and the percentage of payments that auto match today.
Target one stage with a number. Example: "raise cash application auto match from 60 to 85 percent" or "cut DSO by five days on the top 100 accounts." Pick the stage bleeding the most time.
Stand up autonomous receivables. Use an AR platform or a custom matcher over your remittance and bank data, and add an assistant that drafts collections outreach tuned to each account's history.
Gate the money moves. Auto apply clean matches, route the ambiguous ones to a human, and keep approval on credit limits, write-offs, and anything customer facing.
Measure DSO and cost to collect. Track days sales outstanding, auto match rate, and collector capacity freed. Expand from the first account tier to the full ledger.
| Tool | What it does | Where to access | Typical cost | ROI signal to watch |
|---|---|---|---|---|
| HighRadius | Autonomous receivables across cash application, credit, and collections | highradius.com | Enterprise custom | Auto match rate; DSO |
| Billtrust | AI assisted billing, payments, and cash application | billtrust.com | Subscription (tiered) | Time to cash; invoice accuracy |
| Tesorio | Cash flow performance and collections prioritization | tesorio.com | Subscription (tiered) | DSO; forecast accuracy |
| Esker | Order to cash automation across the full cycle | esker.com | Enterprise custom | Touchless order rate; cost to collect |
| Sidetrade | AI agents (Aimie) for collections and cash prediction | sidetrade.com | Enterprise custom | Cash collected; promise to pay kept |
| BlackLine | Cash application and reconciliation automation | blackline.com | Enterprise custom | Unapplied cash; close time |
Many of these integrate with Sage Intacct, NetSuite, and SAP; confirm your ERP connector and remittance formats before scoping.
Order to cash is the clearest automation win in the back office. Straight through processing applies the clean payments, posts the invoices, and sends the first two reminders without anyone touching them. AI handles the gray area: predicting who will pay late, drafting the right toned outreach, and matching messy remittances. Your collectors stop doing data entry and start working the accounts that actually need a human.
An accounts receivable team spends two days a week matching payments to invoices because remittances arrive in a dozen formats. They add an AI matcher that applies clean payments automatically and queues the rest for a person, plus drafted collections emails tuned to each account. Auto match rises from 60 to 85 percent and DSO falls by five days.
QuickBooks and Xero include automated reminders and payment matching, and tools like Upflow or Chaser add collections workflows without an enterprise contract.
What you need to build the in-house version
Open invoices, bank and remittance feeds, payment history, and contact records, joined to your ERP ledger.
A matching engine for cash application, a model to rank collection risk, and an assistant that drafts outreach from account history.
Auto apply only high confidence matches, log every write-off, and require approval on credit and customer messages.
An AR lead who owns the rules, plus integration help to connect the bank, ERP, and remittance sources.
Conclusion
What to do with everything in this book, starting Monday.
If you take only one idea from these pages, make it this one. AI is not a project you launch, it is a practice you build. The companies pulling ahead are not the ones with the biggest models or the largest budgets. They are the ones who picked a real workflow, put a tool on it, measured the result, and then did it again, team after team, until it became simply how the work gets done.
Everything here points back to a few simple moves. Start where the pain is sharpest and the number is clearest. Redesign the workflow instead of bolting AI onto a broken one. Govern from day one, so trust and safety are built in rather than patched on later. And hold every effort to a return you would be willing to defend out loud. The Edge Framework, the department playbooks, the tool tables, and the build it yourself panels are all just structure around that one discipline.
The cost of waiting is the quiet part worth repeating. Every quarter spent debating is a quarter a competitor spends compounding. The gap between the companies that move and the companies that stall does not hold steady. It widens, month after month, as one side reinvests its freed hours and the other keeps paying for the same friction.
You do not need a perfect plan. You need a first workflow and the will to measure it.
So close the book and pick the workflow. One team, one task, one number to beat. Put a tool on it this week, govern it properly, and watch what your people do with the hours they get back. Then turn to the next department and do it again. That is how AI goes from a headline to a habit, and that is how it finally pays.
Rocky Hill
Industries and their top AI tools
Ten industries, and the three tools that tend to move the needle fastest in each. This is a starting shortlist, not a ranking. Confirm fit, pricing, and data requirements before you commit, and read it next to the Buy / Build / Assemble page.
Field Service (HVAC & Refrigeration)
Faster quotes, smarter dispatch, and fewer repeat truck rolls by putting service history and pricing at the technician's fingertips.
Construction
Catch schedule slip and rework early with vision based progress tracking and AI over your project record.
Retail & Ecommerce
Personalize at scale and turn product and customer data into copy, merchandising, and timely offers.
Financial Services
Compress research, surface answers from filings, and flag fraud in real time.
Healthcare
Give clinicians their time back with ambient documentation and faster triage.
Real Estate
Respond to every lead instantly and focus agents on the deals most likely to close.
Manufacturing
Stop unplanned downtime and catch defects before they ship.
Logistics & Route Services
See the whole network, route smarter, and coach drivers with context, not guesswork.
Legal & Professional Services
Cut hours of review and drafting while keeping a human accountable for the work product.
Hospitality
Answer guests instantly and price rooms to demand without a revenue team working overnight.
Tool names reflect each vendor's stated capabilities. Availability, pricing, and best fit vary by company size and data maturity; treat this as a research starting point, not an endorsement.
How the in-house tools get built
Every department ends with a "build it yourself" panel. They all draw on the same small set of building blocks. Learn these once and the patterns repeat across marketing, finance, support, and the rest.
1. Retrieval augmented generation (RAG)
A general model knows the public internet, not your pricing, contracts, or last week's tickets. RAG closes that gap. You split your documents into passages, turn each into a numeric fingerprint called an embedding, and store them in a vector database. At question time you embed the question, pull the closest passages, and hand them to the model as context so it answers from your material instead of guessing. This one pattern powers most internal assistants in this book. The discipline that makes or breaks it is chunking and source citation: every answer should point back to the passage it came from.
2. Agents that take steps, not just talk
An agent is a model given a goal, a set of tools it may call, and permission to loop until the job is done. Tools are just functions you expose: look up an order, draft an email, file a ticket. Agents move you from "draft me a reply" to "resolve this end to end." They are also where governance matters most, because an agent that can act can act wrongly at speed. Start read only, earn write access, and gate any action that spends money or sends external messages behind a human.
3. The Model Context Protocol (MCP)
Connecting every model to every system used to mean custom glue for each pair. MCP is an open standard that defines one consistent way for AI applications to reach data and tools.8 Build an MCP connector for your CRM once, and any MCP aware assistant can use it through a single, auditable layer.
4. Automation platforms
Not every workflow needs custom code. Low code tools like Zapier, Make, n8n, and Power Automate now embed AI steps directly. A nightly report that pulls from your field service platform, summarizes it with a model, and emails a formatted digest can be built in an afternoon here. Use these to prototype and to handle the long tail; graduate to code only when volume or control demands it.
5. Evaluation and monitoring
A feature that works in a demo can quietly degrade in production. Before launch, build a test set of real inputs with known good answers and score every change against it. After launch, log inputs and outputs, track accuracy and escalation rates, watch cost per task, and sample outputs for review. This is the same discipline the Edge Framework enforces in Phases 05 and 06.
The order matters as much as the pieces. Buy to build literacy, assemble to prove value, build only where it creates a lasting edge, and add agents last. The Buy, Build, or Assemble chapter walks through that sequence in full.
Proving it paid off
A pilot nobody measured is a pilot nobody can defend. The rule here is to lock the baseline before work begins and measure throughout, so there are no arguments at budget review.
Set the baseline before you launch
You cannot claim time saved if you never measured the old way. Before any rollout, capture the current state of the workflow you are changing: how long the task takes, how many people touch it, the error or rework rate, and the volume per week. In the Edge Framework this is Phase 01, and it carries CFO sign-off so the number is not disputed later.
The four numbers that matter
Most AI business cases come down to four levers. Time saved per task times volume times loaded labor cost. Throughput, meaning more work handled without adding headcount. Quality, measured as fewer errors, lower rework, or higher resolution and conversion rates. And cost to serve, especially in support where a price per resolved contact compares directly against a human handled one. Pick the one or two that fit the workflow. Resist claiming all four.
Count the full cost, not just the license
Seat prices are the visible cost. The real cost also includes usage fees that scale with volume, build and integration time, data cleanup, training and change management, and the ongoing human review that keeps quality honest. The software license is often the smallest line in year one. Budget for adoption, because workflow redesign, not the tool, is what separates the high performers from everyone else.1
Prove the AI caused it
Use a holdout group or an A/B design so you can show the AI drove the improvement, not market conditions. This attribution step is what turns a plausible story into a defensible one, and it is standard in any disciplined ROI measurement.
The payback test
Before you scale, write down the date the initiative should pay for itself and the single metric that will prove it. If a support assistant is meant to deflect a third of tickets, set the target, instrument the deflection rate, and check it on the date. Initiatives that hit their number get more budget. Initiatives that miss get fixed or stopped.
For each AI project, keep a single page: the workflow, the baseline, the target metric, the all in cost, the payback date, and the named owner. If it will not fit on one page, the initiative is not clear enough to fund yet.
References and sources
Statistics, pricing, and regulatory points are drawn from the primary sources below. Pricing and product details reflect publicly available information as of the October 2026 revision and change often. Verify current figures with each vendor before you budget.
- J. Ryseff, B. F. De Bruhl, and S. J. Newberry, "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed," RAND Corporation, 2024. Based on interviews with 65 experienced data scientists and engineers, it estimates that more than 80% of AI projects fail, about twice the rate of IT projects without AI, and traces most failures to unclear problems, poor data, and weak infrastructure. rand.org
- McKinsey & Company, "The State of AI" global survey, on adoption breadth and workflow redesign as the strongest predictor of financial impact. mckinsey.com
- National Institute of Standards and Technology, "AI Risk Management Framework (AI RMF 1.0)," NIST AI 100-1, 2023. The Govern, Map, Measure, and Manage functions. nist.gov
- International Organization for Standardization, "ISO/IEC 42001:2023, Artificial intelligence management system," the first certifiable AI management system standard. iso.org
- European Union, "Regulation (EU) 2024/1689 (the AI Act)," a risk tiered framework with certain employment uses classed as high risk. digital-strategy.ec.europa.eu
- New York City, "Local Law 144 of 2021," requiring bias audits and candidate notice for automated employment decision tools. nyc.gov
- GitHub and Forrester Consulting, research on GitHub Copilot developer productivity, including task completion speed and a commissioned Total Economic Impact analysis. github.blog
- Anthropic, "Introducing the Model Context Protocol," 2024, with subsequent open source adoption. modelcontextprotocol.io
- Microsoft and Forrester Consulting, "The Total Economic Impact of Microsoft 365 Copilot," and Microsoft Work Trend Index research on time saved per user. microsoft.com
- World Economic Forum, "Future of Jobs Report 2025," January 2025, projecting 170 million jobs created and 92 million displaced by 2030, a net gain of 78 million, with 22% workforce churn. weforum.org
- International Monetary Fund, Cazzaniga and others, "Gen AI: Artificial Intelligence and the Future of Work," Staff Discussion Note SDN2024/001, 2024, on ~40% global and ~60% advanced economy exposure, with about half of exposed roles positioned for augmentation. imf.org
- Goldman Sachs Global Investment Research, "The Potentially Large Effects of Artificial Intelligence on Economic Growth," 2023, estimating AI could affect the equivalent of about 300 million full-time jobs. goldmansachs.com
- OWASP, "Top 10 for Large Language Model Applications (2025)," the community standard list of the most critical LLM application security risks. genai.owasp.org
- MITRE ATLAS, "Adversarial Threat Landscape for Artificial Intelligence Systems," a knowledge base of real world adversary tactics and techniques against AI. atlas.mitre.org
- Alan M. Turing, "Computing Machinery and Intelligence," Mind, 1950, which posed the question of machine intelligence and the imitation game. Mind, Oxford
- J. McCarthy, M. Minsky, N. Rochester, and C. Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence," 1955, the 1956 workshop that named the field. dartmouth.edu
- Association for Computing Machinery, "2018 ACM A.M. Turing Award" to Yoshua Bengio, Geoffrey Hinton, and Yann LeCun for breakthroughs in deep learning. acm.org
- A. Vaswani and others, "Attention Is All You Need," 2017, the transformer architecture underlying modern large language models. arxiv.org/abs/1706.03762
- Boston Consulting Group, "The Widening AI Value Gap: Build for the Future 2025," September 2025, a survey of 1,250 companies finding 60% achieving no material value from AI and 5% achieving value at scale, plus the 10-20-70 rule. bcg.com
- IBM Institute for Business Value, "2025 CEO Study," May 2025, a survey of 2,000 CEOs reporting that 25% of AI initiatives delivered expected ROI and 16% scaled enterprise wide. newsroom.ibm.com
- MIT NANDA, "The GenAI Divide: State of AI in Business 2025," on the learning gap behind stalled generative AI pilots; its 95% headline figure has been widely disputed. finance.yahoo.com
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," June 2025. gartner.com
- DXC Technology, "DXC Launches AdvisoryX and Releases New Global Insights on the AI Execution Gap," December 2025. prnewswire.com
- Deloitte AI Institute, "The State of AI in the Enterprise: 2026 Report," on the skills gap as the top integration barrier and the maturity of agentic AI governance. deloitte.com
This field guide is original work. Where it draws on outside research, regulation, or vendor documentation, those sources are cited above and summarized in my own words. Tool descriptions reflect each vendor's stated capabilities; pricing is indicative. Nothing here is legal advice.