The Enterprise AI Use-Case Landscape

Suppose it is Monday morning and you run operations. Four proposals are waiting. Marketing wants an AI content generator. Finance wants an AI anomaly detector. The warehouse team wants an AI route optimizer. Human resources wants a tool that summarizes interviews. Each arrives with a polished demo, a keen sponsor and a budget request.

Before approving any of them, notice what all four have in common. Each names a capability: generate, detect, optimize, summarize. None says whose work will change, which step of that work, or what number will move if it succeeds. They are descriptions of what a technology can do, not of what the business will do differently.

That gap is a common reason enterprise AI disappoints. When the Boston Consulting Group surveyed 1,000 senior executives in 2024, 74 percent said their companies had yet to show tangible value from AI. The small group that had moved beyond proofs of concept pursued, on average, only about half as many opportunities as everyone else, and scaled more than twice as many¹. The leaders were not the ones who saw the most possibilities. They were the ones who read the landscape well enough to choose.

Start with the work, then choose the capability

The conviction of this chapter is simple to state. Start with the business activity, decision or workflow. Then ask which AI capability, if any, can improve it.

In practice many organizations run the sequence backwards. Someone sees a capability in a demonstration, and the organization goes looking for places to use it. The result is a collection of tools in search of problems, and a progress report that counts pilots rather than results.

The evidence for running it the right way round is strong. In McKinsey’s March 2025 survey, the redesign of workflows had the biggest effect on whether an organization saw an impact on earnings from generative AI, out of 25 attributes tested. Yet only 21 percent of organizations using generative AI said they had fundamentally redesigned even some of their workflows². The value sits in changed work, and roughly four in five of those organizations had yet to fundamentally redesign any. As From AI Capability to Business Outcome argues in detail, a capability becomes valuable only when it causes a measurable change in how the business performs.

A use case is a capability placed in real work

A capability is what the technology can do: generate text, classify, predict, retrieve, recognize images, optimize. A use case is a specific application of a capability to improve a business activity, decision or workflow. Text generation is a capability. Drafting a reply to a customer that cuts the time to resolve a complaint is a use case.

The difference is easiest to see as a chain. Take an illustrative example: a field-service engineer who repairs equipment at customer sites.

The role is an engineer on a repair visit. The problem today is that manuals and records of past fixes sit in several systems, so the engineer often leaves without a fix and has to come back another day. The capability is retrieval: find similar past repairs and suggest the likely cause. The new workflow has the engineer confirm the diagnosis and make the repair, so a person still owns the fix. The outcome is more repairs completed on the first visit, measured against how the work runs today.

Hold the distinction grammatically. Capabilities are verbs: generate, predict, optimize. A use case is a whole sentence, with a subject who does the work, an object the work acts on and a result that can be checked. “Generate” on its own says nothing about the business. “The service planner drafts the delay notice before the customer calls, so complaints fall” is a sentence a leader can fund, staff and measure. If a team can only offer you verbs, it does not yet have a use case.

Five patterns cover most of the landscape

Executives do not need a catalog of fifty tools. They need a small number of categories that make any proposal legible. The best-known published set is older and smaller. In 2018, Thomas Davenport and Rajeev Ronanki studied 152 corporate AI projects and found they fell into three types: process automation (71 projects), cognitive insight from data (57) and cognitive engagement with customers and employees (24)³.

The five patterns this course uses are a teaching device that builds on those three types; they are not a separately researched taxonomy. Cognitive insight splits into understand, predict and decide, because executives fund them differently. Generative AI made create a category of its own. And act stretches process automation toward systems that carry out work with some independence. Four of the five already have measured results in real deployments. The fifth, act in that wider sense, is emerging.

Create produces content or artifacts: first drafts of reports, proposals, documentation and code. Understand extracts meaning from information: it summarizes a contract, classifies a complaint or answers a question from approved sources. One of the best-measured examples combines the two. When a software firm gave about 5,000 customer-support agents an assistant that read the conversation and suggested replies, the number of issues resolved per hour rose by 15 percent on average, with the largest gains among less experienced agents⁴.

Predict estimates what may happen: demand, churn, equipment failure, credit risk. It carries a trap. A prediction creates no value by itself; it pays only when it changes a decision someone actually makes. Economists Ajay Agrawal, Joshua Gans and Avi Goldfarb describe AI as a sharp fall in the cost of prediction, but prediction is one input to a decision, alongside human judgment about what the outcomes are worth⁵. AI in Prediction, Forecasting and Optimization, later in this module, takes that further.

Decide recommends or optimizes under constraints: the next best action, a delivery schedule, an allocation of stock or staff. Act carries out work or coordinates systems, with decision rights and controls that match the consequence of getting it wrong. At its fullest, act means agents, which have their own chapter at the end of this module. It is listed as a pattern because proposals of this kind are already arriving, not because results are in: act has the thinnest record of measured results in real deployments so far. Treat it as an emerging capability still being proven, not an outcome already shown.

The labels overlap, and that is fine. Their job is to make two questions routine. Which patterns does this problem really need? And is the team proposing the one it needs, or the one it already owns?

Same patterns everywhere; the context is what changes

The same five patterns appear in almost every function. What differs is not the species of AI but the setting it works in.

Finance tends to forecast, detect and reconcile. Sales and marketing tend to generate and predict. Operations tends to predict and optimize. Customer service and human resources lean on understanding and drafting. The table is an illustrative map of tendencies, not survey data or a prescription; any function can use any pattern.

What changes from one function to the next is the data available, the shape of the workflow, the harm if the AI is wrong and the measure that defines value. A draft that is slightly off costs little in marketing and a great deal in a regulatory filing. What stays the same is the method: the five patterns, read through the same chain from role to outcome. That is why the function chapters that follow can each be short. They apply one method to different settings.

Where the value concentrates

A landscape this wide invites the conclusion that AI belongs everywhere. The evidence points the other way: the potential value is uneven, and it clusters.

McKinsey’s 2023 estimate of the economic potential of generative AI examined 63 use cases across 16 business functions. About 75 percent of the estimated value fell in just four areas: customer operations, marketing and sales, software engineering, and research and development⁶. BCG’s 2024 survey, which covered AI of all kinds, found a similar tilt toward the core of the business.

In that survey, 62 percent of the value came from core functions, led by operations at 23 percent, sales and marketing at 20 percent and research and development at 13 percent. Support functions together accounted for 38 percent⁷. Both figures are estimates, one modeled and one reported by executives, so treat them as a compass rather than a forecast for your company. The direction is what matters: value concentrates where the business makes and delivers what it sells. How to size a value pool once you have found it was the work of Finding Strategic AI Opportunities.

The seams between functions

The second concentration is the one this chapter most wants you to see, because the usual way of drawing the map hides it. Some of the largest opportunities do not sit inside any one function at all. They sit on the seams between functions, in workflows such as order to cash, procure to pay or hire to productive, where work passes from one team to the next.

Three things make the seams valuable. First, errors compound there. Information that one team holds and the next team needs, such as a substitute agreed with a customer, is lost or arrives late at each handover, and the next team works on a stale picture. Second, each function improves the part it can see. A warehouse that picks faster and a fleet that drives fewer miles have both done their jobs, and the order can still arrive late or wrong. Third, the cost lands in a measure that no department reports as its own: customer credits, cash tied up in disputes, days from order to cash. Those can be the very numbers the board watches.

Three things also make the seams invisible. Budgets, dashboards and calls for AI ideas are drawn by department, so each list is complete on its own terms and blind to the handovers. The data for a seam sits in several systems owned by several teams. And nobody has the authority to change all the steps at once, so even a seam that everyone complains about rarely becomes anyone’s proposal.

The patterns that pay at a seam are usually not the ones that dominate department lists. Most of the value comes from moving the right information forward in time: predict a shortage while the customer is still on the line, decide a route with the delivery window as a constraint, understand why disputes arise so that the cause is fixed upstream. Drafting tools help inside a lane; they rarely close a gap between two lanes. That is also why McKinsey’s finding that workflow redesign matters most applies here with special force: a seam can be closed only by redesigning the workflow across it, and that needs one owner with authority over every step².

Finding the seams takes three questions, none of them technical. Follow one order, claim or hire from first contact to the end and count the handovers. Ask where work waits, or where two departments hold different numbers for the same thing. And ask which measure the board cares about that no department presents. Where the three answers meet, there is usually a seam worth an owner.

Three ways to misread the map

The landscape is easy to draw and easy to misread. Three misreadings recur.

Tool first starts from a product already bought and asks where to use it. The list that results describes the tool, not the business, and anything the tool cannot do never makes the list. Department only asks each function to nominate its own use cases. Every list is sensible, and the workflows that cross functions belong to nobody. One pattern fills the portfolio with whatever is easiest to try. Since generative AI arrived, that has usually meant drafting and summarizing, which are real gains, while the prediction and optimization problems that drive operating results go unexamined.

The diagnostic question is the same in every case. Not “can AI do this?” but “where can AI materially improve a business workflow or decision, and what measurable outcome will prove that it did?”

Breadth is not a portfolio

Reading the landscape well leads to fewer use cases, not more. BCG’s leaders pursued about half as many opportunities as their peers and scaled more than twice as many¹. Breadth of exploration is cheap; breadth of delivery is not, because every use case needs an owner, data, a changed workflow and a measured outcome.

Finding Strategic AI Opportunities already showed why the easiest ideas are a poor guide to the valuable ones. The opposite trap is moonshots only: Davenport and Ronanki’s best-known example, MD Anderson Cancer Center’s AI oncology advisor, cost more than 62 million dollars and was put on hold in 2017 without being used on patients³. What this chapter adds is where the middle of the portfolio should sit: on core workflows with clear owners, including at least one seam. Hold one rule: count the workflows AI has materially improved, not the use cases you have launched.

Story: the distributor that mapped by department

What follows is an illustrative composite, built from a pattern common in distribution businesses. A regional food distributor supplies restaurants and grocers from several depots, and most of what it moves is perishable. Its leadership team wanted to avoid the tool-first trap, so it took what looked like the disciplined route: each of four department heads would nominate one AI use case, own it, fund it and measure it.

Sales chose an assistant that drafted order confirmations and suggested substitutes when an item was out of stock. The warehouse chose a tool that optimized picking routes. Transport chose route optimization for the delivery fleet. Finance chose an assistant that read customer complaints and drafted credit notes. By month six, all four dashboards were green: more orders handled per sales representative, more picks per hour, fewer miles per drop and faster dispute resolution.

At month twelve the board asked about the measure it cared about most: credits paid to customers for short, wrong or spoiled deliveries. It had not moved. The post-mortem found why. Substitutes that sales had offered were not visible to the warehouse until picking, so some orders were picked as originally placed. Routes optimized for distance pushed some chilled deliveries later in the day. And the dispute assistant meant credits were paid faster, not that fewer were needed. Each tool had improved one department’s piece of a single workflow, from order to delivered as promised, and nobody owned that workflow. It had never appeared on any list, because the lists were drawn by department.

The restart, led by the chief operating officer, began from the workflow instead.

The company kept most of its tools. What changed was the unit of design. One owner held the whole order. The assistant in sales now predicted shortages while the customer was still on the line, so substitutes were agreed before picking began. Chilled time windows became a constraint in the routing tool rather than an afterthought. The dispute assistant’s most useful output became a weekly list of reasons, read by the owner, rather than a faster credit note. The lesson is not that departments should not own AI. It is that the landscape has to be read by workflow as well as by function, or the most expensive work stays invisible.

What this means for leaders

The landscape gives you a vocabulary, not a shopping list. Five patterns, applied in every function, read through one chain from role to outcome. Four practical consequences follow.

First, refuse capability-only proposals. Send back any proposal that names a tool but not a role, a workflow and a measure; it is not ready to be judged. Second, look at the mix. If almost every use case in your portfolio is a drafting tool, the problems that drive your operating results are probably not on the list. Third, look across the seams. Name the two or three workflows that cross functions and cost the most, and give each an owner before you ask anyone for use cases. Fourth, report workflows improved, not pilots launched, so that progress is measured where the value is.

Check yourself

1. Generative AI is the right answer for almost every enterprise use case.MythFact
2. Most of generative AI’s estimated potential value sits in a handful of functions.MythFact
3. Companies that get the most from AI pursue more use cases than their peers.MythFact
4. The same capability in a different workflow is a different use case.MythFact
5. Asking each department to choose its own AI use case will surface the biggest opportunities.MythFact
6. Workflow redesign was the attribute most strongly linked to earnings impact from generative AI.MythFact

Reflection: read your own map

What comes next

A map shows where AI can create value. It does not tell you which opportunities to pursue, or how to shape one so that it survives contact with real work. The next chapter, AI Use-Case Discovery and Design, turns the landscape into a method for finding, framing and testing the use cases worth building.