Finding Strategic AI Opportunities

In July 2025 a team at MIT’s Project NANDA published a preliminary study of how companies were spending on generative AI. When the researchers asked executives to divide a hypothetical budget across business functions, about half of it went to sales and marketing. That figure is a stated preference, not measured spending. Yet some of the most dramatic savings the team documented came from the back office: finance, procurement and operations, often by replacing outsourced processing and external agencies. The authors’ explanation was blunt. The tilt toward the front office reflected “easier metric attribution, not actual value”¹.

The study is small and early, built on 52 organization interviews and 153 survey responses, and its headline claim that 95 percent of organizations were getting no return has been widely questioned. We know of no larger, independent study that tests its specific claim, that AI budgets follow ease of measurement rather than value, so the NANDA result is best read as a hypothesis, not a finding. It is still a hypothesis worth testing, because the mechanism behind it is familiar. Money tends to flow to the AI ideas whose results are easiest to show on a slide, and nothing guarantees that this is where the value sits.

Finding strategic AI opportunities is the discipline of looking in the second place.

The core idea

A strategic AI opportunity is found by walking down from the strategy, not up from the technology. You start with what the business must achieve, find where value is created, captured or lost, locate the capability and the bottleneck that hold the result back, and only then ask what AI could do there.

The aim is a short list of named opportunities, each one sized roughly and tested against the alternatives, not a long list of possibilities. Richard Rumelt’s observation that good strategy concentrates effort, rather than spreading it across every worthy goal, applies with full force here².

Many organizations run the funnel upside down. They start from a demonstration, a product or a capability, and search the business for somewhere to put it. Think of an oil company that drills wherever its rig happens to be parked. It sinks many shallow holes, produces a great deal of activity and finds little. The disciplined company surveys first: it reads the terrain, finds the fault line where pressure collects and then drills one deep well. Vendors arrive with rigs, and good ones. Only you can survey the ground, because only you know where your business makes and loses money.

Opportunity, use case, solution

Three words are often mixed up in steering committees, and separating them is a useful habit.

An opportunity is a meaningful business improvement that AI may help deliver, stated in business terms. Take an illustrative equipment-rental company, the example this chapter follows from here on. Its opportunity might be: cut the cash and staff time lost to disputed invoices. No technology is named. A use case is one specific way to pursue it, such as matching delivery and collection records to each invoice before it goes out. One opportunity usually has several use cases, and some of them need no AI at all: a mandatory checklist at delivery might deliver much of the gain. A solution is how you obtain the use case, whether you build it, buy it or partner for it. That question comes last.

The order matters because each level constrains the next. A conversation that starts at the solution has already decided the opportunity without discussing it. A conversation that starts at the opportunity can still reach the same product, but it reaches it for a reason, and it can see the cheaper alternatives on the way.

Find where value is created, captured or lost

Before anyone names an AI idea, map where value lives. In 1998 Orit Gadiesh and James Gilbert of Bain described an industry’s profit pool, the total profit earned at each step of its value chain, and argued that the pattern of profit is often very different from the pattern of revenue³. Their best-known example was the US truck-rental market. Renting trucks generated most of the industry’s revenue but little of its profit, because customers shopped hard for the lowest daily rate. The accessories that came with the rental, such as boxes, insurance, trailers and storage, earned a large share of the profit. U-Haul priced rentals to win volume and made its money on the rest, earning an operating margin of about 10 percent when the industry averaged less than 3.

The lesson for AI is direct. The part of the business with the most visible activity is not necessarily where an improvement is worth most. Walking the activities of the value chain one by one, as Michael Porter proposed, remains a dependable way to see where value is made⁴. For AI opportunity work, four pools cover most of it.

The pool that matters most is set by the strategy, not by the technology. A business that competes on price should look hardest at the cost pool. One that competes on service should look at the customer pool. Inside the chosen pool, the question becomes where value leaks away: rework, write-offs, slow decisions, scarce experts doing routine work, outsourced processing nobody has examined in years. If the NANDA hypothesis holds, this is where it bites: back-office leakage is real money, but it is quiet, so it is easy to underfund. Testing that in your own budget costs little: list where AI money went last year, and where the largest leaks sit, and see whether the two lists overlap.

Five lenses for finding candidates

Once the pool is chosen, a small set of questions generates candidates without starting from a vendor’s catalog. Each lens points at a different way AI can change the economics of the work.

The decisions lens deserves particular attention. Ajay Agrawal, Joshua Gans and Avi Goldfarb argue that the simplest way to understand AI’s economics is as a fall in the cost of prediction. When prediction becomes cheap, decisions that were once made by rule of thumb, or not made at all, become worth making well, and the value of the human judgment that acts on the prediction rises⁵. The useful question is therefore not “where could we use a forecast?” but “which decision would we make differently if we knew more?” A better prediction that changes no decision is worth nothing.

The competition lens asks a larger question. Marco Iansiti and Karim Lakhani argue, from cases such as Ant Financial, that firms built around data and AI can escape the traditional limits on scale, scope and learning that bind conventional operating models⁶. That is shown in a handful of digital firms, not yet across most industries. An opportunity seen through this lens is less about saving cost than about where a rival could learn faster than you. The lenses are a workshop tool, not a second framework. Use them to fill the funnel, then let the value pools and the strategy decide what stays.

Size the pool, not the return

Each surviving candidate needs a rough size before it competes for money. At this stage the honest answer is a range, built from a few plain numbers: how many times the work happens, what each occurrence costs or earns, and how much of that AI could plausibly change.

Return to the equipment-rental company, and suppose it sends 200,000 invoices a year, and 8 percent of them are disputed. That is 16,000 disputes. If each one costs about 150 in staff time and delayed cash, the pool is 2.4 million a year. Now narrow it. Perhaps half the disputes come from missing delivery paperwork, the cause AI could plausibly address by matching documents before the invoice goes out. That leaves 1.2 million addressable. If the company then captures 60 percent of that once adoption and exceptions are allowed for, the realistic prize is about 720,000 a year.

Two things make this useful. First, it separates a big opportunity from a small one quickly, and order of magnitude is what matters when comparing candidates. Second, it keeps the pool and the prize apart. The pool is what is at stake. The prize is a hypothesis about how much of it you will capture. Treating the first as the second is how weak ideas get funded with false precision.

AI advantage: beat the best alternative, not the status quo

A large pool does not mean AI is the right tool. The test is whether AI is materially better than the credible alternatives: a fixed rule, conventional automation, a process redesign, better data, simple analytics, outsourcing or a hire. The test asks how much of the gain AI adds on top of the best non-AI option.

In the rental example, a mandatory delivery checklist might remove most of the paperwork disputes with no AI at all. AI matching on top removes some more. The AI advantage is the second gap, not the whole distance from today. That smaller number is the one that should be weighed against AI’s cost and risk. Sometimes it is still large, because the problem involves language, documents or prediction that rules cannot handle. Sometimes it is close to zero, and the opportunity is real but the answer is a process change. Either way, the opportunity was worth finding. AI’s edge is also uneven across tasks that look alike, which is another reason to measure it for each opportunity rather than assume it.

Six tests, read as one row

Potential is not enough. Each candidate faces six tests together: strategic fit, business value, AI advantage, feasibility, risk and time to value.

Here is why the scores fall where they do, in the illustrative case of the rental company. Dispute matching serves the strategy’s cash and margin goals, the pool is 2.4 million, the data already exists in delivery and billing systems, and a wrong match is caught before the invoice goes out, so the risk is low. Its AI edge is only medium, because the checklist takes most of the gain; that is exactly what the advantage test found, and the row records it honestly. The marketing copy generator is easy and fast, but it touches no value pool the strategy depends on. Feasible is not the same as important. Autonomous pricing, in which the system sets and changes rental rates by itself, looks attractive on fit and value, but it needs clean demand data the company does not yet have, a wrong price reaches every customer at once, and it would take long to pay back. It becomes a later candidate once the foundations exist.

Reading the whole row guards against two common mistakes: funding whatever is most feasible, and funding whatever has the biggest number. A score structures the debate; it does not settle it.

Write the opportunity in one sentence

An opportunity that survives the six tests should fit in one sentence that anyone on the executive team can check.

For the rental company, an illustrative statement might read: to protect cash and margin, we will cut the cost of disputed invoices by relieving missing delivery paperwork, where AI will match delivery and collection records to each invoice before it goes out, because it catches the mismatches a checklist alone misses. The sentence is not a business case and not a design. It is the test of whether the thinking is finished. Notice that the sentence carries the advantage test with it: anyone reading it can ask how many disputes the checklist alone would remove, and how many more AI would.

Story: a ship-repair yard runs the search twice

The following is an illustrative composite, drawn from a pattern common in project businesses rather than from a single company. Picture a ship-repair yard that bids for repair and refit work on cargo ships and ferries.

Its leadership decides it needs an AI program and asks every team for ideas. Seventy arrive. A committee ranks them by how easy they are to deliver and funds the eight most feasible: meeting summaries, proposal formatting, a marketing copy tool, an intranet chatbot and four more like them. A year later the yard has eight pilots and a good slide about them. Its margins have not moved, and nobody can say which strategic priority any pilot served, because nobody asked.

The second round starts at the top of the funnel. The yard’s strategy is to win more fixed-price repair and refit jobs profitably. The value pool is margin lost on bids priced too low. Suppose the yard wins 150 fixed-price jobs a year at an average of 1.2 million each, and one in five overruns its budget by 15 percent. That is 30 jobs losing about 180,000 each, a pool of roughly 5.4 million a year. Much of the overrun is extra work found only once a hull is opened, such as corroded steel to renew. The capability is bid estimating, and the bottleneck is specific: experienced estimators spend days searching old survey reports and job records for comparable vessels and for the surprises that hurt the yard last time.

Now the opportunity is clear: make past job knowledge available at bid time. The AI role is to read and retrieve. The use case is a bid-review assistant that finds comparable jobs and flags likely extra work in the owner’s specification. The AI assembles the evidence; an estimator still judges it and sets the price. The yard also tests the alternative. A better filing system would help, but it would not read thousands of past reports, so AI keeps a real advantage. The measure the yard agrees is fewer loss-making jobs, tracked against the overrun rate before the change.

The first round was not foolish. Every idea was reasonable. The list simply never touched the place where the yard wins or loses money.

What this means for leaders

Strategic AI opportunities rarely announce themselves. They sit in quiet places: a write-off line, a dispute queue, an expert’s inbox. Leaders find them by insisting on the order of the funnel, by asking where value leaks before asking what AI can do, and by making every proposal show its chain from priority to value. Two numbers deserve skepticism whenever they appear: a pool presented as a prize, and an AI gain measured from today rather than from the best alternative.

Check yourself

1. A comprehensive list of AI use cases is a good first draft of an AI strategy.MythFact
2. An opportunity should be stated without naming any technology.MythFact
3. The functions where AI results are easiest to measure are usually where the most value lies.MythFact
4. AI’s advantage should be measured against today’s process.MythFact
5. A large opportunity pool tells you how much value the organization will capture.MythFact
6. A better prediction that changes no decision has no business value.MythFact

Reflection: where your value leaks

What comes next

This chapter sits between others in the course. It takes its starting point, the constraint and the capability map, from Start With Business Strategy, and its sense of how far AI should change the business from Defining AI Ambition. What it hands on goes to several places: measuring realized value to From AI Capability to Business Outcome, the full cost side to Module 08, the controls a risky opportunity needs to Module 07, the one-page design of a use case to AI Use-Case Discovery and Design, weighing many candidates against capacity to Prioritizing the AI Portfolio, and capabilities that several teams need to AI Platform Strategy.

The nearest question comes first. Once an opportunity is named, sized and tested, who should provide the capability? Some capabilities are worth owning, some are better bought, and some are best built with a partner. The next chapter, Build vs Buy vs Partner, sets out how to decide.