Prioritizing the AI Portfolio

Suppose you are the chief operating officer, and next year’s AI proposals have arrived. There are 38 of them. Every one has a sponsor, most have a business case, and a few already have a vendor waiting. Your data, integration and security teams can properly support perhaps six at a time. Whatever you choose, 32 sponsors will think you chose wrong.

The tempting answer is to fund a little of everything and let the winners emerge. Recent evidence shows how often such starts end before they deliver. S&P Global Market Intelligence’s 2025 survey of more than 1,000 enterprises in North America and Europe, as reported by the trade publication CIO Dive, found that 42 percent of companies had abandoned most of their AI initiatives, up from 17 percent in the previous year’s survey. A second figure counts something different. The 42 percent is a share of companies; the other is a share of projects inside them: the average organization had scrapped 46 percent of its proofs of concept before they reached production¹.

Stopping work is often the right decision, and a portfolio that never stops anything is not being managed. The expensive pattern is different: too many starts, too little capacity to finish them, and no agreed point at which a weak idea would be dropped. The cure is not a better list of ideas. It is a better way of choosing among them.

A portfolio is a capital-allocation decision

An AI portfolio is the full set of AI experiments, products, automations and shared capabilities that the enterprise funds, managed together against one budget and one pool of people. Prioritizing it means choosing the combination of initiatives that best advances the strategy within those limits, and then revisiting that choice as evidence arrives.

Two words in that definition do most of the work. Combination means the initiatives are not independent. They compete for the same engineers, depend on the same data and carry risks that add up. A list ranked one idea at a time cannot see any of that. Revisiting means a portfolio decision is never final. Every initiative is funded only to its next decision point, where it is funded further, deferred or stopped. As The AI Maturity Model showed, managing AI this way, with initiatives started, scaled and stopped on purpose, is one of the marks of the scaled level.

The discipline is older than AI. In a study of new-product portfolios, Robert Cooper, Scott Edgett and Elko Kleinschmidt found that financial methods were the most popular way to select projects and produced the worst portfolios, while the best performers applied a formal, explicit method consistently to every project².

The starting point is strategy. Name the three to five enterprise outcomes that AI is meant to serve, and require each proposal to say which one it advances. As Start With Business Strategy argued, an idea that serves no stated priority is not a candidate, however impressive the demonstration. Whether one designed use case is ready to test is a separate question, answered by the gates in AI Use-Case Discovery and Design. This chapter starts where those gates stop: with a set of credible candidates and not enough capacity for all of them.

Value against feasibility: the two-by-two

A useful first picture of a portfolio has two axes. Value is the strategic and economic prize if the initiative works: revenue, cost, risk reduction, customer or employee benefit, stated as a hypothesis with a measure. How to estimate and discount that prize is the subject of Building the AI Business Case. Feasibility is how close the organization is to delivering it: data, integration, skills, process change and the controls the risk requires. The readiness assessment from Assessing Enterprise AI Readiness is the evidence for this axis.

Each quadrant implies a different decision. High value with high feasibility is funded and sequenced within capacity. High value with low feasibility is not rejected; it is staged. The organization funds the readiness work or a small test that retires the biggest uncertainty, and holds back the full program until the path exists. Low value with high feasibility is done cheaply, usually with a standard product, without spending scarce engineering on it. Low value with low feasibility is stopped or parked, and revisited only if the facts change. The next chapter turns these quadrants into distinct kinds of investment, each with its own funding logic.

Two cautions keep the matrix honest. First, each position is an estimate, so place proposals as ranges rather than points, and treat a proposal whose range spans two quadrants as a question to resolve, not a placement. Second, the matrix judges each initiative on its own. It cannot see that six proposals need the same data, or that the team can carry only four. Those are the next questions.

Risk belongs on the feasibility axis, because a higher-risk use needs stronger evidence and more controls before it can run. Sometimes the law sets that bar.

A score is structure, not a verdict

Many organizations go one step further and score each proposal: weights for value, strategic fit, feasibility, risk and time to value, summed into a number. Scoring is worth doing. It forces every proposal to answer the same questions, and it makes disagreements visible. It also invites a particular mistake: treating the number as the decision.

Three habits keep scoring useful. Choose the weights deliberately for your own strategy rather than importing someone else’s, and say them out loud, because the weights are where the strategy hides. Record confidence beside every score: a high value estimate resting on a vendor’s claim is not the same as one resting on a measured pilot, as AI Value Scorecard showed. High value with low confidence usually means a staged test, not a full program. And when two proposals score close together, stop reading the decimals. Break the tie on dependencies, time to value, strategic differentiation and the balance of the whole portfolio.

The Cooper finding is the warning against one particular scoring habit. A ranking by return alone favors small, certain projects and starves the work that is larger, slower and often more important, including the shared capabilities that make other initiatives possible².

Capacity is the real budget

Money is often not what limits an AI portfolio. People are: data engineers, integration specialists, security and risk reviewers, change managers, and the time of the business owners who must redesign their processes. These teams are shared, so every initiative added to the active list slows every other one.

Queueing theory makes the cost precise. Little’s law states that, in a stable system, the average amount of work in progress equals the rate at which work is completed multiplied by the average time each item spends in the system⁵. Rearranged, average time to finish equals work in progress divided by throughput.

Suppose, as an illustration, that the shared team finishes about four initiatives a year. With twelve running at once, each takes three years on average. With four running, each takes about one. The team delivers the same number of finished initiatives either way, but with four in progress, value and learning arrive two years sooner, and a failing idea is discovered while it is still cheap. The arithmetic even flatters the crowded portfolio, because it assumes a team works as quickly on twelve things as on four.

The practical rule follows. Set an explicit limit on the number of initiatives in active build for each shared team, and make a new start wait for a finish or a stop. Twenty priorities with capacity for five is not a portfolio. It is a queue.

Dependencies and enablers

Some initiatives cannot start until something else exists: clean master data, an integration layer, an evaluation service, a security review route. Those enablers rarely have a convincing standalone business case, which is why a return-ranked list pushes them to the bottom, and why the initiatives above them then stall.

Value an enabler by what it unlocks: the number and importance of funded initiatives that need it, the duplicated work it avoids, and the time it saves each of them. Then sequence it ahead of the initiatives that depend on it. The opposite error is just as real. As AI Platform Strategy argued, a shared capability built ahead of demand becomes a cost looking for a use. A workable test is to fund an enabler when at least two or three initiatives that are already funded need it, and to size it to their needs rather than to every possible future one.

Stage the money, write the stop rule first

Uncertain initiatives should not receive their full budget up front. In 1995 Rita McGrath and Ian MacMillan proposed discovery-driven planning for ventures whose outcomes cannot be forecast with confidence: write down the assumptions the plan depends on, test the most important ones first, and release funding as milestones are met⁶. Robert Cooper’s stage-gate model, built for new products, makes the same point structurally. Work proceeds in stages separated by gates, and at each gate management decides go or kill and commits resources only for the next stage⁷.

A small, early stage has option value: it buys the information that decides whether a much larger investment should follow. That is a reason to start small, not a reason to skip evidence. Each stage needs a learning objective and, before it starts, a stop rule: the quality, adoption, unit cost or risk level at which the initiative will be redesigned or ended. How a single system is taken through production and scale is the subject of From AI Pilot to Production to Scale; the portfolio question is only whether it has earned the next tranche. Gate outcomes and portfolio decisions are different things: a gate judges one initiative’s evidence at a stage boundary (proceed, redesign, defer or stop), while the four decisions below place every initiative in the portfolio at the quarterly review.

The stop rule must be written first because people are poor judges of their own failing projects. In a classic experiment, Barry Staw found that people who were personally responsible for an earlier investment that had gone badly committed more additional money to it than people who had not made the original choice⁸. He called it escalating commitment. Written thresholds, agreed before the money moves and checked by someone other than the sponsor, are the practical defense.

Four decisions, revisited every quarter

The output of prioritization is not a ranked list. It is a decision for each initiative, in a small vocabulary.

Stop and defer matter as much as fund. Every stopped initiative releases people for a better one. A deferral with a named trigger, such as “when the customer data is clean”, keeps a good idea alive without letting it occupy capacity. For each initiative, the decision record can be short: why this, why now, what value, what evidence, what cost, what risk, which dependencies, who owns it, and when the next gate falls. Review the whole portfolio on a fixed cadence, quarterly for most organizations, because evidence, costs and strategy all move.

Story: Johnson & Johnson narrows 900 use cases

Johnson & Johnson began its generative AI program the way many large companies did: broadly. In March 2025 its chief information officer, Jim Swanson, said that over the previous two years much of the company’s generative AI work had been proofs of concept and pilots, aimed at learning, and that a group had been set up to review nearly 900 generative AI use cases across the organization for ethical, regulatory and compliance risk⁹. Employees had been encouraged to experiment, and a central governing board oversaw the work¹⁰.

Then the company measured. As the Wall Street Journal reported in April 2025, J&J found that 10 to 15 percent of its use cases delivered about 80 percent of the value. It changed course. Responsibility for AI projects moved from the central board to the business functions, such as commercial, supply chain and research, which were to prioritize and scale the highest-value work, and pilots that were redundant or not delivering were reported to have been shut down¹⁰. Projects were judged on three things: how readily they could be implemented, how useful they would be across the company and how much they would benefit the business¹¹. The short list that survived included a copilot that coaches sales representatives on their conversations with health-care professionals, an internal assistant that answers employees’ questions about policies and benefits, tools for drug discovery and a system to anticipate supply-chain risks¹⁰.

Read as a portfolio decision, the case shows several of this chapter’s ideas at once. Broad exploration was not in itself the error: Swanson described the early phase as a deliberate period of learning, and a small early stage can buy information, as staged funding argues. What turned the exploration into a portfolio was its end, with value measured across the whole set and a decision taken for each item. Value was concentrated, as it often is, and no organization’s specialists can carry 900 initiatives toward production, while a short list can be finished. The three reported criteria map onto the value-feasibility matrix: ease of implementation is feasibility, business benefit is value, and usefulness across the company is the reuse that makes shared work worth funding. And the choice moved to the functions that hold the knowledge, the people and the processes that must change.

One caution keeps the case honest. It comes from the company’s own account, largely through the press. The public record does not show J&J’s stop rules, how value was measured, or what the shut-down pilots cost. The lesson for a leader is the shape of the decision, not the figure of 80 percent: measure the whole portfolio, fund what can be finished, and stop the rest on purpose.

What this means for leaders

Treat AI prioritization as capital allocation. The question is not which ideas are best, but which combination this organization can finish, given its strategy, its people and its appetite for risk. Start from the outcomes the strategy names, place each candidate on value and feasibility with honest ranges, and use scores to structure the debate rather than to end it.

Then make the constraints visible. State the capacity of every shared team and hold the active list to it. Find the enablers that many initiatives need and fund them first, but only when real demand exists. Stage the money behind evidence, write the stop rules before the money moves, and give someone other than the sponsor the job of checking them. Finally, judge the portfolio by what it finishes and what it stops, not by how many initiatives it has started.

Check yourself

1. The initiative with the highest projected ROI should always be funded first.MythFact
2. Running more AI initiatives at once gets more value delivered sooner.MythFact
3. A high-value idea with low feasibility should be rejected.MythFact
4. A shared data enabler can deserve funding without a standalone business case.MythFact
5. Stop rules work best when they are agreed before an initiative starts.MythFact
6. A careful weighted score makes the portfolio decision objective.MythFact

Reflection: test your own portfolio

What comes next

Choosing the initiatives answers only part of the question. The funded set will still contain very different kinds of investment: some meant to return value within months, some meant to test a strategic possibility, and some meant to change how a whole part of the business runs. Each needs its own expectations and its own way of being judged. The next chapter, Quick Wins, Strategic Bets and Transformation Initiatives, shows how to balance that mix on purpose.