A List of AI Projects Is Not a Portfolio

An executive committee reviews a long list of AI initiatives. Every project has a sponsor. Most are green or amber. The teams can explain what they have delivered and what comes next.

Then leadership asks three questions:

• Which initiative deserves the next investment?

• Which should stop?

• Which is genuinely ready to scale?

The room has no defensible answer.

I have sat through versions of this meeting inside global organisations: different priorities, different slide templates, the same pause when activity had to become a decision.

That is not an AI portfolio. It is an inventory of activity.

The distinction is becoming critical. AI investment is accelerating while enterprise financial impact remains uneven. McKinsey’s August 2026 global survey found that nearly nine in ten respondents report regular AI use in at least one business function, yet only 37% attribute any EBIT impact to AI and just 6% qualify as high performers. At the same time, 60% expect their organisations to increase AI investment over the next year. McKinsey, The State of AI in 2026

BCG reports the same tension from another angle: companies expected AI spending to rise from about 0.8% to 1.7% of revenue in 2026, while nearly three-quarters of CEOs identified themselves as the main AI decision maker. BCG AI Radar 2026

The leadership problem is no longer whether AI deserves attention. It is whether organisations can allocate capital across competing AI opportunities with discipline.

An inventory tracks work. A portfolio forces choices.

An inventory can tell leadership what exists: the business unit, use case, technology, sponsor, budget, stage, milestone and delivery status. That information is useful. It creates visibility and may expose duplication.

But visibility alone does not create portfolio management. A portfolio exists only when leaders can compare unlike investments and make consequential choices between them.

That means deciding whether to:

• stop an initiative whose evidence no longer supports the investment;

• continue a bounded test because a material uncertainty is still worth resolving;

• scale an initiative because the evidence and operating conditions justify greater commitment;

• redirect money, talent or executive attention from one initiative to another.

A steering meeting that reviews every initiative but changes no allocation may be governance theatre: orderly, informed and economically inert.

The four questions I use

To examine whether an AI initiative is ready for another investment decision, I use four questions.

1. What decision must be made now?

“Review progress” is not a decision. Neither is “continue monitoring.” The required call should be explicit: approve another eight-week test, release the next funding tranche, expand to two markets, pause deployment, or stop.

If no decision is due, the initiative may not belong on the steering agenda.

2. Who owns the measurable business outcome?

The delivery lead may own the model, platform or implementation. That is not automatically ownership of the business result. A named executive must be accountable for the outcome being pursued: reduced cycle time, lower loss, higher conversion, improved service quality, better working-capital performance or another defined result.

Shared ownership often means that nobody has the authority—or incentive—to make the trade-offs required for value.

3. What evidence justifies further investment?

Evidence should match the decision. Technical accuracy may justify continued experimentation. It rarely justifies enterprise scale on its own. A scale decision may require evidence of user adoption, workflow integration, realised unit economics, control effectiveness, data readiness and operational capacity.

Deloitte’s 2026 enterprise research reinforces the gap: AI access and production ambitions are rising, but only 34% of organisations report truly reimagining the business, and only one in five has a mature governance model for autonomous AI agents. Deloitte, The State of AI in the Enterprise 2026

The question is not simply, “Do we have evidence?” It is, “Is this the right evidence for the decision we are about to make?”

4. What would cause us to fund, stop or scale?

This should be answered before the next review—not improvised inside it. Predetermined thresholds reduce the tendency to protect sunk costs, reinterpret weak results or keep pilots alive because no one wants to be the person who ends them.

The missing step: every initiative must leave with a decision

The test is not complete when the four questions have been discussed. It is complete only when leadership makes and records one of three calls:

• FUND — authorise a bounded, time-limited test with explicit evidence gates to resolve a material uncertainty.

• STOP — end or pause investment because the evidence, relevance or risk no longer supports it.

• SCALE — expand a sufficiently evidenced initiative under controlled operating gates.

This is the point at which a governance conversation becomes a portfolio decision.

Running two initiatives through the test

Consider two hypothetical AI initiatives competing for the same funding.

Initiative A: customer-service assistant

The model performs well in a controlled test, but agents bypass it in live work because its answers arrive outside their normal workflow. No operational owner has committed to redesigning the process.

• Decision required now: approve or reject wider deployment.

• Outcome owner: not established.

• Evidence: strong model results, but weak workflow adoption and no demonstrated service outcome.

• Decision threshold: no agreed conditions for stopping or scaling.

Exit from the test: FUND—or stop.

Wider deployment should not be approved. If leadership believes the use case remains strategically relevant, it could fund one bounded workflow-adoption test with a named operational owner, a fixed time limit and explicit adoption and service thresholds. If those conditions cannot be created, the initiative should stop.

Initiative B: maintenance-risk prediction

Its technical performance is less impressive, but a plant leader owns the downtime outcome, technicians use the signal in an existing planning meeting, and an eight-week test shows a credible reduction in avoidable stoppages.

• Decision required now: approve or reject controlled expansion to another asset group.

• Outcome owner: plant leader accountable for downtime reduction.

• Evidence: operational adoption, early outcome evidence and integration into an existing workflow.

• Decision threshold: reliability, adoption and downtime thresholds can be defined for the next gate.

Exit from the test: SCALE—under controlled evidence gates.

A controlled expansion could be approved, but not an unlimited rollout. The next commitment should specify the asset group, funding, accountable owner, evidence thresholds and review date.

The test does not declare Initiative A a bad project or Initiative B a good one. It determines what investment decision each initiative is ready for now.

That is the practical difference between project status and portfolio discipline. Initiative A may deserve a deliberately bounded test to resolve one uncertainty. It does not deserve scale funding disguised as continued progress. Initiative B has stronger grounds to scale, but it still leaves with evidence gates—not a blank cheque.

Record the call

Each initiative should leave with a short decision record:

• Decision: fund, stop or scale.

• Rationale: the evidence that drove the call.

• Owner: the executive accountable for the resulting business outcome.

• Commitment: the capital, people, data or access being approved.

• Evidence gate: what must be true at the next review.

• Review point: when the decision will be revisited.

Without this record, the same arguments often return at the next meeting and initiatives drift forward by default.

Use the test carefully

The test is a decision filter, not a complete AI governance system. It does not replace technical assurance, data governance, cybersecurity, legal and responsible-AI controls, architecture standards or change management; nor does it pretend that regulatory, growth and foundational initiatives can be compared through one universal score. Its purpose is narrower: to expose whether leadership has a real decision, an accountable outcome owner, decision-grade evidence and an agreed threshold for the next call.

The uncomfortable research question

Many organisations can explain how an AI initiative gets approved. Far fewer can explain how one gets stopped.

This decision structure reflects what I learned from enterprise transformation portfolios. I am now testing where it holds for AI.

When was the last time your organisation stopped an AI initiative—and what evidence made stopping it possible?

If you have recently made—or reversed—one of these decisions and would compare notes confidentially, message me. I am examining where formal AI governance helps leaders make the call, and where it still fails them.

The answer may reveal more about the maturity of an AI portfolio than the number of pilots, models or users ever could.