Most organisations do not have too few AI ideas. They have too many initiatives that nobody is prepared to stop.
The list usually looks healthy: customer-service pilots, internal assistants, document automation, agent experiments and a platform programme underneath them all. Each item has a plausible sponsor. Each promises time, quality or growth. Together, however, they compete for the same data, technical attention, change capacity and trust.
That is where a use-case backlog stops being evidence of momentum. It becomes a substitute for strategy.
An AI portfolio is not a ranked list of possibilities. It is a recurring leadership decision about where the organisation will learn, where it will commit and where it will deliberately withdraw.
More demand than capacity is useful information
When every initiative appears important, the organisation has not yet made its value logic explicit.
The problem is rarely a lack of scoring criteria. Teams can produce weighted matrices for value, feasibility and risk. The harder question is what the organisation is trying to become better at. Faster service? Better commercial decisions? More reliable operations? A new product capability?
Without that direction, scoring rewards whichever proposal arrives with the best presentation. It does not create a coherent portfolio.
Start with three portfolio intentions:
- Explore: Which uncertainty is worth reducing before we make a larger commitment?
- Exploit: Which proven pattern deserves investment, integration and adoption capacity?
- Enable: Which shared foundation will make several valuable workflows easier to build and govern?
Every initiative should have one primary intention. If it claims all three, it is probably still too vague.
Make start, stop and scale equally legitimate
Many leadership teams have an intake process. Far fewer have a serious exit process.
Starting is visible and optimistic. Stopping can feel like admitting failure. Scaling is often assumed to be the natural reward for a successful demo. None of those instincts is good portfolio management.
Use three different decision bars:
- Start when the problem matters, the learning question is explicit and the experiment is bounded enough to produce evidence quickly.
- Stop when the underlying need is weak, the data or operating conditions make the idea uneconomic, or the initiative is no longer reducing meaningful uncertainty.
- Scale only when value, repeatability, ownership and guardrails have survived contact with real work.
A technically successful pilot may still deserve to stop. A failed experiment may deserve to continue if it revealed a valuable constraint early. The decision should follow the evidence, not the emotional status of the project.
Give every initiative a one-page decision contract
Before an initiative enters the portfolio, require one page with six fields:
- the business outcome and accountable owner;
- the workflow or decision being changed;
- the uncertainty the next stage must reduce;
- one leading value signal;
- one guardrail that must not deteriorate;
- the date and evidence required for the next start, stop or scale decision.
This is deliberately smaller than a business case. Early AI work contains uncertainty. Pretending otherwise creates precise financial projections around untested assumptions.
The one-page contract does something more useful: it makes the next decision visible before enthusiasm, sunk cost and stakeholder politics accumulate.
Run a portfolio rhythm, not an annual review
AI initiatives move too quickly for an annual investment cycle and too consequentially for an informal innovation meeting.
A useful rhythm is a short portfolio review every four to six weeks. Do not invite every project to present progress. Review only the evidence needed to make a decision:
- What changed in the workflow?
- What did we learn that we did not know before?
- Did the value signal move?
- Did the guardrail hold?
- What new dependency or risk appeared?
- Is the next investment still justified?
The output is not a status colour. It is a decision, an owner and a date.
This rhythm also exposes a different kind of bottleneck. If several promising initiatives wait for the same identity pattern, data product or evaluation capability, the portfolio has discovered an enabling investment. That is more valuable than treating each delay as a local project problem.
The strongest objection: “We cannot measure innovation this early”
You often cannot calculate durable ROI after a few weeks. You can still measure whether uncertainty is falling.
Can users complete a meaningful task? Does the workflow produce less rework? Can the team describe the exceptions? Is the required data accessible with acceptable risk? Would the accountable owner choose to continue with their own budget and change capacity?
Early measurement should not pretend to prove the final business case. It should produce enough evidence for the next responsible commitment.
The portfolio test
Take the ten most visible AI initiatives in your organisation. For each one, ask:
What exact evidence would make us stop this within the next six weeks?
If the answer is “none,” you do not have ten experiments. You have ten promises.
A strong AI portfolio is not the one with the most activity. It is the one that converts uncertainty into decisions, releases capacity when evidence weakens and concentrates investment where a repeatable capability is actually emerging.
Deutsche Ausgabe: Die AI-Portfolioentscheidung: Was Führung starten, stoppen und skalieren muss
Sources and framing
- Microsoft 2025 Work Trend Index: the emergence of the Frontier Firm
- NIST AI Risk Management Framework
- Microsoft guidance: manage the agent lifecycle
Editorial note: This is an independent operating framework. It does not replace financial, legal or risk assessment for a specific investment.
