Prioritize AI opportunities by comparing specific pieces of work against the same business objective, the effort needed to make them usable and the evidence behind each proposal. Keep unresolved conditions visible. An attractive benefit estimate cannot answer whether the necessary records exist, whether people can check the output or whether the proposed use is permitted.
The immediate decision is often which opportunity deserves the next investigation or bounded trial. You can make that choice without pretending to know which idea will ultimately deliver the greatest return. The wider AI adoption strategy guide covers ownership and the path from trials to an enduring program.
Separate the delivery route from the priority
In a May 2025 Reddit discussion, a Reddit user working in IT at a large manufacturing company described a growing queue of generative AI requests. They were sorting requests into using existing tools, composing low-code solutions and building custom systems. The custom-build queue kept growing.
That account identifies a useful distinction. Sorting an idea by how it might be delivered helps route the work. It does not establish why that idea deserves attention before another. A custom build may address an important problem, but its position in that queue tells you little about the evidence for its value.
Before comparing proposals, name the decision you can actually make. Are you choosing what to investigate, what to trial or what to fund for wider use? A proposal ready for a small investigation should not be presented as ready for production. Nor should a potentially valuable idea disappear merely because it needs more investigation than another.
If a candidate is still described as “AI for service operations,” return to discovering use cases from everyday work. You need a task, inputs, useful output and a person who can check it before a comparison means much.
Put the evidence beside each judgment
Use the same questions for every candidate, with enough context to explain your answers. The table below is a discussion aid. It does not calculate an objectively best opportunity.
| Lens | Question to answer | Evidence to bring |
|---|---|---|
Business value | What would improve, for whom, and against which current problem? | Recent work, a baseline or a clearly labeled estimate |
Feasibility | Can the proposed approach produce a usable result with the available inputs and systems? | Representative records, technical checks and known dependencies |
Risk and conditions | What could go wrong, and what must be resolved before the proposed activity can proceed? | Relevant information-use decisions, failure consequences and a workable review or recovery method |
Adoption effort | What must people change, learn, maintain or check? | The performer’s and reviewer’s account of the work, including available capacity |
Evidence confidence | Which judgments are observed, estimated or still unknown? | The source, its limits and the unanswered question that could change the recommendation |
Microsoft's planning guidance for AI agents distinguishes business impact, technical feasibility and user desirability. Its setting is agent planning, but the distinction is useful more broadly: a technically possible solution can still be unwanted or impractical for the people doing the work.
Keep confidence attached to the other judgments. “High value” based on a sponsor's estimate and “high value” based on repeated observed errors are different claims. Writing “unknown” is more informative than assigning a middle score that makes them look comparable. Weak evidence is not proof of low value; it tells you what remains uncertain.
Separate conditions from tradeoffs. Some weaknesses can be weighed against a benefit. Others prevent the proposed activity from proceeding. If the necessary customer records are not permitted in the proposed tool, an appealing value score cannot grant that permission. Define a narrower permitted investigation or resolve the condition with its owner.
Compare the next commitment on real work
Suppose Claire Bennett coordinates an industrial-parts distributor's operations shortlist. Service coordinators want help assembling repair-case summaries. Inventory planners want better forecasts of demand for spare parts. A third request proposes AI to find missing equipment serial numbers on repair-intake forms.
Claire asks the service supervisor and inventory planner to explain their proposals using recent work. The service team can show technician notes and examples of accepted summaries. The planner sees a potentially important stock problem, but does not yet know whether historical service records identify the part actually used. These are different evidence positions, even if both sponsors describe the potential value as high.
| Candidate | What the team can establish now | Next commitment to consider |
|---|---|---|
Draft repair-case summaries from approved technician notes | The supervisor can compare a draft with its sources, checking symptoms, parts and open issues. The amount of correction work remains unknown. | A bounded comparison of accepted summaries and total effort, if the inputs and tool are approved |
Forecast spare-part demand | The stock problem merits investigation. Whether the historical records support a useful forecast remains unresolved. | Have the planner and data owner inspect record coverage and part identifiers before proposing a forecast trial |
Find missing serial numbers in intake forms | An omitted required field can be detected with an ordinary form rule. | Assess that simpler process fix rather than commissioning an AI model to detect blank fields |
Claire can recommend investigating the forecast records while preparing a limited summary trial, subject to available people and approvals. She does not need to declare forecasting a bad idea or promise that summarization will save time. The serial-number problem can follow a simpler route.
The comparison also exposes a possible conflict: the same supervisor may be needed to validate historical repair records and check trial summaries. Put that dependency in front of the person allocating their time. The guide to AI steering committee decisions explains how to handle shared capacity and decision authority.
Ask what would change the order
A recommendation becomes more useful when you can explain what would overturn it. For each leading candidate, finish the sentence: “We would change this recommendation if…” Then identify the evidence needed to answer that question.
For Claire's shortlist, the summary trial becomes less attractive if checking and correcting its drafts takes as much work as the current method. The forecasting investigation becomes more promising if the planner finds sufficiently complete, relevant records and can explain how a forecast would change stocking decisions. These are conditions to examine, not results already achieved.
The UK Analysis Function's introductory guide to multi-criteria decision analysis recommends examining how uncertain inputs affect option rankings. It also warns that unclear scoring can mix performance with preference. That guidance supports challenging a ranking's assumptions; it does not validate an improvised AI scorecard or turn this qualitative comparison into formal decision analysis.
Use the disagreement to choose the next evidence-gathering step:
- Disagreement about the work. Inspect a relevant example with the performer or reviewer. Check whether both proposals describe the same problem and acceptable result.
- Disagreement about a forecast. Name the assumption that changes the recommendation and find evidence for it. Keep the estimate labeled until checked.
- Disagreement about priorities. Ask the responsible sponsor to state the tradeoff. More data cannot decide how much customer service matters relative to inventory cost.
Some assumptions affect several judgments together. Poor repair records might weaken both forecast feasibility and the expected benefit. Do not count the same hoped-for improvement under several headings and then treat the total as independent confirmation.
For a consequential funding choice, build the fuller adoption-program business case, including credible alternatives, costs and uncertain benefits. A shortlist comparison helps decide where to look next; financial appraisal and trial evidence must support larger commitments.
Leave a recommendation someone can revisit
Record the recommendation in a form the next reviewer can understand:
- Proposed commitment. Investigate the forecast records, prepare a bounded summary trial or make the form change.
- Reason. Explain the relevant value, effort and evidence, including the strongest reason to prefer another candidate.
- Open condition. State what must be resolved, who owns it and what activity is permitted meanwhile.
- Return point. Name the evidence or event that triggers reconsideration and who will make that decision.
Keep promising but weakly evidenced ideas visible with a bounded next question. Keep apparently easy ideas accountable for a useful result. You are choosing how to spend the next unit of attention, while preserving enough reasoning to change your mind when the evidence changes.
Questions about prioritizing AI opportunities
Should we use a weighted AI scorecard?
A weighted AI scorecard can help structure a discussion when participants understand the criteria, scales and evidence. A total is much less useful when one person's “high value” means a measured problem and another's means a large untested forecast. It can conceal disagreement rather than resolve it.
Start with the evidence comparison, keep unknowns explicit and separate conditions that require resolution from benefits you can trade off. If the decision warrants formal numerical appraisal, involve someone qualified to design it. An arbitrary set of weights does not establish the best investment or authorize the proposed use of information.
What if the highest-value AI idea has weak evidence?
Keep the potential benefit visible and name the uncertainty that could change the decision. A spare-parts forecast might address an expensive stock problem while depending on historical records no one has checked. That can justify a bounded investigation of those records, rather than immediate implementation or rejection.
Identify an owner, permitted inputs and a return point for the investigation. Compare its effort with other calls on the same people. The recommendation reversal test shows how to state what evidence would make an uncertain candidate more or less attractive.
How is adoption effort different from technical feasibility?
Technical feasibility asks whether the proposed system can produce a usable result with the available inputs and integrations. Adoption effort asks what people must change to use, check and maintain it. Both affect whether an AI opportunity is worth pursuing.
A repair-summary tool might generate plausible drafts while requiring a service supervisor to correct every part identifier and unresolved issue. Record that checking burden alongside the technical work, and involve the supervisor before committing their time. Use the five comparison lenses to keep the two questions visible.
Should every first opportunity be a quick win?
No. An opportunity that creates useful foundations or addresses an important longer-term problem may deserve investigation even if it cannot deliver an immediate result. Equally, a cheap prototype deserves a clear purpose and a limit on the effort people will spend assessing it.
Compare the next commitment and what it can establish, rather than requiring every candidate to promise the same delivery horizon. Keep dependencies and shared capacity explicit. Use the strategy guide's pilot criteria when an investigation is ready to become a trial with an owner, measures and an expansion decision.
