An impact-effort matrix for prioritizing AI projects

When every department proposes an AI project, the loudest sponsor can become the priority by default. An impact-effort matrix gives the team a common comparison, provided that the numbers represent explicit assumptions. Use it to choose what to investigate next. The scores should not make uncertain benefits look precise. The OpenAI use-case guide presents impact versus effort as a prioritization tool. Deloitte's agentic AI framework also considers implementation complexity and suitability. For a small business, I would add two visible checks: can the required information be used, and can errors be contained during a test?

Define the axes before scoring ideas

Use a five-point scale. For impact, assess observed workload, effect on an important service objective and reach across users. A score of one means a minor local inconvenience; three means a recurring measurable bottleneck; five means a major business constraint supported by evidence. Do not score a task five merely because its sponsor mentions revenue. For effort, include source preparation, integration, evaluation, staff training and ongoing support. One means a bounded task using accessible approved inputs and existing review; three means several sources and some integration; five means substantial dependencies or consequential actions across systems. These anchors are an example. Adjust them before reviewing the backlog, then keep them consistent. Have the process owner propose impact and the technical lead propose effort. They should record the evidence and uncertainty together. Investigate a disagreement of two points before taking an average.

A hypothetical distributor's backlog

All scores describe an invented company. Suppose delivery replies occur frequently and delay the support queue; that supports impact four. Policy drafts require no live system writes, supporting effort two. Technical substitution could affect more valuable orders, but missing compatibility evidence and specialist review make it a poor first experiment. The table does not imply that four divided by two is a reliable return on investment. Ordinal scores express a judgment about relative priority; they do not express money or delivery time. Record the underlying volume, delays and dependencies separately so decision makers can inspect the reasoning.

Check prerequisites before choosing a quadrant

A candidate needs a named business owner, permitted inputs, a defined reviewer and a way to detect unacceptable outcomes. Mark each as ready, unresolved or blocked. A high-impact idea with unresolved authorization remains an investigation candidate. An attractive average score does not justify moving to a live pilot without authorization. Then use four broad positions: high impact and lower effort deserves an early bounded test; high impact and higher effort needs staged discovery; lower impact and lower effort may suit personal assistance; lower impact and higher effort usually waits. Keep the boundary between low and high visible. In this example, scores four and five are high.

Challenge the result with a sensitivity check

Ask what happens if estimated frequency is half as large or required review is twice as demanding. If the leading idea immediately loses its advantage, gather better evidence before committing. Also ask whether fixing a form, clarifying a policy or enabling an existing system feature would resolve the bottleneck more simply. For the distributor, the immediate decision is to test delivery drafts while checking inventory integration access in parallel. The support lead owns acceptance of replies. The systems owner obtains evidence about the inventory connection. Substitution remains outside the test, regardless of its potential impact.

Keep a decision record

Save the scoring definitions, evidence, participants, unresolved prerequisites and reason for selecting one next step. Revisit when a dependency changes or new operating evidence appears. The NIST generative AI profile supports ongoing assessment; the matrix is only one planning document. For help applying these criteria to your actual backlog, bring three candidate tasks and their available evidence to an AI strategy assessment.