How Founders Should Score AI Opportunities

Founders should score AI opportunities with a weighted rubric tied to operating evidence, then apply vetoes that can overrule the total. Use a five-point scale for each factor, attach a written observation to every score, and compare candidates under the same assumptions. The score ranks attention. It does not predict ROI or authorize a deployment.

Build the rubric from operating facts

Use six factors whose weights add to one hundred percent. Business value carries thirty percent because a technically elegant idea that changes nothing should not lead the backlog. Frequency carries fifteen percent. Evidence confidence carries fifteen percent. Implementation effort carries fifteen percent and is scored in reverse, so easier work receives the higher score. Adoption carries fifteen percent. Risk and reversibility carry ten percent.

• Business value, thirty percent. Score whether the result can change revenue, cost, cycle time, or a defined risk outcome.
• Frequency, fifteen percent. Score how consistently the task occurs and whether enough representative cases exist.
• Evidence confidence, fifteen percent. Score the quality of workflow observations, records, labels, and baseline information.
• Implementation effort, fifteen percent. Score a simple, bounded, low-dependency change higher than a broad integration with uncertain upkeep.
• Adoption, fifteen percent. Score whether a named team can use, review, and improve the output inside its normal process.
• Risk and reversibility, ten percent. Score higher when permissions are clear, external actions are constrained, and rollback is practical.

Multiply each factor score by its weight and divide by five, then add the results. Every score needs a note such as observed in the sales log, confirmed by the process owner, or assumed pending a sample. The note matters more than false precision. When evidence is absent, score confidence low and state what would change it.

The local AI test contributes a useful effort check: some capability questions can be tested on existing hardware before procurement. That may improve the effort and reversibility scores, but it should not raise business value or adoption without workflow evidence.

Apply vetoes before trusting the winner

A total score cannot compensate for a critical disqualifier. Veto a candidate if the required data cannot be used lawfully, no business owner accepts the outcome, an external action would be irreversible during the test, there is no safe fallback, or a reviewer cannot inspect the evidence behind a consequential result. Record the veto separately. Do not hide it as a slightly lower risk score.

Risk-sensitive work shows why this rule exists. The medical data security case study contributes a control pattern based on full-surface audit, workflow-specific defenses, and continuous monitoring. For scoring, its contribution is not a transferable outcome claim. It is evidence that operational controls and ownership must be designed around the actual data path.

Challenge the highest-ranked option harder than the losers. Ask which assumption contributed most to its position, who supplied that assumption, and what observation could reverse the ranking. If the winner survives only because unknown integration effort was scored as easy, lower the confidence and gather evidence before allocating a team.

Worked example: follow-up versus document generation

Assume a founder compares sales follow-up drafting with general document generation. Internal evidence shows missed follow-ups in an existing queue, a named sales owner, reusable message history, and a review step already inside the customer relationship system. On the five-point scale, assign follow-up scores of five for value, four for frequency, four for confidence, four for effort, four for adoption, and three for risk and reversibility. The weighted result is eighty-four out of one hundred.

For document generation, assume requests vary widely, owners disagree on acceptable output, and staff would review drafts outside their normal tools. Assign three for value, three for frequency, three for confidence, two for effort, two for adoption, and four for risk and reversibility. The weighted result is fifty-six. These are explicit hypothetical scores, not company results. Sales follow-up ranks first because its path from output to action is clearer.

The CFO payback guide contributes the next challenge: connect the chosen job to a metric, cost of inaction, decision makers, and an acceptable payback window. The ranking chooses where to investigate. Only observed cost and benefit inputs can support a later investment case.

Convert rank into a founder decision

The winner should receive a first experiment, not automatic funding. Define the trigger, allowed data, draft output, reviewer, fallback, success evidence, and stop condition. Re-score after the experiment. If confidence rises and the priority remains stable, the founder has a defensible sequence. If the ranking changes, the rubric has done its job by exposing a weak assumption early.

Frequently Asked Questions

Keep the scorecard with its evidence notes, veto log, worked comparison, and proposed experiment. For a structured opportunity inventory and independent challenge of the leading candidate, request an AI Opportunity Report.