A convincing AI demo is not an acceptance test: what should executives require?
Define the decision before choosing the score
An AI system answering demonstration questions well is not enough to approve all uses. Executives should first define the task, input types, affected people and downstream use of outputs. Approval to draft for human review differs from approval to send answers directly to customers, even with the same model.
NIST’s AI RMF describes evaluation before deployment and during operation, with documented test sets, measures and contextual limitations. The following acceptance approach applies that principle; it is not NIST certification or a Thai legal requirement. A pass threshold should reflect task risk rather than a pleasing number followed by examples selected to reach it.
Separate correctable defects from critical failures
List desired outcomes and unacceptable events before reviewing results. A policy-search assistant might tolerate wording that needs editing, but should not expose another person’s information or invent a material condition. The process owner should define how each category is judged, who reviews it and when the system must defer to a person, so fluency does not conceal high-impact errors.
Hypothetical example: out of 100 test cases, the system passes 95 ordinary cases but fails five cases that all concern data-access permissions. A 95% aggregate score does not answer whether deployment is acceptable. Review critical groups and severity separately. These figures illustrate interpretation only; they are neither a recommended threshold nor a client result.
Choose tests that expose limitations
Acceptance tests should cover routine tasks, ambiguity, incomplete inputs and cases requiring refusal or an admission of uncertainty, based on actual organizational use. Separate development examples from acceptance cases and restrict data access as needed. If languages, user groups or document types remain untested, state those limits in the approval instead of assuming coverage.
Record the system version, knowledge set, relevant configuration, test date and scoring method alongside results. Expert reviewers need shared criteria and a procedure for disagreements. If criteria change after results are seen, document why and retest affected scope. Renaming failures to make a report appear to pass is not a valid acceptance process.
Account for review effort and stage the decision
Measure total work time, including review, corrections, returns and exception handling, rather than only generation speed. Faster drafting may not deliver the demonstrated savings if reviewers must verify every assertion. Executive reports should distinguish measured benefits from projections and specify the human capacity still needed at larger scale.
The decision may be approval for limited scope, continued trials with review of every case, or no deployment yet. Identify who accepts residual risk, stop conditions and retesting triggers when models or data change. For discussions with Kerwin & Vance, bring business objectives, test cases and categorized results to connect AI decisions with internal controls and actual workload.
