KCM GUIDE

The AI Worked Yesterday. What Changed?

Test AI changes before the business relies on them.

An AI assistant may give clearer answers after a prompt update while dropping an instruction the business still requires. This article shows how to compare the current and candidate configurations, inspect individual failures and record a release decision that an owner can defend.

Read the full article ↓
Current and candidate paths have matching start and end positions but different middle turns beneath shared expectations; open cyan brackets mark the comparison.

Compare matched current and candidate evidence.Inspect mandatory failures behind the overall score.Record an accountable release decision.

Imagine a service assistant that gives clearer answers after a prompt update. More test responses pass. The team is ready to release it, until someone notices that the shorter answer no longer tells a coordinator to obtain a supervisor’s approval before promising a replacement.

That is the release decision this article addresses: whether to approve a change to an AI workflow people already use. It is different from choosing a platform or deciding whether an initial pilot is ready.

A better overall score does not resolve that decision. You need to know what improved, what became worse, and whether any failure breaks a mandatory business rule.

Treat the change as a testable proposal

AI regression testing checks whether a change has damaged behavior the business still requires. A change may involve a prompt, model, source document, retrieval setting, permission, tool or application component. Testing only the model can miss a failure elsewhere in that chain. Microsoft’s guidance explicitly distinguishes model evaluation from testing the complete system when a change is introduced. Microsoft: Test and evaluate AI workloads.

Write a short change statement before comparing answers: what is changing, what benefit is expected, which users and tasks are affected, and what must remain true. For example: “Replace prompt P1 with P2 to make service instructions easier to scan. Mandatory approvals, source citations and access boundaries must remain intact.”

Record the current and candidate configurations. Compare them using the same approved source versions, test identities, cases and scoring rules, except for the differences the proposal deliberately introduces. If several components must change together, test that bundle, but do not attribute the result to the prompt alone.

The current system also needs a fresh run. Last month’s score may describe different documents, permissions or service behavior.

A decision: more passes, but a hold

Consider a North American equipment distributor. Its internal assistant explains return procedures; it cannot authorize a replacement or change a customer record. Coordinators use the answer to decide what to do next.

The team tests a shorter prompt against the current prompt. Both use the same model deployment, retrieval settings, approved policy RP-4 and role assignments. Twelve cases run three times each under each configuration, producing 36 trials per configuration. A trial passes only when every required check for that case passes.

Case group Trials per configuration Current passes Candidate passes
Routine instructions 18 14 18
Mandatory escalation 6 6 4
Access and refusal 6 6 6
Rules revised in the current policy 6 4 6
Total 36 30 34

The figures are not product benchmarks. The candidate passes four more trials overall. Yet it fails two of three trials for one mandatory escalation case.

The coordinator asks: “The unit is damaged and the customer cannot provide its serial number. Can I promise a replacement today?”

Current answer: “No. Record the order number and damage photos, then send the request for supervisor review before promising a replacement. See RP-4, section 3.2.”

Candidate answer in a failing trial: “Record the order number and damage photos, then start the replacement request. See RP-4, section 3.2.”

The candidate cites the right document and includes useful steps. It omits the required restriction and review. A citation check alone would not catch that omission.

The predefined rule is that every mandatory escalation trial must pass. The decision is therefore hold the candidate, correct it and retest. Better routine answers cannot compensate for a new failure that could lead a coordinator to make an unauthorized commitment. The guide provides the case-level results and a reusable decision worksheet.

Separate an improved answer from a changed requirement

Regression testing should preserve required behavior, not yesterday’s wording.

Suppose the distributor legitimately changes the deadline for attaching damage photos from five business days to three. An answer saying “three” should not fail because an old reference answer says “five.” The policy owner must approve the revised expected answer, its effective date and the cases it affects before the release comparison.

Both current and candidate runs should then be judged against the same current requirement. Keep historical results under their original policy version; do not silently rewrite them. If old and new policies apply to different transactions, include the transaction date in the case and score each against the applicable rule.

Also separate source support from completeness. Microsoft’s RAG evaluation documentation treats groundedness, meaning support from the supplied context, and response completeness as different dimensions. KCM’s practical implication is simple: ask both “Can the source support this answer?” and “Did the answer include every requirement the user needs?” Microsoft: RAG evaluators.

Look at repeatability and the failed cases

One attractive answer is weak release evidence. Generative outputs can vary, so repeat important cases and retain every result. OpenAI recommends task-specific evaluations, ongoing comparison after changes and human calibration of automated scoring. OpenAI: Evaluation best practices.

In the example, the candidate succeeds on the escalation case once and fails twice. Reporting its best answer would conceal the problem. Three trials make the comparison readable; they are not a recommended sample size or proof of a production failure rate. Choose coverage and repetition according to the workflow’s consequences and observed variability.

Inspect the available evidence behind each disputed result: input, user role, retrieved passages, response, citations, tool events and final business state where relevant. A failed score may reflect an actual regression, an outdated expected answer or a faulty scoring rule. Anthropic’s evaluation guidance emphasizes reading trial records rather than accepting aggregate scores at face value. Anthropic: Demystifying evals for AI agents.

Treat unresolved mandatory cases as a hold. Do not revise a scoring rule merely because the candidate failed it.

Make approval specific, and recovery realistic

The accountable business owner should approve a named configuration for named users and uses. The technical owner supplies the comparison evidence and a tested recovery path; security or data owners review changes within their responsibility.

Approval should require all mandatory rules to pass, agreed quality and performance limits to be met, and outstanding issues to be resolved or explicitly accepted where policy permits. Compare response time and cost per acceptable outcome, including retries and human correction. A cheaper response that creates more rework may not be cheaper service.

A narrower release is possible only if the excluded cases can actually be kept out through enforceable routing or access controls and the remaining scope passes its own tests. “The user will know when not to trust it” is not a boundary.

Recovery must preserve current business truth. Reverting a prompt does not justify restoring an obsolete policy or revoked access. Nor does it undo messages sent or records changed by a tool. If the previous configuration is no longer valid, pause the affected workflow and use the agreed manual route while a safe replacement is tested.

Monitor after release as well. NIST’s March 2026 report explains why controlled predeployment tests cannot substitute for observing deployed systems under changing real-world conditions. NIST: Challenges to monitoring deployed AI systems.

Start with the next change already in your queue

Choose one proposed update. Ask its owner for the configuration difference, current expected behavior, matched test results, mandatory failures and recovery limits. If those pieces are missing, the next step is to assemble that evidence before approving the update.

KCM’s proposed release method is an operating decision aid, not a vendor certification or guarantee. Its purpose is to make a small change understandable enough that someone can responsibly approve it, hold it or reduce its scope.

Explore KCM AI consulting to connect an AI change to the data, controls and evidence your business needs.