The feature’s existence is corroborated by Microsoft Learn’s archived release-plan history, which records “Multi-grader comparison” as added to Power Platform governance and administration on June 30, 2026. That history establishes an earlier planning milestone, not current availability.
What the comparison experience would show
The roadmap record describes four principal capabilities:
- Compare graders side by side.
- Examine their scores and pass/fail outcomes.
- Review each grader’s explanation and supporting evaluation evidence.
- Analyze agreement, differences, and grading consistency.
The record lists Web as the platform and Worldwide (Standard Multi-Tenant) as the cloud scope. Its product labels include Microsoft Copilot (Microsoft 365) and Microsoft Power Platform governance and administration. Those labels alone do not establish which authoring surfaces, subscriptions, or permissions will expose the experience.
Availability remains unconfirmed. The record still says “In development,” alongside estimated preview availability in September 2026 and general availability in October 2026. As of October 21, 2026, the preview month has passed and the general-availability month is underway, but those dates are not proof of deployment. The live roadmap page’s retrieved content did not expose item 577183’s detailed fields; Microsoft also explicitly warns that roadmap dates are estimates and subject to change.
Running several graders is not the same as comparing them
Microsoft already documents a related Copilot Studio capability: applying multiple graders to one test set in a single execution. Its release-plan documentation describes adding graders such as general quality, text similarity, and exact match, setting individual passing criteria, and displaying results in separate grader columns with an aggregate evaluation summary.
That distinction matters. Multiple-grader execution gathers different judgments; the proposed comparison experience emphasizes interpreting those judgments together. The existing documentation is useful context, but it does not prove that roadmap item 577183’s explanation-and-evidence comparison interface has shipped. Microsoft’s Copilot Studio blog likewise describes combining graders to obtain a multidimensional assessment, rather than treating agent quality as a single measurement.
The practical opportunity is therefore narrower—and more useful—than “another AI score.” If delivered as described, the feature could help makers investigate why graders disagree.
A disagreement is a question, not a verdict
Microsoft Learn’s assertion guidance explains why differing outcomes need interpretation. It recommends matching the grading mechanism to the requirement: exact matching for exact values, text similarity for semantic meaning, and LLM-based judgment for qualitative assessments. Assertions should also be atomic, binary, outcome-focused, and verifiable.
Consider a hypothetical support agent whose answer conveys the expected meaning but uses different wording. An exact-string check and a semantic assessment could reasonably produce different outcomes. That does not automatically make either grader defective; the first question is whether the test actually requires identical wording.
A useful review approach, based on Microsoft’s assertion guidance, is to:
- Identify the requirement. Is the test checking a precise value, meaning, completeness, or another observable outcome?
- Check the grader’s fit. Does its mechanism measure that requirement?
- Review the explanation and evidence. Look for the reason behind the result, not merely which score is higher.
- Clarify ambiguous assertions. Separate factual correctness from other requirements instead of bundling them into one judgment.
This is an interpretation workflow, not a verified sequence of clicks for the upcoming interface.
The governance value—and the limit
For enterprise teams, the potential benefit is a more explainable evaluation review: understanding disagreement before accepting a result or changing an agent. But the roadmap description does not promise automatic conflict resolution, grader validation, or root-cause diagnosis.
The sensible takeaway is to watch 577183 as a planned comparison capability, not a confirmed launch. Better visibility can make an evaluation easier to challenge. It cannot, by itself, turn a poorly defined test into a good one.