AI Performance Summaries Need More Than a Warning

A manager opens a quarterly review and finds a new card labeled “communication summary.” It says an employee was slow to respond, contributed little in meetings, and showed uneven follow-through. The card cites several chat threads and meeting transcripts, but it does not show which passages produced each conclusion.

That card is the governance problem.

Products including Microsoft 365 Copilot, Slack AI, and Zoom AI Companion can summarize workplace content in different contexts. A summary used to catch up after a meeting is one thing. The stakes change when an employer copies it into a performance file, uses it to shape a rating, or presents it to a promotion committee as evidence about reliability or contribution.

The disclosure question is therefore narrower than whether the employer “uses AI.” Workers need to know when a system crosses from helping someone read communications into evaluating the people who produced them.

How the summary card gets made

A typical system first selects material from connected sources, such as email, chat channels, meeting transcripts, calendars, or task records. It passes that material to a language model with instructions to identify themes, missed commitments, participation patterns, or examples of completed work. A separate application may then store the output in a manager dashboard or human resources system.

Every stage changes what the card means. The retrieval layer, which selects material for the model to read, may omit private channels, old messages, attachments, edited posts, or meetings that were never transcribed. The prompt may treat message frequency as engagement. Speaker identification can assign a meeting comment to the wrong person, while sarcasm, disagreement, and work completed outside the connected tools remain difficult to interpret from text alone.

Permissions do not solve the evaluation problem. A manager may already have lawful access to a team channel, yet workers may not expect messages written for day-to-day coordination to be compressed into a persistent judgment about performance. The system also creates a new record: even if the underlying chat is deleted on schedule, the summary card may remain in an HR platform for years unless someone sets a separate rule.

An employer can keep this workflow narrow. The model can draft a summary for the worker to review, preserve links to supporting passages, and prevent the output from entering a personnel file until a person confirms it. That setup costs staff time and makes the workflow slower. It also creates a record that can be inspected rather than an unattributed conclusion that gains authority because it appears in a dashboard.

Put the notice where the purpose changes

A general employee handbook statement that AI may be used for productivity does not explain the communication summary card. Disclosure should arrive before the output is used for evaluation, with enough specificity that a worker can understand the input, purpose, audience, and consequence.

For the card, that means naming the communication sources and the period searched, including material the system cannot see. The notice should say whether the model is asked to summarize events, infer traits, score behavior, compare employees, or flag departures from a manager-defined expectation. Those are different capabilities. “Summarization” is an incomplete label if the prompt asks the model to judge responsiveness or leadership.

The audience matters just as much. Workers should be told whether the output remains with their direct manager, moves into an HR record, reaches a promotion panel, or becomes available to a vendor. If managers can regenerate the card with different instructions, the disclosure should cover that discretion and the system should log the prompt, source window, output, editor, and time of use.

Disclosure is not the same as consent. It does not settle whether consultation, collective bargaining, a lawful basis for processing, or another authorization is required in a particular workplace. It creates the minimum information needed to inspect the workflow and contest its output.

Access must include the evidence trail

Showing a worker the final paragraph is weak access. The employee needs the version that influenced the decision, along with the passages or records cited as support and a description of material the system excluded.

Return to the communication summary card. “Slow to respond” could refer to unanswered direct messages, delayed email, or a meeting transcript containing a promised deadline. Without citations, the worker cannot tell whether the model found a real pattern, confused two people, or interpreted an after-hours message as requiring an immediate response.

Source access still needs boundaries. A cited thread may contain another employee’s confidential information, and a manager’s private notes may be restricted. The system can expose the relevant excerpt, record identifier, date range, and reason for inclusion without opening every connected mailbox. Where even an excerpt cannot be shown, the employer should record that limitation and avoid treating the hidden source as uncontestable evidence.

Versioning matters. If a manager edits the summary before saving it, the system should retain both the generated text and the human revision, rather than presenting the combined result as purely automated or purely human. That receipt establishes which actor introduced a claim and which version reached the decision-maker.

Correction should preserve the dispute

A correction mechanism should handle two different failures. Factual errors include a message assigned to the wrong worker, a task marked late despite a changed deadline, or a transcript that dropped a speaker. Interpretive disputes arise when the source is accurate but the conclusion is contested, such as treating fewer meeting comments as weaker contribution.

Silently replacing the card is the wrong fallback. The original output should remain in an audit record, which is a history of the system’s inputs, outputs, edits, and uses, while the personnel-facing version carries the correction or employee response. If the system regenerates the summary, it should link the new card to the prior one and show what changed.

This design prevents a practical failure: a worker wins a correction in one dashboard, but the old summary has already been exported to a review document or promotion packet. A correction policy needs downstream propagation, meaning every active copy receives the update or a visible dispute marker. Otherwise the correction button is cosmetic.

Retention should follow the decision, not the chat platform

Derived records need their own retention rule. An email may disappear under one schedule while a performance summary based on it survives elsewhere, stripped of the context needed to test it.

The employer should specify how long the card, its source references, the prompt, the model or system identifier, human edits, and any appeal outcome will remain available. Keeping everything indefinitely makes later reconstruction easier, but increases privacy exposure and allows stale judgments to follow workers into unrelated decisions. Deleting the evidence immediately is no better if the summary remains in use.

A defensible operational trigger is the period during which the employment decision can be reviewed, followed by deletion or restricted archival handling under the organization’s applicable record rules. The exact period will depend on jurisdiction, contracts, and the decision involved. The governance requirement is consistency: the summary should not outlive the ability to inspect its basis.

Appeal needs authority to stop reliance

An appeal channel must reach someone who can change the decision. Sending feedback only to the product team may improve a prompt later, but it does nothing for the worker whose current rating contains the disputed card.

The reviewer should be able to pause reliance on the summary, inspect the cited communications, obtain context from the worker and manager, and remove or qualify unsupported claims. The system should record the outcome and whether other summaries produced by the same configuration need review. A repeated speaker-attribution error is not confined to one employee.

Human review alone is not a safeguard if the reviewer sees only the model’s conclusion. The human needs the source trail, uncertainty, exclusions, and correction history. Otherwise the model frames the issue before the reviewer begins.

Existing rules do not reduce to one AI notice

The European Union’s AI Act treats certain systems used to monitor or evaluate worker performance as high-risk, subject to its scope and exceptions. Its workplace provision says: “Before putting into service or using a high-risk AI system at the workplace, deployers who are employers shall inform workers’ representatives and the affected workers that they will be subject to the use of the high-risk AI system.” The regime has phased application, and not every communication summarizer automatically falls into that category.

The EU General Data Protection Regulation separately addresses access, correction, storage limitation, and certain solely automated decisions with legal or similarly significant effects. Those duties turn on the processing and its effects, not on whether a vendor markets the feature as a summary.

In New York City, Local Law 144 covers certain automated employment decision tools used in hiring or promotion and requires a bias audit plus notices. A tool used only for ordinary performance coaching may sit outside that law, while a summary that substantially assists a promotion decision may raise a different analysis. Across the United States, existing employment, discrimination, privacy, labor, and records rules can still apply without a single federal disclosure template for AI-generated performance summaries.

This is a design checklist, not legal advice. The practical procurement test is whether the employer can configure the communication summary card to expose sources, preserve versions, enforce deletion, accept corrections, and suspend use during an appeal. If the product exports only polished prose, with no durable link to its evidence or configuration, it is not worth using for performance decisions.

Questions people ask

Should workers see every AI-generated performance summary?

They should be able to see any summary relied on for a rating, promotion, discipline, work allocation, or another consequential employment decision. Access should cover the operative version and its supporting record, subject to necessary limits for other people’s confidential information.

Is a general notice that the company uses AI enough?

Usually not for governance purposes. A useful notice identifies the communication sources, evaluation purpose, intended recipients, retention period, human role, and correction or appeal route. A broad productivity notice does not reveal that meeting transcripts may later become performance evidence.

Can a manager correct the AI summary before showing it to HR?

Yes, but the system should preserve the generated version and identify the manager’s edits. Keeping both versions makes responsibility visible, supports later correction, and prevents a human-edited judgment from being represented as an untouched model output.

What should happen while a worker challenges a summary?

The employer should mark the card as disputed and pause reliance on contested claims where the decision can still be held. An authorized reviewer should inspect the cited communications, record the outcome, and propagate any correction to copies already placed in review or promotion materials.