When AI Observes Performance, What Counts as Evidence?

Performance management once depended largely on what managers and employees could observe, remember and document. AI changes that.

Meetings, calls, teaching sessions, written communication and system activity can now be recorded, transcribed and analyzed at scale. AI can identify patterns, flag behavior and generate summaries or scores for managers to review.

This potentially gives organizations a richer evidence base, but creates a much more fundamental question:

When AI observes something, what makes it legitimate performance evidence?

A recent case involving UK training company Multiverse illustrates the problem. Its AI system analyzes transcripts of online teaching sessions, flags possible performance issues and assigns risk and confidence indicators for managers. Multiverse says that human managers remain responsible for performance reviews, but employees have raised concerns about accuracy, transparency, continuous scrutiny and the future use of their information.

The distinction between AI and human decision-making is important – but it is not enough. A human decision can still be questionable if it relies on evidence that is irrelevant, incomplete, taken out of context or collected without appropriate safeguards.

Organizations therefore need to make decisions about more than whether AI should make the final judgment. They need to decide:

• Which employee behaviors/actions should AI observe or track
• Whether those behaviors are genuinely relevant to the work
• How this performance context will be retained and considered
• Whether employees can see and challenge the evidence
• When an AI-generated observation should inform coaching
• Whether it is sufficiently valid to contribute to formal assessment
• Where the underlying information accessed by the AI is processed and stored
• What else that information may be used for

These are performance-management questions, evidence questions and privacy questions. They cannot be resolved by adding a human reviewer at the end of the process.

AI can detect behaviors that previously went unnoticed or were too time-consuming to examine consistently. It may count interruptions, identify filler words, measure response times, analyze sentiment or flag departures from an expected process.

But observing and counting something doesn’t make it a meaningful measure of performance.

Frequency is an example.  Speaking more often doesn’t necessarily mean contributing more value. Responding quickly does not mean the response is correct.  Standard scripts support consistency, but departing from them may be necessary to address an unusual situation or meet an individual customer’s needs.

A behavior can therefore be:

• Accurately detected but unrelated to effective performance
• Relevant in one role or situation but not in another
• Useful as a prompt for discussion but unsuitable for assessment
• Meaningful only when considered alongside its outcome
• Misleading when separated from the circumstances in which it occurred

AI also sees only the data available to it.  Activities such as Planning, preparation, problem-solving, informal support and offline work may not be captured.  The resulting picture may be highly detailed but still incomplete.

Organizations therefore must distinguish between observable activity and valid performance evidence. The question is not simply, “Can we measure this?” It is, “What does this observation tell us about performance against a legitimate expectation?”

The availability of data can easily reverse the logic of performance management. Instead of defining effective performance and then identifying suitable evidence, organizations may begin with the information their systems can collect and treat it as important because it is available.

The risk is measuring what is technically visible rather than what matters.

Valid performance evidence begins with a clear understanding of the role:

• Why the role exists
• What the person is accountable for
• What outcomes and standards are expected
• Which behaviors are genuinely important to performing the work
• What evidence would demonstrate effective performance

Goals provide evidence about particular priorities and results, but they do not define the whole job. Continuing accountabilities, quality standards, professional judgment, collaboration, customer outcomes and relevant organizational values also matter.

Any AI-observed behavior must therefore be traceable to an established performance expectation. If you can’t explain why the behavior or action matters to the work, it should not be used as a performance measure just because the technology can detect it.

Even where the connection appears credible, the organization should check.

• Does the behavior reliably distinguish effective from ineffective performance?
• Could there be different but equally effective approaches?
• Might the observation process result in a narrow focus on this target, behavior or action – with adverse consequences?

AI may accurately detect that an employee responded slowly, departed from a standard process, spoke less than expected in a meeting or had an unsuccessful task outcome.   What it cannot safely infer is why.

The same observable behavior can have very different causes.  A delayed response might reflect insufficient knowledge, an unavailable system, conflicting priorities or the need to investigate a complex issue properly. A departure from an expected process might indicate poor understanding—or sound professional judgment in circumstances the process did not anticipate.

Possible causes of an apparent performance gap include:

• Expectations that were unclear, inconsistent or recently changed
• Insufficient knowledge, skill or practical experience
• Inadequate tools, information or resources
• Workload pressure or conflicting priorities
• Lack of authority to make the required decision
• Failures elsewhere in the process or system
• Adaptation to unusual circumstances
• Weak or no feedback or support
• Consequences that unintentionally discourage the expected behavior
• A genuine problem with effort, judgment or performance

Some of these causes sit with the employee. Others sit with the manager, the process or the organization.

An AI-generated observation should not enter a performance record simply because the system produced it. Its value depends on whether it can support a fair and meaningful conclusion about the person’s work.

Before using an observation as performance evidence, organizations should apply several tests.

Relates directly to an established requirement of the role, an agreed goal or core behavioral expectation, with a clear explanation of why it matters.  Correlations with performance or assumptions are not enough.

The underlying event is detected and reported correctly. Checking means access to the source data. Transcription errors, missing information and incorrect interpretation can all create an apparently precise but inaccurate record.

There must be context.  A transcript may show what was said but not why it was said.  Response time measures may not show the complexity of the issue.  The employee’s explanation and other relevant evidence should be considered before a conclusion is reached.

AI may analyze only the activities captured by a particular system.  Visible and easily measured behavior is reported while other important contributions are not.  A large quantity of data does not correct a biased or incomplete sample.

Employees should know what information is being collected, what the AI is assessing and how the resulting observations may be used.  They should be able to see relevant evidence, identify inaccuracies, provide context and challenge an interpretation before it contributes to a consequential decision.

The strength of the evidence should match the significance of the decision.  The more consequential the decision, the stronger, more complete and more independently reviewed the evidence should be.

The position changes when the same information is used to:

• Contribute to a formal performance rating
• Compare or rank employees
• Influence pay, promotion or access to development opportunities
• Identify someone as a performance risk
• Begin a disciplinary or capability process
• Support decisions about continued employment

Evidence used for these purposes requires a much higher standard of relevance, accuracy, context and review.

Telling employees that AI is being used is not the same as making its use transparent.  An employee may know that meetings or work activities are being analyzed. But they may not know what actions, behaviors or results are involved, how they are converted to flags or scores, who can see them, how long they are kept and what decisions they influence.

If AI-generated observations are used in performance management, employees need access to the relevant underlying evidence and a meaningful opportunity to respond, including the option to:

• Correct factual or transcription errors
• Explain the circumstances surrounding an event
• Identify important work or outcomes the system did not capture
• Challenge whether the observed behavior is relevant to their role
• Question how the AI reached its interpretation
• Ask for an appropriately qualified person to review the evidence
• Add their response to the formal record

AI performance monitoring creates risks beyond the accuracy of the resulting assessment.

To analyze work, the system may receive recordings, transcripts, messages, documents, customer interactions and system activity. This material can contain personal information about the employee and other people. It may also reveal customer information, internal processes, operating problems, commercial decisions and the organization’s intellectual property.

Removing an employee’s name or replacing it with an identifier does not necessarily resolve these concerns. The content itself may identify the people involved or disclose sensitive information about how the organization operates.

Before implementing an AI monitoring tool, organizations need clear answers to questions such as:

• What information is collected and from which sources?
• Is continuous collection necessary, or could a smaller sample meet the purpose?
• Where is the information processed and stored?
• Does it leave the organization’s systems or jurisdiction?
• Which AI models, vendors and subcontractors receive it?
• Is the information retained by the provider?
• Can it be used to train, test or improve external models?
• Who can access the original material, generated observations and performance reports?
• How long is each form of information retained?
• Can the information be combined with data collected from other systems?
• How will records be corrected or deleted?
• What happens to historical observations if the model, criteria or expected behaviors change?

Information collected to create meeting notes, support coaching or improve service quality should not automatically become performance evidence. Nor should performance data quietly be reused to train an automated replacement, build broader employee profiles or support unrelated workforce decisions.

Consent may not be enough to resolve this issue. Employees may have little practical choice about using a system introduced by their employer.  Organizations must demonstrate that collection is necessary, proportionate, secure and consistent with the stated purpose.

These decisions should be made before data begins to accumulate. Once detailed behavioral information has been collected, it becomes much harder to prevent use expanding beyond its original purpose.

The question is therefore not only, “Can the AI assess this information?” It is also, “Should this information leave the work environment, who gains access to it, and what future uses are we permitting?”

Before AI-generated observations are used in performance management, organizations should be able to answer eight questions:

1. Purpose: What problem are we trying to solve?
2. Necessity: Is monitoring necessary, and is there a less intrusive way to achieve the same result?
3. Relevance: What legitimate work expectation does each observation relate to?
4. Validity: Has the behavior been identified accurately, and does it provide meaningful evidence of performance?
5. Use: Is the information for reflection, coaching, quality assurance or formal assessment?
6. Employee input: Can employees see, correct, explain and challenge the evidence?
7. Information governance: Do we know what is collected, where it goes, who can access it, how long it is retained and whether it can be reused?
8. Ongoing review: Will we monitor accuracy, unintended effects and expansion beyond the original purpose?

The answers should be established before information begins influencing managers’ judgments. Higher-consequence uses require stronger evidence, greater transparency and more rigorous human review.

AI can make aspects of work more visible. Governance determines whether that visibility supports better performance – or becomes surveillance without a sound performance management purpose.

AI can help employees reflect, help managers identify patterns and bring more evidence into performance conversations. But increased visibility is not the same as increased understanding.

An AI system may accurately identify a behavior while misunderstanding its importance, context or cause. Human oversight provides little protection if the manager relies on incomplete evidence or accepts the AI’s interpretation without question.

The central question is not whether AI can observe performance. It is whether the organization can justify turning what it observes into evidence.

Centranum connects ongoing job expectations, goals, organizational values, feedback, competency evidence, development activity and formal reviews within one performance and capability framework.

This gives managers and employees a shared reference point for performance conversations and a traceable evidence history across the review period. It also helps organizations distinguish between role requirements, demonstrated competence, development needs and performance outcomes, rather than reducing them to a single automated signal.

Explore Centranum Performance and Development