AI can summarize a record. It should not become an unaccountable participant in judging someone’s motives, competence, or character.
Work systems contain tempting raw material for automated conclusions: response times, missed dates, estimated and actual effort, message history, review cycles, help requests, and changing assignments. A model can find patterns across that material faster than a person can.
The existence of a pattern does not make every interpretation responsible.
Operational data describes events inside a system. It rarely contains the full human or organizational context required to decide what those events mean about a person.
A fact and a judgment are different products
There is a meaningful difference between saying that a review waited three days and saying that someone is disengaged. Between showing that work exceeded an estimate and concluding that the worker is inefficient. Between identifying repeated help requests and inferring a lack of competence.
The first statement can be grounded in a record. The second introduces an interpretation. The interpretation may omit changing scope, conflicting feedback, an unavailable decision-maker, an unrealistic plan, or help that the organization failed to provide.
AI can make an unsupported interpretation sound unusually complete. Fluent language creates confidence even when the underlying evidence is partial.
This loss of context is not a theoretical edge case. NIST’s AI Risk Management Framework warns that turning complex human and social phenomena into measurable quantities can remove context needed to understand their effects. It also notes that AI can amplify human bias under some conditions and calls for human roles and responsibilities in decision-making to be clearly defined. A model’s ability to calculate consistently does not make the thing being calculated a complete representation of a person.
That is precisely why the boundary must be explicit.
The more consequential the judgment, the less acceptable it is to hide responsibility behind a model.
AI may assist with the record
There are useful roles for AI in work coordination.
It can summarize a long project history. It can identify an unanswered question, a missing dependency, or a conflict between the current plan and a new commitment. It can draft a reviewable plan, surface relevant context, and help someone see what may need attention next.
Each of those roles should remain connected to evidence a person can inspect. The output should be reviewable, correctable, and subordinate to human judgment.
The model can say: these facts appear related. It can ask: does this change the plan? It can propose: here is a draft next step.
It should not quietly convert those facts into a verdict about a person.
Some decisions require an accountable human
Decisions affecting employment, compensation, promotion, discipline, reputation, access, or opportunity require more than pattern recognition. They require someone who can understand context, explain the standard, hear a response, and remain accountable for the consequence.
AI cannot hold responsibility in the human sense. It does not live with the result, repair trust, or answer morally for a mistaken inference.
Placing a human approval button after an automated judgment does not automatically solve the problem. If the system frames the person as high-risk, low-performing, disengaged, or unreliable, that framing shapes the decision before the human begins.
Existing employment guidance reinforces where responsibility belongs. The U.S. Equal Employment Opportunity Commission and Department of Justice have warned that algorithmic employment tools can disadvantage workers and applicants with disabilities. An employer remains responsible for how a tool affects people, including obligations related to reasonable accommodation. The presence of an algorithm does not transfer accountability away from the organization using it.
The responsible design choice may be not to generate the judgment at all.
Surveillance does not become humane when summarized
AI also increases the temptation to collect more information because the system can analyze it.
Keystrokes, screenshots, presence indicators, message sentiment, and inferred attention can be assembled into a detailed portrait of activity. That portrait is not the same as understanding the work.
Knowledge work includes thinking, uncertainty, conversation, revision, and pauses that may be essential to quality. A system optimized to detect visible activity will privilege what it can count and penalize what it cannot see.
Cards should not use AI to create behavioral scores, emotional assessments, productivity rankings, or speculative profiles of workers. It should not make surveillance acceptable by wrapping it in a helpful tone.
Reciprocal visibility is a better standard
If a system surfaces information about a worker’s commitment, it should also preserve the conditions surrounding that commitment.
What changed? What was waiting? What help was requested? Which decisions arrived late? What did leadership add or remove? What did the client need to provide?
This does not guarantee a favorable interpretation. It creates a more complete basis for a fair one.
The person should be able to see the record, correct material errors, and understand how information is used. Hidden scoring and unchallengeable inference have no place in a system built around reciprocal accountability.
Product boundaries are part of the product
Teams building AI features often describe boundaries as limitations. In practice, a clear refusal can be a capability.
It tells customers what the system will protect even when a more invasive feature might be technically possible or commercially tempting. It gives designers and engineers a standard for rejecting ideas before they become normalized. It lets workers understand the role AI is allowed to play in the environment around them.
For Cards, the boundary is straightforward:
- AI may organize and summarize evidence.
- AI may propose a plan or next step for human review.
- AI may surface a conflict, dependency, or unanswered question.
- AI should not infer motives, character, or emotional state.
- AI should not make consequential judgments about a person.
- AI should not become a justification for surveillance.
Assistance should increase human agency
The useful question is not whether AI can produce a conclusion. It is whether producing that conclusion helps people act with more context and responsibility.
Good assistance reduces the burden of finding and carrying information. It gives people a clearer basis for judgment. It preserves the distinction between what happened and what someone believes it means.
AI should help the system remember, connect, and ask better questions. The people affected by the work should remain participants in the truth—not objects of an invisible assessment.
That boundary is not resistance to progress. It is a decision about whom progress should serve.