Drafted by Machine, Signed by a Manager: A Pre-Filing Audit for AI-Assisted Performance Documentation

Generative AI has moved quietly into the manager's drafting workflow. A supervisor short on time may ask a chatbot to turn rough notes into a performance review, a written warning, or a thirty-day improvement plan. The draft often reads polished, balanced, and professional. That polish is precisely why HR should slow down before the document becomes an official record.

The compliance question is not whether the tool is clever. It is whether the document that lands in the personnel file can withstand scrutiny from an employee, a plaintiff's lawyer, a regulator, or a jury. An AI tool can produce language quickly, but it cannot assume the employer's legal obligations, verify its own factual claims, or attest to the accuracy of a manager's judgment. That attestation still belongs to the human who signs the document.

The Record Is the Risk, Not the Prompt

Performance reviews, disciplinary warnings, and improvement plans are business records. They may be used to support promotion decisions, defend against discrimination claims, justify a termination, or rebut an unemployment benefits challenge. In litigation, opposing counsel may seek not only the final document but also earlier drafts, manager notes, prompts, and metadata. A careless AI-assisted drafting process can create a trail of inconsistent statements that undermines the final record.

The manager's signature converts a draft into an attestation. If the manager cannot explain which facts came from personal observation, which came from the AI tool, and which came from an unrelated template, the document's credibility is at risk. HR should treat every AI-assisted performance document as a draft until a reviewer has tested it against the record.

Five Failure Modes to Screen For

1. Coded and Biased Phrasing

AI models learn from large volumes of text, including language that carries stereotypes and coded assumptions. A draft may describe an employee as abrasive, emotional, not a culture fit, or lacking executive presence. Such phrases can function as proxies for protected characteristics and may invite scrutiny under Title VII, the ADA, the ADEA, and state fair employment laws. Reviewers should ask whether the same language would be used for an employee of a different background, gender, age, or disability status.

2. Invented Facts and Borrowed Specifics

Generative tools can hallucinate. A draft may include a project name, a date, a metric, or a client interaction that never occurred. It may also borrow specifics from another employee's review if the prompt included template language or prior examples. Every factual assertion in the final document should trace to a verifiable source: attendance records, project logs, prior reviews, customer feedback, or the manager's contemporaneous notes.

3. Inconsistent Standards Across Managers

Two managers may use the same tool with different prompts and receive dramatically different tones and standards. One review may emphasize missed deadlines; another may ignore the same issue for a different employee. Inconsistent documentation can support a claim of disparate treatment. HR should compare AI-assisted drafts against the employer's rating rubric, prior reviews for the same employee, and documented standards for similar roles.

4. Confidential Data Leakage

Prompts may include medical leave details, accommodation requests, compensation data, customer information, trade secrets, or personally identifiable information. Some consumer AI tools retain inputs or use them for model training. Before any manager pastes employee information into a tool, the employer should confirm that the tool is approved, that the data handling terms are acceptable, and that the prompt contains only the minimum information necessary.

5. Discoverability and Retention

Drafts, prompts, and vendor logs may be discoverable. If the employer cannot produce a coherent record of how a decision was made, the absence may be explained by the plaintiff. HR should establish retention rules for AI-assisted drafts and decide which materials belong in the personnel file. A litigation hold can expand those obligations quickly.

A Pre-Filing Audit Checklist

Before an AI-assisted review, warning, or improvement plan becomes an official record, a reviewer should complete a structured audit:

1. Separate drafting from decision-making. The manager must independently decide the rating, the discipline, and the outcome. The AI tool may assist with wording, not judgment.
2. Verify every factual assertion. Match names, dates, metrics, incidents, and policy references to source documents. Remove anything that cannot be substantiated.
3. Screen for protected-class language. Flag coded terms, stereotypes, references to health or disability, age-related comments, and gendered or racialized phrasing.
4. Test consistency. Compare the draft to the employer's rubric, the employee's prior reviews, and comparable documentation for similar conduct.
5. Review data handling. Confirm the tool is approved, the prompt contains no unnecessary personal data, and the vendor's retention terms are understood.
6. Document provenance. Record that AI assistance was used, which tool was used, and what human review occurred. Keep the record factual and consistent with the employer's policy.
7. Obtain manager attestation. The manager should confirm in writing that the final document is accurate and reflects personal knowledge.
8. Preserve the employee's response. If the document is shared with the employee, follow the employer's usual process for comments, rebuttal, or acknowledgment.

Policy Guardrails for 2026

Employers adopting AI-assisted performance documentation should pair the checklist with clear policy controls. A useful policy identifies approved tools, prohibits entering sensitive personal data into unapproved tools, requires human review before filing, sets retention rules for drafts and prompts, and explains how the employer will monitor compliance. Training should cover prompt hygiene, bias screening, factual verification, and the difference between drafting assistance and decision-making. Periodic sampling of AI-assisted documents can reveal calibration drift before it becomes a pattern.

Vendor contracts deserve attention as well. Employers should understand whether the tool retains inputs, whether outputs may be used for model training, and how the vendor responds to legal process. These terms affect both confidentiality and discoverability.

The Bottom Line

AI can reduce the time managers spend drafting performance documentation, but it cannot reduce the employer's accountability for the result. The safest approach is disciplined and unglamorous: treat every AI-assisted document as a draft, verify the facts, screen the language, test for consistency, control the data, and require a human attestation before the document reaches the personnel file. The algorithm may write the first version, but the employer owns the record.