HRM Info All articles
HR Compliance

The Draft Before the Record: A Generative AI Data-Handling Review for Lean Teams

HRM Info
The Draft Before the Record: A Generative AI Data-Handling Review for Lean Teams

Photo: Constanceofficiel, CC BY-SA 4.0, via Wikimedia Commons

Generative AI has entered small-business HR through the side door: a manager drafting a job posting, a supervisor summarizing a performance conversation, an administrator softening a leave denial. Each use can save time. Each can also move employee names, medical details, complaints, pay data, and customer information into a tool that no one approved. For a lean employer, the answer is not a blanket ban. It is a short, repeatable review that separates safe drafting from risky disclosure and leaves a record of the final decision.

Map the Work Before You List the Tools

Start with use cases, not vendors. Ask each manager and HR-adjacent role to name the last five tasks for which they used a generative AI tool. Group the answers by function: recruiting, onboarding, performance management, employee relations, benefits and leave, payroll support, and customer communication. For each task, note the data that entered the tool. A job posting draft may contain only a title and public company description. A performance summary may contain names, ratings, medical accommodations, and allegations. The second example requires a different control than the first.

Sort Data Into Three Practical Tiers

A simple classification scheme is easier to enforce than a long policy. Use three tiers and publish examples.

Red data should never enter an unapproved generative AI tool. This includes Social Security numbers, bank and payroll details, medical or leave records, accommodation requests, workers compensation information, background check results, investigation notes, complaint allegations, disciplinary drafts, and any document from a personnel file. It also includes trade secrets, unpublished financials, customer lists, and credentials. If the tool is a free consumer product that may train on inputs, treat red data as a public disclosure.

Amber data requires an approved tool and human review before use. This might include de-identified performance themes, engagement survey comments with names removed, or draft customer responses that contain no personal data. Amber work should use a business account with contractual protections, not a personal login.

Green data can be drafted in most approved tools. Public job descriptions, general policy language, training outlines, and generic email templates fit here, provided the final text is checked against current company policy.

Approve the Tool and the Task Together

A tool approval alone is too broad. A task approval alone is too narrow. Use a one-page request that captures the tool name, business owner, task, data tier, vendor terms, retention settings, and whether inputs are used for model training. Prefer business versions with administrative controls and a data processing agreement. If the vendor cannot confirm that employee data will not train the model and will be deleted on a set schedule, the tool is not ready for amber or red work.

Keep a register of approved tools. Include the approval date, approved uses, prohibited uses, and the person who can answer questions. Review the register when a vendor changes terms or a team adopts a new feature. A chat tool with a new file-upload button is a new data pathway, even if the logo has not changed.

Put a Review Station Between Draft and File

Generative AI output should be treated as a draft from an enthusiastic assistant who has not read the personnel file. Before any text becomes an official record, a human should verify facts, remove invented details, check for bias or discriminatory language, and confirm that the draft does not promise something the company cannot deliver. For performance warnings, discipline, leave decisions, and accommodation responses, use a second reviewer. The reviewer should compare the draft to source documents and the applicable policy.

Recordkeeping matters. Save the final approved version in the official HR or business file. Do not rely on a chat history as the record. If the tool retains prompts, assume the prompt may be discoverable. Write prompts with the same care as an email to counsel. Avoid pasting entire documents when a summary of the issue will do.

Train Managers on Three Habits

Training does not need a half-day seminar. It needs repetition and examples. Teach three habits: pause before pasting, verify before sending, and file the final version. Use real scenarios from the business, with names removed. Show how a harmless request for a performance note can become a red-data event when medical details are included. Explain that deleting a chat does not necessarily delete the data from the vendor. Ask managers to route uncertain requests to a named contact rather than guessing.

Write a Policy That Fits on One Page

The policy should state its purpose, who it covers, which tools are approved, what data is prohibited, how to request a new use, who reviews outputs, how records are kept, and how to report a mistake. Add an acknowledgment for employees who use AI in HR-related work. Date the policy and review it at least annually, and after any new tool, vendor term change, or reported incident. A short policy that managers can find and follow is stronger than a long policy no one reads.

Document the Review and Set a Date to Repeat

Keep an audit worksheet with columns for task, data tier, tool, approval status, reviewer, and next review date. In the first pass, focus on the five highest-risk uses rather than every prompt. Correct shadow use by offering an approved alternative, not only by prohibiting the old one. Set a repeat date for six months out. Generative AI tools change quickly, and a control that worked last quarter may not match the feature released this month.

For small employers, the objective is proportionate control: let people draft faster, but keep confidential data out of unapproved windows and keep the final record where it belongs. This article provides a management framework, not legal advice; employers should confirm obligations under applicable state and federal law with counsel.

All Articles

Related Articles

Snow on the Lot, Questions in the Ledger: A Weather-Closure Pay Framework for Small Employers

Snow on the Lot, Questions in the Ledger: A Weather-Closure Pay Framework for Small Employers

When the Business Expense Lands on the Employee: A Reimbursement Audit for Lean Teams

When the Business Expense Lands on the Employee: A Reimbursement Audit for Lean Teams

Counting Before You Cross: A Headcount Threshold Audit for Small Employers

Counting Before You Cross: A Headcount Threshold Audit for Small Employers