Questions to ask about AI at work.

Why a written AI policy matters.

"We are testing AI" can describe almost anything. It may mean an optional writing assistant, a hiring screen, a system that scores calls, a camera that tracks movement, or software that recommends discipline. Employees cannot give useful feedback or catch hidden risks if the use remains vague.

The U.S. Equal Employment Opportunity Commission's worker guidance notes that automated systems may be used in recruiting, hiring, monitoring time or tasks, measuring performance, and evaluating facial expressions, voice, or movement. Calling all of those systems "productivity tools" hides the different consequences.

A written policy creates a shared record. It lets workers compare the promise with the actual use, gives managers clear limits, and makes it harder for a high-stakes tool to spread through the organization without review.

1. Where is AI being used?

Ask for a list of tools and use cases, including systems purchased from outside vendors. The list should distinguish low-stakes support from consequential decisions.

• Does it draft, summarize, recommend, rank, score, monitor, predict, or make a final decision?
• Is it used in hiring, scheduling, evaluation, promotion, discipline, pay, or termination?
• Are employees required to use it, or can they choose a non-AI process?
• Is the use a time-limited pilot? If so, what result would stop the pilot?

2. What worker data does it collect?

AI features can sit inside software that already handles messages, documents, calls, video, location, keystrokes, or work records. Ask what goes into the system, where it is stored, who can access it, how long it is kept, and whether it trains any model.

Do not accept "the data is anonymized" as the whole answer. Ask what identifiers are removed, whether people can be reidentified through context, and whether a vendor can use workplace content for another purpose.

3. Which decisions can it influence?

A recommendation can shape an outcome even when a manager technically clicks the final button. Ask what the human sees, whether they receive information beyond the score, and how often they reject the automated recommendation.

The EEOC and U.S. Department of Justice have warned that employers may still be responsible when automated tools discriminate against applicants or employees with disabilities. A vendor contract does not transfer the employer's duty to follow employment law.

4. How can a worker challenge an error?

The policy should name a person or office that can review the underlying information and reverse the result. The worker should be able to correct bad data, explain missing context, request an accommodation, and receive a timely answer from a human.

A generic help desk is not enough when the system affects hiring, pay, performance, discipline, or access to work. The appeal route should match the seriousness of the decision.

If a worker cannot challenge the system, the system is not advice. It is authority.

5. What changes for the job?

Efficiency is not a complete workplace plan. Ask which tasks will disappear, which new tasks will be added, who checks the output, and whether the time spent correcting errors is counted as work.

• Will staffing, hours, pay, quotas, or performance targets change?
• Will entry-level work disappear without a new path for learning?
• Who receives training, and is that training paid?
• Who receives the value created by higher output?
• Will employees be disciplined for refusing unsafe or inaccurate output?

6. Who owns the result and the risk?

Ask who is responsible for checking facts, protecting confidential material, resolving copyright questions, and handling harm. If an employee must sign their name to AI-assisted work, they need enough time and authority to verify it.

The organization should also explain what the vendor promised, what independent testing was completed, and what incident would cause the tool to be paused. "The model made a mistake" is not an accountability structure.

A script for the next meeting.

One person asking may be dismissed as resistant. A group bringing the same specific questions is harder to wave away. Keep notes, save policy versions, and document examples where the real use differs from the written rule.

When to get outside help.

If an automated system may have affected hiring, firing, discipline, pay, disability accommodation, or another protected employment decision, consider speaking with a union representative, worker center, employment attorney, or the appropriate government agency. Deadlines can apply to employment claims.

Sources and further reading.

• 01U.S. EEOC: Employment Discrimination and AI for Workers
• 02U.S. EEOC and Department of Justice: Disability discrimination and AI hiring tools
• 03NIST: AI Risk Management Framework Core
• 04International Labour Organization: GenAI, jobs, productivity, and work organization