Treat an AI-use objection as an accommodation case, not a policy fight.

One accommodation makes the issue concrete, not settled

On October 26, 2026, the American Humanist Association said the University of Kentucky had approved a research administrator's request to be excused from personally using AI because of his Humanist beliefs. The public account is a useful operating example. It is not a court decision, a federal agency ruling, or a universal right to refuse every AI-connected workflow.

The association says David Durman requested an accommodation in late August, met with the university in mid-September, and received approval on September 25. Independent reporting by Religion News Service adds an important boundary: he was still expected to attend meetings and cooperate with colleagues who use AI or submit materials to systems that use it, while being excused from personally using AI to complete his tasks. That distinction is exactly where many employer reviews will succeed or fail.

A request framed as “no AI” can hide several different activities: entering a prompt, reviewing model output, completing AI training, using a product with an embedded assistant, working with a colleague who uses AI, uploading material that a downstream system analyzes, or accepting a decision partly informed by AI. HR cannot evaluate effectiveness or hardship until it separates those actions. Neither employee nor manager shorthand should substitute for a workflow map.

Under U.S. federal law, Title VII defines religion to include aspects of religious observance, practice, and belief unless an employer demonstrates it cannot reasonably accommodate them without undue hardship. In Groff v. DeJoy, the Supreme Court rejected the long-used “more than de minimis” shorthand and said undue hardship requires substantially increased costs in relation to the conduct of the particular business, viewed in overall context. Current EEOC material reflects that standard.

This guide is a workflow, not legal advice. State and local law, public-sector rules, collective bargaining, contracts, and other accommodation frameworks may add duties or stronger protections. Use qualified counsel for the applicable standard and preserve a neutral, confidential process.

Map the action before judging the request

Start with the employee's own words. A request does not need a statute number or the phrase “reasonable accommodation.” The practical trigger is notice that a work requirement conflicts with a sincerely held religious belief, observance, or practice. A manager who hears that conflict should route it to the accommodation process rather than debate the belief or demand immediate compliance.

Next distinguish an essential function from a preferred method. A research administrator may need to prepare a compliant budget, summarize terms, and meet a deadline. “Use the approved chatbot” may be a method selected for efficiency, not the job's essential purpose. Conversely, a role created specifically to evaluate generative models may make direct interaction central. Use actual duty evidence: job records, frequency, consequences, time spent, business purpose, prior methods, and how peers perform the work.

Do not promise a boundary that technology cannot enforce. A word processor may add AI features after an update. A ticketing system may summarize text downstream. An email security service may use machine learning. Ask system owners to map data flows, feature flags, logs, permissions, and fallbacks. Describe the approved boundary precisely—such as “no personal use of generative AI to draft or transform work”—instead of an untestable promise that the employee's work will never touch any AI system.

Use a cooperative exchange, not a one-shot form

EEOC guidance describes an information exchange in which employer and employee cooperate to identify the conflict and an effective accommodation. Treat the employee's proposed option as the starting point, not necessarily the only option. The employer may choose another effective accommodation, but effectiveness must be real and the choice should not punish the employee for making the request.

1. Acknowledge. Confirm receipt, case owner, confidentiality, anti-retaliation route, interim measure, and next date.
2. Clarify the conflict. Identify each personal action the employee says conflicts and which surrounding activities remain possible.
3. Verify duties. Ask the manager for the business objective, essential function, frequency, deadline, quality needs, and non-AI history.
4. Map systems. Have IT or the process owner verify feature controls and downstream AI processing.
5. Develop alternatives. Test manual work, feature disablement, task swaps, alternative training, review changes, or another effective route.
6. Measure burden. Record actual or supportable cost, operational impact, mitigations, and the employer's context.
7. Apply human gates. HR, counsel, business owner, technical owner, and authorized decision maker review the same frozen packet.
8. Implement and revisit. Verify access and work allocation, then monitor effectiveness and material changes.

Questions about sincerity require discipline. EEOC materials say beliefs are generally presumed sincere, while a limited inquiry may be appropriate when the employer has an objective basis for questioning religious nature or sincerity. Inconsistency can be relevant, but beliefs can change, and uncommon or unfamiliar beliefs are not automatically insincere. Record the objective reason before follow-up, limit the request to what is reasonably needed, and avoid theological judgment.

Temporary measures matter. If a mandatory training or tool rollout arrives before the review can finish, consider pausing the disputed action while preserving the employee's pay, access, and duties where feasible. An interim arrangement is not a final approval. It is a way to prevent the process itself from forcing the conduct or triggering retaliation.

Compare options with evidence, not intuition

The cost packet should distinguish a number from an adjective. “Too expensive,” “inefficient,” and “the team will resent it” are not measurements. Record one-time and recurring cost, the work actually shifted, schedule effect, service or safety impact, mitigation, and the employer's scale and operating context. Under Groff, impacts on coworkers matter when they affect the conduct of the business; hostility to the religious practice or the accommodation itself cannot count.

Consider combinations. An AI-disabled account plus an alternative training module and a limited task swap may be effective where no single control works. Ask whether an option preserves opportunity, credit, compensation, schedule, access to information, and evaluation under neutral performance standards. An accommodation should not quietly isolate the employee or remove developmental work without review.

Worked example: procurement analyst asked to use a drafting assistant

A procurement analyst is told that all contract summaries must be drafted in an approved generative AI assistant. The employee says personal generation conflicts with a Humanist belief about authorship and environmental responsibility. The employee can read source documents, write summaries manually, attend meetings, and work with colleagues who use AI, but asks not to prompt or edit generated text.

1. Intake. HR records the request in the employee's words, tells the manager not to enforce the new prompting requirement while the review is open, and limits case access.
2. Duty analysis. The essential output is an accurate, timely contract summary. Direct use of the assistant was introduced two weeks ago as a standard method; it is not in the role description and the prior manual process remains available.
3. Technical map. IT confirms the assistant can be hidden for the employee. The contract repository separately runs automated classification after upload, which cannot currently be disabled per user. HR returns to the employee to clarify whether downstream classification creates a conflict.
4. Options. The employee proposes manual summaries. The manager proposes that a colleague generate a draft for the employee to check. The employee says review of generated prose also conflicts. The team tests manual work with the same source, template, and quality review.
5. Evidence. Across ten representative contracts, manual completion takes an additional 22 minutes on average and meets the deadline and error threshold. The team can absorb the measured difference without overtime or service failure. The AI-generated route still needs 14 minutes of fact checking.
6. Decision. The employer approves manual drafting, alternative AI-policy training without hands-on generation, and continued use of the repository after disclosing its downstream classifier. The arrangement does not prevent meetings or ordinary collaboration with AI-using colleagues.
7. Implementation. IT disables the assistant, the manager receives neutral performance instructions, the employee confirms access, and HR schedules a 30-day check.

The result turns on the evidence, not the label “AI exemption.” A different role, system, requested boundary, operational burden, or governing law could produce a different effective option or decision. The case record should explain that reasoning without declaring the employee's worldview correct or incorrect.

Write the arrangement so it can be implemented

A decision letter is not an implementation plan. The case record should name covered actions, duties that remain, systems, feature states, alternative method, schedule, quality and security checks, training, owners, effective date, review date, change triggers, incident route, confidentiality, and reconsideration path. It should state what is outside scope to prevent a narrow request from becoming either an accidental blanket exemption or an accidental denial.

case: RA-AI-2026-041
approved_boundary:
  - no personal prompting of approved generative assistant
  - no editing of assistant-generated contract summaries
continued_duties:
  - read source contracts
  - draft manual summaries using approved template
  - meet existing accuracy and deadline standards
  - collaborate with team members regardless of their tool use
systems:
  drafting_assistant: disabled_and_verified
  contract_classifier: active_downstream_disclosed
controls:
  quality_review: unchanged
  performance_credit: equal
  confidentiality: need_to_know
review_at: 2026-11-30
reopen_on: [job_change, tool_change, data_flow_change, effectiveness_issue]
decision_owner: named_hr_authority

If the request is denied, document the specific option assessed, the evidence of substantial burden under the applicable standard, alternatives and mitigations considered, why they were ineffective, the approving authority, and the reconsideration route. Do not use a generic template that says “business needs” or cites inconvenience without the underlying facts.

Failure modes worth checking before closure

Monitor effectiveness and neutrality

Check the arrangement after implementation, not only at annual review. Confirm that the disputed AI action remains disabled or reassigned, essential duties are completed, quality and safety remain within standards, actual cost matches the estimate, transferred work is sustainable, and the employee has not lost credit, schedule access, training, advancement, or ordinary collaboration.

• Effectiveness: employee confirms the arrangement addresses the stated conflict.
• Work: deadlines, quality, safety, and service outcomes are measured under the agreed method.
• Implementation: feature flags, permissions, training, and task assignments match the written plan.
• Neutrality: performance ratings, opportunities, pay, workload, and manager conduct show no retaliatory pattern.
• Burden: actual cost and operational impact are compared with the original estimate.
• Change: new tools, embedded AI features, duties, data flows, or legal guidance trigger review.

Close the administrative case only when the arrangement is implemented, the employee knows the contact and reconsideration route, and required records are complete. An active accommodation can remain in force after case administration closes. Reopen it when material conditions change or either party reports that the arrangement is no longer effective.

FAQ

Does the Kentucky case mean every employee can refuse AI?

No. It is a reported employer accommodation involving particular beliefs, duties, and boundaries. It does not create court precedent or decide other employers' cases.

May HR require proof from a religious organization?

Not as a universal rule. Title VII protects sincerely held beliefs that may be unfamiliar or not part of an organized religion. When an objective basis supports limited follow-up, request only reasonably necessary information and use qualified guidance.

Can we require the employee to work with colleagues who use AI?

Possibly, depending on the request and facts. The reported Kentucky arrangement preserved meetings and cooperation with AI-using colleagues. Clarify the employee's actual conflict and test whether the proposed boundary is effective.

Is extra coworker work automatically undue hardship?

No. Measure how work affects the conduct of the business in the employer's overall context and consider alternatives. Coworker hostility toward the belief or accommodation is not a valid burden.

Should this process be placed inside the general AI policy?

The AI policy should point to the existing accommodation and anti-retaliation process. Keep individualized belief information in restricted accommodation records, not the general AI-use log.