hr.cafe
Share your story
Workplace AI AI and changing work Practical people operations Manager support Moderate-High — Growing neurodivergent assistive-AI requests vs. corporate DLP bans

When an Employee Requests ChatGPT or Copilot as an ADA Accommodation: HR Runbook

Your Acceptable Use Policy bans generative AI, and an engineer with dyslexia and ADHD asks for ChatGPT to draft client reports. Blanket policy denials violate the ADA—run this zero-retention interactive process instead.

When an Employee Requests ChatGPT or Copilot as an ADA Accommodation: HR Runbook — hr.cafe field guide banner
Controlling authorities: 42 U.S.C. § 12112(b)(5)(A) (ADA Reasonable Accommodation Mandate) · 29 CFR § 1630.2(o)(2)(ii) & (o)(3) (Policy Modification & Interactive Process) · 42 U.S.C. § 12111(10) & 29 CFR § 1630.2(p) (Undue Hardship Standard) · 29 CFR § 1630.2(n) (Essential Job Functions Boundary) Portable tool: ADA Assistive-AI Security Substitution Matrix & 30-Day Trial Agreement
Jurisdiction
US Federal · ADA Title I (42 U.S.C. § 12112; 29 CFR § 1630.2(o)) · UK Equality Act 2010 s.20
Employer threshold
15+ employees (US ADA Title I) · Day-1 interactive duty
Decision trigger
Employee with dyslexia, ADHD, autism, low vision, or TBI requests ChatGPT, Claude, or Copilot
Review cadence
Reviewed on EEOC assistive-AI guidance, JAN updates, or federal ADA rulings
Direct Operational Answer

Direct operational answer

Employers with 15 or more employees cannot summarily deny a disability request to use generative AI by pointing to a company-wide "no AI" policy. Under Title I of the Americans with Disabilities Act, 42 U.S.C. § 12112(b)(5)(A), and 29 CFR § 1630.2(o)(3), HR must run an individualized interactive process—yet the employer retains full legal authority to reject insecure consumer accounts and substitute a zero-data-retention enterprise seat or local assistive tool under the hr.cafe four-tier security matrix.

Why "InfoSec Blocked AI for Everyone" Fails Under 42 U.S.C. § 12112

A neutral Acceptable Use Policy does not override federal disability law. Closing an accommodation ticket with a copy-pasted IT ban hands the EEOC an open-and-shut interactive-process violation.

Look at the exact ticket landing in people operations queues across mid-market firms. Six months ago, your security team blocked `chatgpt.com`, `claude.ai`, and `gemini.google.com` at the corporate firewall after a vendor scare. On Monday, a senior cybersecurity consultant—or a clinical project manager—submits a clinician note documenting dyslexia and ADHD. Reading 40-page client RFPs and structuring blank-page written summaries takes her twice as long due to mechanical decoding and working-memory friction. She asks HR for an exception to use ChatGPT Plus on her work laptop.

HR forwards the ticket to IT Security. IT replies in four minutes: "Denied. Corporate Policy SEC-14 bans external LLMs for all staff."

Stop before relaying that refusal. Under the Americans with Disabilities Act (42 U.S.C. § 12112(b)(5)(A)) and 29 CFR § 1630.2(o)(2)(ii), modifying a neutral workplace policy is a textbook reasonable accommodation—the exact principle confirmed by the Supreme Court in US Airways, Inc. v. Barnett, 535 U.S. 391 (2002). Just as a "no animals in the building" rule yields to a service dog and a "no recording devices" rule yields to deaf captioning, a blanket "no generative AI" rule requires individualized assessment under 29 CFR § 1630.2(o)(3).

Check the unit economics too. Multi-year employer survey cohorts published by the U.S. Department of Labor's Job Accommodation Network (JAN) covering more than 3,000 workplace accommodations show that roughly half of all accommodations cost $0, and among accommodations carrying an outlay, the median one-time or annual cost sits around $300. Provisioning a single $30-per-month ($360-per-year) zero-data-retention enterprise AI license—or a $150–$250 local assistive literacy tool—never approaches financial undue hardship under 42 U.S.C. § 12111(10).

Employer Choice: Saying No to Consumer ChatGPT and Yes to Zero-Retention Tools

The ADA entitles a worker to effective barrier removal, not their preferred consumer brand. InfoSec keeps full veto power over public model-training pipelines when HR offers a secure equivalent.

Here is the rule that brings your CISO onto HR's side in five minutes. Under long-standing EEOC Enforcement Guidance on Reasonable Accommodation (Question 9), when more than one accommodation is effective, the employer has the ultimate discretion to choose between effective accommodations, selecting the option that is less burdensome, easier to administer, or safer for company data.

Allowing an employee to paste unredacted client source code, HIPAA patient charts, ITAR export-controlled schematics, or privileged legal memos into a free consumer chatbot creates genuine operational and statutory undue hardship. You have every right to say no to a personal consumer ChatGPT account—provided you diagnose the exact functional barrier and say yes to a secure tool that solves it.

Decompose the employee's request by functional barrier rather than software brand:

  • Barrier A — Mechanical Writing, Spelling & Blank-Page Structuring (Dyslexia / Dysgraphia): The employee understands the technical subject matter completely, yet translating spoken or bulleted thoughts into polished prose takes double the hours. Secure fix: an isolated single-seat enterprise LLM license (Microsoft 365 Copilot inside your existing E3/E5 tenant, ChatGPT Enterprise/Team with training contractually disabled, or Grammarly Enterprise with zero-retention DLP) or local dictation-plus-rephrasing software.
  • Barrier B — Dense Text Decoding & Reading Fatigue (Dyslexia / Low Vision / TBI): Processing 60-page PDFs causes visual tracking fatigue. Secure fix: local on-device text-to-speech readers (such as Texthelp Read&Write in offline/local mode or native OS Spoken Content) paired with an enterprise document summarizer inside your existing cloud storage perimeter.
  • Barrier C — Auditory Working Memory & Task Sequencing (ADHD / Autism): Multi-speaker video calls overload working memory while taking manual notes, or large ambiguous projects trigger executive-function stall. Secure fix: internal zero-retention meeting recaps (restricted to approved internal calls) plus structured task-breakdown software containing zero customer PII.

Essential Job Functions vs. Assistive Scaffolding

Providing an assistive drafting tool never excuses inaccurate work product. Keep human verification and professional judgment squarely inside the employee's essential functions.

Line managers frequently push back during manager support intakes: "If I let one analyst use Copilot while the rest of the team writes reports manually, isn't that lowering performance standards or giving an unfair edge?"

Draw the clean statutory boundary under 29 CFR § 1630.2(n) before issuing approval:

  1. You Never Eliminate an Essential Job Function: If the role's core purpose is auditing tax citations, certifying structural load math, or editing final regulatory filings for factual precision, the employee—not the AI—remains 100% accountable for accuracy, citations, and professional judgment. Generative AI removes the mechanical input/output bottleneck (just like speech-to-text dictation removes typing barriers for an employee with severe carpal tunnel); it does not lower production quality standards.
  2. Hallucinations Remain Performance Accountable: Specify in the written accommodation letter that submitting unverified AI-fabricated facts, fake citations, or unreviewed client deliverables violates standard work-quality expectations on the exact same terms applied to every peer.
  3. Protect Approval Continuity Across Manager Changes: Once HR and InfoSec approve a provisioned assistive license, log it in your central accommodation register so a future reorganization does not trigger the trap covered in our new-manager accommodation revocation guide.
Portable worksheet · hr.cafe

Four-Tier ADA Assistive-AI & InfoSec Substitution Matrix

Security Environment Employee's Raw Request Approved Secure ADA Substitute InfoSec / DLP Guardrail
Tier 1: Standard Corporate SaaS Seat
(Internal business & general client work)
Personal ChatGPT Plus or free Claude web tab for drafting & summarizing. Single-seat enterprise license (M365 Copilot, ChatGPT Team/Enterprise, or Claude Enterprise) under corporate SSO. Zero-Data-Retention + No-Training DPA; disable external plugin/web sharing; ~$30/mo HR accommodation budget.
Tier 2: HIPAA / Regulated Financial Seat
(PHI, NPI, or payroll records)
AI assistant to structure clinical case notes or compliance summaries. BAA-covered enterprise AI tenant (e.g., HIPAA-configured Azure OpenAI / enterprise clinical scribe) OR local offline dictation/TTS. Executed HIPAA BAA + RBAC folder scoping; block pasting SSNs/PHI into non-BAA chat windows.
Tier 3: Export-Controlled / Air-Gapped Seat
(ITAR, EAR, SCIF, or defense contracts)
Cloud LLM for dyslexic report editing or code commenting. 100% on-device local assistive software (offline Dragon, local Read&Write, or self-hosted air-gapped GovCloud model). Zero external internet egress; passes DFARS / NIST SP 800-171 boundary audit.
Tier 4: Multi-Party External Call Restriction
(Client or HR calls in all-party states)
AI meeting bot to record and summarize external client calls for ADHD memory. Local live OS closed captioning (no audio stored) + 15-minute protected post-call buffer block to dictate structured notes into enterprise ZDR tool. Prevents unconsented third-party wiretapping (Cal. Penal Code § 632) while solving working-memory barrier.

Copy-Ready 30-Day ADA Assistive-AI Trial & Security Agreement

Issue a 30-day provisional trial rather than debating edge cases for two months. A bounded written agreement satisfies the ADA interactive process while locking InfoSec rules in ink.

CONFIDENTIAL MEDICAL / ADA ACCOMMODATION RECORD (SEGREGATED FILE — 29 CFR § 1630.14(c))
Employee: [Name & Role]                  Accommodation ID: [ADA-YYYY-NNN]
Effective Date: [YYYY-MM-DD]             30-Day Effectiveness Check-In: [Date + 30 Days]

1. Approved Assistive Accommodation (Policy Exception to AUP Section [X])
   To address functional barriers documented during our interactive review, the Company is provisioning a single-seat license to [Approved Tool, e.g., Microsoft 365 Copilot / Zero-Retention Enterprise AI / Local Assistive Suite] at Company expense, paired with [e.g., native OS text-to-speech & 15-minute post-meeting note buffers]. Because this secure enterprise tool provides effective support, your request to use a personal consumer ChatGPT/Claude account on Company data is declined for data-security compliance.

2. Mandatory Information-Security Boundaries
   - Use ONLY the Company-provisioned SSO account above; personal or free browser AI tools remain blocked.
   - Do NOT enter [restricted data carve-outs, e.g., third-party export-controlled schematics / unredacted SSNs] outside approved [BAA/GovCloud] workspaces, and do NOT connect AI recording bots to external client calls without prior all-party consent.

3. Accuracy & Essential Job Functions
   This tool supports mechanical drafting, reading, and organization only. Reviewing, verifying, and standing behind the factual accuracy, calculations, and professional quality of all work deliverables remains an essential function of your role.

4. Confidentiality & Continuity
   Supervisors are informed only of the approved software access—never your underlying diagnosis. At Day 30, HR will check in with you to confirm whether this tool effectively removes the workplace barrier or needs adjustment.

HR Accommodations Lead: _____________________   Employee Acknowledgement: _____________________

Aligning HR, InfoSec, and Line Supervisors

Accommodation breakdowns happen when IT auto-revokes single-seat licenses during quarterly license sweeps or supervisors gossip about why one employee got Copilot. Lock role rules across AI and changing work.

HR Accommodations Lead

Focus medical documentation strictly on functional barriers (reading, drafting, working memory), fund single-seat licenses from a central ADA budget rather than team P&Ls, and run the Day-30 check-in.

InfoSec / CISO Team

Pre-approve one Tier-1 enterprise zero-retention AI SKU and one Tier-3 offline local literacy suite so HR can fulfill ADA tickets in 72 hours without custom firewall battles.

Line Manager

Evaluate final deliverable quality identically across peers, never disclose to coworkers that a teammate's AI seat is a disability accommodation (42 U.S.C. § 12112(d)), and never revoke IT access unilaterally.

Primary statutes, regulations, and enforcement dockets

  1. 42 U.S.C. § 12112 & 29 CFR § 1630.2(o) — Americans with Disabilities Act Reasonable Accommodation Regulations — Defines reasonable accommodation to include acquisition of assistive equipment/software and modification of workplace policies, alongside the mandatory interactive process.
  2. EEOC Enforcement Guidance on Reasonable Accommodation and Undue Hardship Under the ADA — Confirms employer obligation to modify neutral workplace policies absent undue hardship and affirms employer discretion to choose among effective accommodations (Question 9).
  3. U.S. Department of Labor / Job Accommodation Network (JAN) — Workplace Accommodations: Low Cost, High Impact — Empirical multi-year employer survey data establishing low direct costs of assistive software accommodations and practical guidance on dyslexia, ADHD, and executive-function tools.
  4. U.S. Department of Justice & EEOC — ADA.gov Assistive Technology & Effective Communication Resources — Federal technical standards on real-time captioning, text-to-speech, speech-to-text, and accessible workplace communication tools.

Practitioner questions on the edge cases

If coworkers ask why one analyst has a Copilot license when everyone else was told AI is banned, what can the manager say?

Under 42 U.S.C. § 12112(d) and EEOC Enforcement Guidance (Question 42), a manager must never reveal that an employee received software as a disability accommodation. Coach managers to give a neutral privacy-preserving reply: "We do not discuss individual administrative or workspace configurations; if you have a specific workflow request, please submit it through standard IT/HR channels."

Can we require limited medical documentation when an employee asks for an AI tool due to ADHD or dyslexia?

Yes, when the disability and functional need are not obvious. Keep the request narrow: ask a qualified professional to confirm only the nature of the functional limitation (for example, mechanical written expression or auditory working memory) and how assistive drafting or summarizing software addresses that barrier—never demand lifelong school IEPs or full psychiatric records.

What if our company does not have an enterprise AI contract and vendors require a 50-seat minimum?

Check actual vendor tiers before claiming undue hardship: major providers offer self-serve business/team tiers starting at 2 to 5 seats (roughly $60 to $150/month total) with model training contractually disabled, and cloud providers (AWS Bedrock, Azure OpenAI, Google Vertex) allow zero-retention single-user deployments for pennies per query, alongside zero-cloud local assistive suites like Read&Write or Dragon.

Does UK law treat assistive AI accommodation requests similarly for London staff?

Yes. Under Section 20 of the UK Equality Act 2010, employers have a positive legal duty to make "reasonable adjustments"—explicitly including providing auxiliary aids (such as assistive software) and adjusting provisions, criteria, or practices (such as a blanket AI ban) where they place a disabled worker at a substantial disadvantage.

Get the Thursday edition

One email a week with new practitioner stories and primary-sourced HR playbooks. Free, and one click to unsubscribe.

Weekly on Thursdays. Your email is used for nothing else. Privacy.

Keep reading

Where this decision connects to neighboring statutory clocks and workflows.

All field guides
Workplace AI US Multi-State (CA / NY / NJ / IL / CO / WA)

"Meaningful Human Review" as a Record: Building an Auditable AI Override Log

Clicking "Approve All" on 400 AI-ranked rejections in twelve minutes is rubber-stamping, not oversight. Build a verifiable override-and-evidence log before 2027 state and EU mandates hit.

8 min

First-person practitioner stories on these topics

Field Note

The investigation finding I got wrong

A detailed, consistent account felt like enough. One question from my manager changed the finding, and two months later it mattered.

HR business partner (example) 2 min read
Esc