An HR consultancy was mid-way through a harassment investigation when it discovered the employer’s handbook had no anti-harassment policy. Not a weak one; none. The handbook had been drafted by ChatGPT, and nobody had checked what the model left out. Carly Holm, the consultancy’s CEO, put the consequence plainly: “If the workplace does not have appropriate policies in place like a zero tolerance policy for sexual harassment, workplace violence, etc, the investigation will then look at the employer, and there will be consequences.”
That story, reported by Futurism, is the shape of AI for employment lawyers in 2026. The same tools that let you audit a handbook in an afternoon are producing the handbooks you will be paid to fix, and employers deploying AI need a new category of advice. Seven prompts below cover both sides. Every one assumes a no-training tier, an anonymised employee, and a lawyer who opens each cited statute before the client sees it.
What employment lawyers automate first
Spellbook names first-pass employment agreements from approved templates, compliance monitoring for “Minimum wage changes, Non-compete restrictions, Paid leave requirements”, severance and release review, policy drafting, and compensation-consistency checks across salary, bonus and severance formulas. Thomson Reuters’ CoCounsel team describes three workflows: an executive-agreement review for IRC § 409A, a handbook compliance audit, and a wage-and-hour survey for remote work. Anthropic’s open-source Claude for Legal repository ships an Employment plugin with a termination-review command.
All of these are review and extraction tasks against a standard you supply; none is “what does the law say”. That distinction is the safety model, and the contract review guide explains why playbook-based review works where open research fails. The termination review is the simplest first prompt.
Act as a [state] employment lawyer advising the employer. We plan to terminate [EMPLOYEE], a [role] who [recently filed a workers' compensation claim / requested leave / raised a complaint]. From the anonymised facts and documents below only, list: the protected characteristics or activities engaged; timeline facts that support or undercut a retaliation inference; documentation gaps to close before any decision; and the questions to ask HR. No conclusions on the merits, no case citations.
[paste]Handbook audits with a Topic / Coverage / Section / Recommendations table
CoCounsel’s handbook audit for a nonprofit returns a four-column table, a format worth copying in any tool because it turns review into spot-checking. Give the model the checklist, not the law.
Act as a [state] employment lawyer advising the employer. Audit the handbook below against this checklist, which I wrote and which is your only source of legal requirements: <checklist>[paste]</checklist>. Output first a table with exactly these columns: Topic | Coverage (Present / Partial / Missing) | Section | Recommendation. Then list every outdated or unlawful provision, quoting the words and the checklist item breached. Add no requirements of your own; anything you think the checklist misses goes under "COUNSEL TO CONFIRM".
Handbook:
[paste]If you let the model add legal requirements from memory, you have converted a review task into a research task, and general models “may hallucinate laws” and lack “state-specific labor code awareness”, in Spellbook’s own words.
Severance and release drafting: the OWBPA prompt
The severance prompt circulating in prompt libraries is short: “Draft a severance agreement for a [state] employee at [level], including release of claims, [X weeks] severance, non-disparagement, and the OWBPA language required for employees over 40.” It works as a skeleton and fails as a finished product: the consideration and revocation periods, the state-specific carve-outs and the deferred-compensation interactions are exactly what a general model gets subtly wrong.
Draft a severance agreement for a [state] employee at [level] (anonymised as [EMPLOYEE]) with [X weeks] severance, a general release, non-disparagement and return-of-property terms, employer side, following our precedent: <precedent>[paste]</precedent>. The employee is over 40: include the OWBPA elements and mark every consideration period, revocation period and statutory reference [VERIFY]. List separately any claim [state] law does not allow to be released. Invent no dates, amounts or plan names; leave [BRACKETS].Executive agreements: 409A triggers and six-month delays
CoCounsel’s executive-agreement review looks for three things: severance trigger dates, the six-month delay for specified employees, and acceleration clauses. A reading task with a fixed target is where a model is at its best.
Review the attached executive employment agreement for IRC § 409A issues, employer side. Extract into a table: Provision | Clause reference | Payment trigger and timing (quote the words) | Specified-employee six-month delay addressed? (Yes/No/Unclear) | Acceleration or discretion creating a 409A risk | Recommended fix. Treat "Unclear" as a finding, not a pass. Cite nothing beyond § 409A itself and tag every regulatory reference [VERIFY].Multi-state wage-and-hour surveys for remote work
A permanent remote-work policy for an employer with staff in a dozen states needs a survey of overtime classification, meal and rest breaks and record-keeping rules. Jurisdiction matters most here and models flatten it most, so the prompt asks for a confidence column and the states to check first.
Build a table surveying wage-and-hour rules for a remote-work policy across [list states]: State | Overtime classification rule | Meal and rest break rule | Record-keeping requirement | Governing provision [VERIFY] | Notable 2024-2026 change [VERIFY] | Confidence (High/Medium/Low). Where you are not confident, say so in the confidence column rather than filling the cell. Then name the five states where the law has most likely changed recently, so I verify those first.Every cell gets checked against primary law before it reaches a policy; the citation verification protocol budgets a few minutes per authority, so a twelve-state table is an afternoon.
The client-side risk: the handbook with no anti-harassment policy
The Humani story has a UK twin in the same report: a company drafted a severance agreement with Copilot, its lawyers rejected it because essential terms were missing, and the settlement cost more than it should have. A model asked to “draft a handbook” or “draft a severance agreement” produces something that looks complete, and a non-lawyer cannot see what is absent.
Advising employers on their own AI: FLSA, NLRA, GDPR
The HR Defense Blog lists three issues employers miss when they roll out AI tools:
| Issue | Trigger | What the advice covers |
|---|---|---|
| FLSA reclassification | Employees now “monitor, verify, or supplement AI-generated work”, changing the duties that supported an exemption | Re-run the duties test for affected roles before the tool changes the job |
| NLRA exposure | An AI productivity tool “scans internal emails, chat messages” and captures protected concerted activity | Scope the tool’s data access; carve out channels; train managers |
| GDPR and CCPA | Employee personal data entered into prompts, or processed by a vendor without a lawful basis or processor terms | Impact assessment, vendor terms, an AI-use policy with a no-personal-data rule |
The AI-use policy is the deliverable most clients ask for, and it is a drafting task from a checklist, not a research task.
Draft the AI-use section of an employee handbook for a [500-employee] employer in [jurisdiction], using only these legal minimums, which I wrote: <requirements>[paste]</requirements>, and these business rules: <rules>[paste]</rules>. Plain English at a new-hire reading level, numbered paragraphs, a defined-terms table, a "questions?" contact line, no legal citations in the body. Then a table mapping each paragraph to the requirement or rule it satisfies. Add no legal requirements of your own.Add the evidential point to every client conversation: chatbot histories are discoverable, and in United States v. Heppner a federal court held a defendant’s consumer-Claude exchanges neither privileged nor work product.
A German Fachanwalt’s 50% time saving on an appeal brief
Martin Lorentz is a Fachanwalt für Arbeitsrecht in a five-lawyer firm, not a digital native, and dictates rather than types. From July 2025 he ran roughly 60 files through the legal AI tool Silvernova. For a labour-law appeal brief the tool extracted the parties, summarised the file and pulled together the Bundesarbeitsgericht case law on workplace disruption against freedom of expression. His verdict, in the Anwaltsblatt: “Meine Zeitersparnis lag bei mindestens 50 Prozent.” His rules: he uses AI only in areas where he is himself the expert, and he checks every citation against a public database. “KI ordnet für mich die Sachverhalte, stellt sie zusammen”: the model orders the facts; the lawyer decides.
Verification and jurisdiction traps
Employment is the fourth-largest subject area in Damien Charlotin’s hallucination database, with 185 of 2,039 decisions as of 12 September 2026. The case employment lawyers should know is Mid Central Operating Engineers Health and Welfare Fund v. HoosierVac (S.D. Ind., 28 May 2025): three briefs with non-existent citations in an ERISA-adjacent dispute, a $6,000 sanction reduced from a recommended $15,000.
The quieter trap is jurisdiction. Clio’s prompt-engineering guidance puts it well: “You’ve been thinking about Texas employment law all morning. The AI hasn’t.” Ask “Is this non-compete enforceable?” without a state and you get whichever law dominates the training data. The seventh prompt refuses to let the model start without one.
Apply only [state] law as at [date]. Assess the enforceability of the non-compete in Section [7.2] of the attached agreement for a [role, seniority, access to confidential information] at a business in [industry]. Structure: the governing test, statute and leading case tagged [VERIFY]; each element applied to the facts; the reformation or blue-pencil position; the status of any federal or state non-compete rule [VERIFY: date-sensitive]; the three facts that would change the answer; a narrower, likely-enforceable clause. Where [state] authority does not support a step, write "NO VERIFIABLE AUTHORITY FOUND".Where to go next: the contract drafting guide shows why whole-document prompts leave gaps, the in-house guide covers the legal department’s side, and the practice-area hub puts employment next to the other fields. More prompts of this kind are in the prompt library and the ChatGPT prompts guide; in AI Lab for Lawyers, employment lawyers build the handbook-audit prompt against their own state’s law in the first session.
Frequently asked questions
Can AI draft an employee handbook?
It can draft handbook sections quickly, and that is the problem: a ChatGPT-drafted handbook for one employer omitted the anti-harassment policy entirely, discovered only when a harassment complaint was being investigated. The safe method is to give the model a compliance checklist written by counsel, require a coverage table before the text, and treat every legal requirement it adds on its own as unverified until you have checked the statute.
Is it safe to use ChatGPT for severance agreements?
Only in a no-training tier with the employee anonymised, and only for structure. A prompt naming the state, the employee's level, the weeks of severance and the release can produce a usable first draft, but the OWBPA language for employees over 40, state-specific release carve-outs and any 409A interaction must be verified by you. A UK employer's Copilot-drafted severance agreement was missing essential terms and cost more at settlement.
How do employment lawyers use AI?
Thomson Reuters describes three CoCounsel workflows that cover most of the practice: executive-agreement review for IRC 409A triggers and the six-month delay for specified employees, handbook compliance audits output as a Topic / Coverage / Section / Recommendations table, and surveys of federal and state wage-and-hour rules for remote-work policies. Spellbook adds first-pass agreements from templates, compliance monitoring and compensation-consistency checks.
What AI risks should I warn employer clients about?
Three, according to the HR Defense Blog: wage-and-hour reclassification if employees 'monitor, verify, or supplement AI-generated work' in ways that change their duties; NLRA exposure if an AI productivity tool scans internal emails or chat messages and captures protected concerted activity; and GDPR or CCPA breaches when employee data goes into prompts. Add the evidential point: a chatbot history is discoverable and not privileged.
Can AI check a handbook for compliance?
Yes, as a first pass against a checklist you supply, not as the checker of record. Ask for a table with Topic, Coverage, Section and Recommendation columns, then a list of provisions that are outdated, unlawful or missing with the controlling statute tagged for verification. General models lack state-specific labour-code awareness, so every citation gets opened before the client sees the audit.