How HR Consultants Use AI for Policies, Letters and Cases

Use AI to adapt your approved policy templates to each client's facts, to fill letter templates from case details, and to organise case evidence into chronologies and interview summaries. Keep decisions, legal checks and final wording with you, put as little personal data in as the task allows, and use a business plan with a separate project for each client.

Employment law is where AI's confident tone does most damage. It will write a policy that borrows a statutory entitlement from the wrong jurisdiction, or a dismissal letter that skips a step your client's procedure requires, and both read perfectly. The consultant's value, and what clients pay for, is knowing which sentence is wrong.

Follow me on Instagram@sagnikteaches

Three kinds of work, three levels of risk

Policies, letters and cases use AI in different ways and carry different risks. Treat them separately:

Connect on LinkedInSagnik Bhattacharya

The pattern: the lower the personal data, the more you can hand over. For policies, AI can do most of the drafting. For cases, it organises and you judge.

Subscribe on YouTube@codingliquids

Policies: from your master template to a client's version

Don't ask AI to write policies from nothing. Keep a master library of policies you've drafted and checked, and use AI to adapt them. The adaptation is where the hours go, and it's also where AI is most reliable, because the legal substance comes from your template.

1. Keep the master library current. Review each master when the law changes, not when a client asks. Date each one.
2. Build a client fact sheet once per client, and keep it in their project.
3. Adapt with a tight prompt that forbids changes to legal substance.
4. Compare the result with the master using Word's Compare feature, so you see every change the AI made rather than trusting a summary.
5. Check the operational fit with the client: can their managers actually do what the policy says?

A filled-in fact sheet for an illustrative sports equipment retailer:

CLIENT FACT SHEET - [retailer]                         Updated: [date]
Staff: 28 (2 shops, 1 small warehouse). 11 part-time, 6 seasonal in winter.
Managers: owner, 2 shop managers, 1 warehouse lead. No in-house HR.
Hours: 7 days; Sunday trading; late opening Thursdays.
Pay: hourly plus sales commission for shop staff (monthly).
Systems: rota app, shared email, no HR system. Payroll by outside bureau.
Known issues: commission disputes; lateness on weekend shifts; staff
  using personal phones for store social media accounts.
Tone wanted: friendly, first names, short sentences.
Existing documents: 2019 handbook (to be replaced), contracts from template.

The adaptation prompt:

Adapt the MASTER POLICY below for the client described in the FACT SHEET.
You may change: examples, job titles, who-to-contact details, tone and
length, and add practical examples that fit this business.
You must not change: any entitlement, time limit, procedural step, right of
appeal or legal statement in the master. If the fact sheet suggests one of
these should change, write [CONSULTANT: ...] instead of changing it.
Mark every change you make with a comment in square brackets.

An illustrative extract of what comes back for a lateness and absence policy:

If you're going to be late or can't come in, call your shop manager (not a
text or message) at least one hour before your shift starts. [CHANGED: master
says "your line manager"; fact sheet says shop managers run rotas.]
Weekend shifts start at 8.30am, so call by 7.30am. [ADDED: practical example.]
Commission is not paid for any period of unauthorised absence. [ADDED]

What needs changing: the first two changes are good. The third is a problem the prompt was meant to prevent. The AI saw "commission disputes" in the fact sheet and invented a pay rule that isn't in the master, sits outside a lateness policy and may conflict with the staff's contracts. Delete it and add a [CONSULTANT] note to raise commission rules with the owner separately. Compare mode is what caught it; a quick read might not have.

The same setup makes annual reviews for retained clients quicker. Paste in the client's current handbook contents page and your master library index, and ask for three lists: policies the client has that your master also covers, policies your master covers that the client lacks, and client policies with no master equivalent. For a new client arriving with an old handbook, that comparison typically surfaces missing items (a hybrid working policy, a social media policy that covers staff running the shop's accounts) in minutes rather than an afternoon of reading. The judgement about which gaps matter for a 28-person retailer is still yours, and so is the conversation about priorities and fees.

An illustrative result for the retailer, run against its 2019 handbook:

IN BOTH: disciplinary, grievance, absence and lateness, holidays,
  equal opportunities, health and safety.
IN MASTER, NOT IN CLIENT HANDBOOK: social media, hybrid working,
  staff personal data, bullying and harassment (client has it inside
  grievance), expenses.
CLIENT ONLY, NO MASTER: uniform and staff discount; staff party conduct.

Reading it against the fact sheet sets the order of work. Social media goes first, because staff running the shop accounts from personal phones is a live issue. Hybrid working barely applies to shop and warehouse staff, so it drops down the list. Bullying and harassment deserves its own policy rather than a paragraph inside grievance. The two client-only policies need your eye before they're kept, since nobody has checked them since 2019.

Letters: fixed templates filled with facts

For letters, the rule is simple: AI fills a template, it never decides the template. The procedural content of an invitation to a disciplinary hearing (the allegation, the right to be accompanied, the evidence enclosed, the possible outcomes) is fixed by your template and the client's procedure. The AI's job is to insert the facts clearly and in the right tone.

Here is what the AI adds, from rough note to letter. The manager at an illustrative food truck business sends this:

need to get [employee] in about the till being short 3 times, about 40
missing each time, last one sat 14th, want to do it thursday

The consultant pastes in the investigation-meeting template plus the facts (with the amounts checked against the till reports). The illustrative filled letter reads:

We are looking into cash discrepancies at the till on three occasions in the
last month, the most recent on Saturday 14 [month]. On each occasion the till
was short by approximately [amount]. This is an investigation meeting, not a
disciplinary hearing, and no decision has been made. We would like to meet
you on Thursday [date] at [time] at [place] to hear your account...

What the AI did well: it turned "get [employee] in" into a neutral investigation meeting and added "no decision has been made", which your template requires. What you check: the dates and amounts against the till reports (the manager's "about 40" becomes the exact figure from each report), the meeting details, and that the letter doesn't name the employee as the cause. Never let the AI draft from the manager's words alone; "the till being short" is fine, "[employee] taking money" would be an allegation made before any investigation.

Outcome letters carry a different risk: the AI writes reasons, and reasons can grow. Suppose the same case went to a hearing and the consultant's decision notes read:

Allegation 1 (till short on three dates): upheld. Till reports and rota
  show [employee] sole cashier each time.
Allegation 2 (till left unattended on the 14th): not upheld; CCTV unclear.
Mitigation considered: new to cash handling, no training record on file.
Sanction: first written warning, live for 12 months.
Appeal: in writing within [x] working days to [owner].

An illustrative extract from the AI's draft: "Having taken into account your previous conduct and the seriousness of the matter, and noting that the till was also left unattended on the 14th, we have decided to issue a first written warning." Two faults, both in one sentence. "Previous conduct" appears nowhere in the notes, and a letter that cites it invites the question of what conduct that was. The unattended till was the allegation that was not upheld, yet the draft reads as if it counted against the employee. The fixed version lists each allegation with its outcome, gives the mitigation in the employee's favour, and states the sanction and appeal route exactly as noted. The check that catches this is mechanical: every reason in the letter must point to a line in your decision notes.

Cases: chronologies and interview summaries

Casework is where AI saves the most time and needs the most care. The case used here is illustrative: a grievance at a catering company with 40 staff, where a chef alleges bullying by the head chef over six months. The consultant has 14 documents (emails, rotas, two earlier informal complaints, WhatsApp screenshots) and notes from three interviews.

Building the chronology

Extract the text from each document (retype screenshots if needed), label each with a document number, and use a prompt like:

From the documents below, build a chronology table with columns:
Date | Event | Who was involved | Document number.
Include only events stated in the documents. Quote short phrases where the
wording matters. If a date is uncertain, say so. Do not assess credibility,
motive or whether any event amounts to bullying.

What would take three to four hours by hand takes the AI minutes and you an hour to check, row by row, against the documents. Check especially that events haven't been merged, that "said" and "alleged" survive, and that dates from message screenshots are the dates sent, not the dates screenshotted.

An illustrative extract of the first draft, and what the check turned up:

Date              | Event                                    | Who                 | Doc
2 [month]         | Rota change: claimant moved to four late  | Head chef, claimant | D3
                  | shifts, note says "cover for holidays"    |                     |
"early spring"    | First informal complaint to owner         | Claimant, owner     | D6
(uncertain)       |                                           |                     |
18 [month]        | Head chef shouted across the pass; later  | Head chef, claimant | D9
                  | emailed "sorry about earlier"             |                     |
27 [month]        | WhatsApp: "don't bother coming in Monday" | Head chef           | D12

Three rows needed work. The 18th row merges two documents: the shouting comes from the claimant's written complaint (D9), while the apology email is D10, sent the following morning, and the draft dropped the word "alleged" in the process. The rewrite splits it into two rows, the first reading "Claimant alleges head chef shouted across the pass". The WhatsApp message carries the date the screenshot was taken; the timestamp inside the image shows the 21st, which puts it before a rota meeting rather than after it. The uncertain date is handled correctly, and it goes on your list of questions for the owner.

Summarising interviews

Ask for a neutral summary of each witness's account under fixed headings: what they saw, what they heard, what they did, what they say they don't know. Instruct the AI not to characterise tone ("she was evasive") or credibility. Those are findings, and findings are yours.

An illustrative summary from the same grievance, with the line that had to go:

WITNESS 2 - kitchen porter, interviewed [date]
Saw: head chef and claimant talking at the pass during evening service;
  could not see faces clearly from the wash-up area.
Heard: raised voices "for a minute or so"; could not make out words.
Did: carried on working; mentioned it to the sous chef at the end of shift.
Doesn't know: what the conversation was about, or whether it had
  happened before.
The witness seemed reluctant to say anything critical of the head chef.

The last sentence isn't in the interview notes. It's an inference about credibility, and if the summary were disclosed it would read as the investigator's view of a witness. Delete it. Then check every quoted phrase ("for a minute or so") against your notes word for word, because a quotation is the part a claimant or their representative will test first.

Handling the data

Grievance and disciplinary files often contain special category data under data-protection law such as the GDPR (health, sexual orientation, religion, trade union membership and similar). Include only what the task needs: a chronology of rota changes doesn't need the claimant's medical history. The food truck case shows how health data arrives uninvited: the manager's follow-up message adds that [employee] "has been off a lot lately, something about anxiety". The till investigation letter needs none of that, so it stays out of the prompt. It still matters. It may change how the meeting is run, whether adjustments are worth discussing, and what you advise the client next, so it goes into your case notes and your call with the owner rather than into the chat that fills the letter. Where a document mentions third parties who aren't part of the case, replace their names with roles, using the routine in how to anonymise client data before you paste it into AI. And remember that AI-generated summaries kept on file are notes like any other: they can fall within an employee's request to access their personal data. For long bundles, how to summarise bundles and transcripts safely covers the mechanics.

Where AI must stay out of the process

• The decision. Whether an allegation is upheld, what sanction fits, whether to dismiss. If any of your clients operate in the EU, the EU AI Act lists AI used to make decisions about promotion, termination, task allocation or monitoring of workers among its high-risk uses. A consultant using AI to draft words around a human decision is in a very different position from a tool that recommends the decision, and it should stay that way.
• Credibility. Never ask the AI which witness is more believable.
• Legal currency. AI's knowledge of employment law may be out of date or from the wrong jurisdiction. Every legal statement is checked against current law or by an employment lawyer.
• Anything the employee will see without your read. No letter goes out on the AI's draft alone.

The master-template rule comes from mistakes like this illustrative one. A consultant asked an AI tool to "draft a flexible working policy for a small employer" from scratch rather than from her master. The draft looked polished and was sent to the client for comment. It included a response deadline and a limit on requests per year that came from a different country's rules. The client's operations manager, who had just read an official guidance page, asked why the numbers differed. No harm done, but a credibility dent that the master-template method would have avoided.

Keeping many clients' data apart

An HR consultancy holds staff data for every client, which makes separation the main safeguard:

• One project per client in ChatGPT Business or Claude Team, holding that client's fact sheet, handbook and templates, and nothing else. See how to use ChatGPT Projects to keep client work separate.
• Casework in the case file, not in chat history. Copy the finished chronology and summaries into your case folder, then delete the working chats when the case closes, in line with your retention policy.
• Business plans only. They don't train on your content by default; consumer accounts rely on each user switching training off.
• Your terms. Say what tools you use and what data goes in. Some clients will want to approve them.

What it does to a consultant's week

For an illustrative solo HR consultant with 15 retained small-business clients, the rough timings with this setup:

• A client handbook refresh: from about three days to one and a half, most of the saving in adaptation and formatting; the legal review takes as long as before.
• A routine letter: from 30 minutes to about 10, including the check.
• A grievance chronology from a dozen documents: from three or four hours to about one and a half.

That's time for two or three more retained clients, or for the advisory conversations clients value most. For the drafting side of employment documents in more depth, see how to write an employee handbook with AI and employment contracts with AI: the legal checks.

Questions HR consultants ask about AI on casework

Can I use an AI note-taker in a disciplinary or grievance meeting?

Only if everyone present knows and agrees at the start, including any companion, and the recording and transcript are stored with the case file under the same rules as handwritten notes. Many consultants still take their own notes and use AI only to tidy them afterwards. A transcript is disclosable in the same way as other notes, so check it for accuracy before it goes on file.

Should my clients know I use AI?

Yes. Put a short statement in your terms of engagement: which tools you use, that they're business plans that don't train on client content, what data goes in, and that all advice and documents are reviewed by you. Clients with strict data rules may want to approve the tools, and it's far better to have that conversation at the start than after a case.

Can AI tell me whether a dismissal would be fair?

It can list the questions that usually decide fairness, and it can check whether your file covers each step of the client's procedure. It can't weigh the facts, the evidence and current law the way you or an employment lawyer can, and it will give an answer confidently either way. Use it as a completeness check on the file, never as the opinion.

Further reads

• How to Answer Staff HR Questions With an AI Helpdesk — Letting clients' staff get policy answers without calling you.
• How to Run Staff Surveys and Exit Interviews With AI Analysis — Analysing surveys and exit interviews for retained clients.
• AI Bias in Small Business Decisions: Hiring, Pricing and Credit — How bias creeps into AI-assisted people decisions.
• How to Set Up Claude Projects as a Shared Team Assistant — Setting up one project per client with your templates inside.
• Which Legal Tasks Should a Small Firm Never Hand to AI? — The same never-delegate list, seen from a law firm.
• Who Owns AI Chat History When an Employee Leaves? — Who owns AI chat history, relevant when case notes live in chats.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: EU AI Act Annex III point 4 (employment and worker management); GDPR Article 9 (special category data); OpenAI and Anthropic business plan data terms.