An AI HR assistant for a UK small business: policy questions, onboarding and paperwork, with the lines it must not cross.

01

What it does

Policy questions: the assistant reads the staff handbook, the policies and the procedures, answers in plain English with the paragraph it relied on, and says “ask HR” when the handbook does not cover it. Onboarding: from a template and the new starter’s details, it drafts the offer letter, the welcome pack, the first-week plan and the list of documents to collect, and chases the ones that do not arrive. Correspondence: reference requests, standard letters, meeting invitations, drafted for a person to check and send. Meeting notes for one-to-ones and reviews, within the transcription rules. Each is reading and drafting; none is deciding.

RelatedA knowledge assistant on your own documents →AI meeting notes and transcription: the UK rules →Fixed-price AI automation for UK SMEs →

02

The lines it must not cross

Decisions about individuals: who to interview, who to promote, whose performance is poor, who to dismiss. Under Articles 22A to 22D a significant automated decision needs safeguards, and the ICO’s 2026 recruitment findings expect meaningful human involvement, bias testing and a route to challenge; the practical rule for a small firm is that the assistant never scores, ranks or recommends about a named person. Legal advice: the assistant explains the firm’s policy, not employment law, and says so. Data: it reads the handbook and, when asked by an individual, that individual’s own leave balance or record, never another person’s, and never health, disciplinary or grievance files unless a specific process with a DPIA allows it.

RelatedAI and employment law for UK employers →

03

The set-up

Two routes. Inside a business plan, a shared project or custom GPT with the handbook and policies attached answers questions for staff at the cost of seats, with the permissions of the plan; it suits a firm whose handbook is current and whose questions are mostly policy. A built knowledge assistant on the API, integrated with the HR system for leave balances and documents, with sources shown, access scoped per user and an audit trail, is a first implementation at £4,000 to £15,000; it suits a firm that wants the assistant to answer “how much leave do I have” from the system rather than “here is the leave policy”. Either way the handbook has to be true first: an assistant over an out-of-date handbook confidently gives out-of-date answers.

RelatedClaude Team plan for a UK business →AI acceptable-use policy template →

04

The rules to write down

Seven lines, adopted before go-live and kept with the policy.

1. 01

   The assistant answers from the handbook and policies; anything it cannot find goes to a named person

2. 02

   It never scores, ranks, recommends or decides about a named individual

3. 03

   It explains policy, not law, and says so

4. 04

   It reads an individual’s own record only for that individual, and never health, disciplinary or grievance files outside a DPIA-covered process

5. 05

   Drafts are checked by a person before they are sent

6. 06

   Conversations have a retention period and are not used for performance management

7. 07

   The handbook is reviewed before go-live and every six months

RelatedA DPIA for AI, with a template →

05

What to measure

Questions answered without escalation, questions escalated, hours saved by the person who used to answer them, time to complete onboarding paperwork, and a monthly sample of fifty answers checked against the handbook. If the sample shows wrong answers, the handbook is usually the cause.

RelatedHow to measure AI return in 90 days →How to check AI output before you rely on it →