Most UK small businesses do not have an HR department. They have a director who also does HR, usually on a Sunday evening, usually copying last year's contract and changing the name. That works until someone asks why two people on the same job have different notice periods, or until a new starter's first day arrives with no paperwork ready.
AI is genuinely useful here, but not in the way most vendor pages suggest. It is not a compliance system and it does not know your business. It is a fast, tireless drafter that will produce a confident document with an invented holiday allowance if you let it. This guide covers what to hand it, what to keep, and how to build the checks that make the output safe to issue.
What does "AI HR" actually mean for a small business?
The scale matters for how you think about it. The Department for Business and Trade's business population estimates show that SMEs make up more than 99% of UK private sector businesses and account for around 60% of private sector employment. That is a very large number of employers carrying real legal duties with no HR function at all.
Can AI write a job advert that stays legal in the UK?
Acas advises that terms such as "recent graduate" or "highly experienced" can discriminate on grounds of age, and that words like "barmaid" or "handyman" imply you want to recruit a particular sex. It also recommends defining the job rather than the applicant, giving the example of "German-speaking sales rep" instead of "German sales rep". Acas suggests advertising in more than one place, since online-only advertising can narrow who sees the role.
Practically, that turns into a prompt with a hard rule set: no age proxies, no gendered job titles, no requirement you cannot justify against the actual work, and a note that reasonable adjustments are available at any stage. Then you read it. The model does not know it has just written an age filter.
What should an AI-drafted job description include?
The most useful discipline is to write a short raw brief first, in bullet form, in your own words. Ten scruffy bullets about what the person will actually do produce a far better description than a polished paragraph asking for "a marketing assistant". Ask the model to separate essential requirements from desirable ones and to justify each essential requirement against a task in the brief. Anything it cannot justify usually should not be there, which is the same test Acas applies to advert wording.
Documenting the role properly at this stage also pays off later. The same source material feeds your onboarding plan and, if you keep it current, your process documentation.
Can AI draft an employment contract you can actually issue?
GOV.UK's guidance on employment contracts states that the employer must provide the principal statement on the first day of employment. It must include at least the employer's name, the worker's name, job title or description and start date, pay amount and frequency, "hours and days of work and if and how they may vary", holiday entitlement including public holidays, work location and any relocation possibility, how long the job is expected to last with an end date if fixed-term, probation period length and conditions, other benefits, and "obligatory training, whether or not this is paid for by the employer".
A wider written statement covering pensions, collective agreements, non-compulsory training and disciplinary and grievance procedures is due within two months of the start of employment.
That list is the checklist. Turn it into the review step: paste every item from the GOV.UK list, ask the model to quote back the clause covering each one, and treat any "not present" as a blocker. This is where the time saving is real, because the tedious part of contract work is not the drafting, it is the checking, and a structured check is something a model does well.
What can AI do for onboarding in the first week?
The version that works in a small team is a single onboarding document per role, generated once and reused. Ask for a day-by-day plan for week one with named owners, then a lighter plan for weeks two to four. Include who the new starter should meet and why. Include what "good" looks like at the end of probation, in specific terms, since that is the conversation nobody prepares for and everybody dreads.
Recording those check-ins is also easy to automate. Teams already using AI meeting notes get a written record of probation conversations without anyone typing them up, which matters if a probation outcome is ever questioned.
Where must a human stay in the loop?
There is a practical reason beyond principle. If a decision is challenged, you need someone who can describe how they reached it. "The model ranked them lowest" is not a defence anyone wants to offer. Using AI to summarise application forms into a consistent format is fine. Using it to rank candidates is a different activity with a different risk profile, and small employers gain very little from it.
The same line applies to policy. A model can rewrite your absence policy in plain English. It should not be the thing deciding whether a specific absence was authorised.
How do you handle right to work checks?
The GOV.UK guidance on penalties for employing illegal workers is the authority here, and the protection comes from doing the prescribed check properly before employment starts, using acceptable documents or the Home Office online service. AI has no role in the verification itself.
Where it helps is around the edges: generating the reminder schedule for time-limited permissions, drafting the request email to a new starter, and producing a written record of what was checked and when. The check stays human, the admin around it does not have to.
What does this cost to run, and what should you not pay for?
For a team under about twenty people, the honest position is that a general-purpose assistant plus good templates covers most of it. Dedicated HR platforms start to earn their price when you need holiday tracking, absence records and payroll integration across enough people that spreadsheets break down. Those are database problems, not writing problems, and AI is not what makes that software worth buying.
The saving that does show up reliably is professional time. Paying an employment specialist to review a template you drafted costs less than paying them to draft it from nothing. The same logic applies across the rest of the back office, which is the pattern behind where AI actually cuts admin hours.
How do you build this in an afternoon?
A workable order:
- Write the facts file. Real numbers, real dates, real job titles. One page.
- Paste the principal statement items from GOV.UK into a checklist prompt you reuse for every contract.
- Build the advert prompt with the Acas banned-phrase rules baked in as hard constraints.
- Generate one advert and one statement for a role you have already hired, so you can compare against something you know is right.
- Have an employment specialist review the contract template once. Then reuse it.
Step four is the one people skip and it is the one that catches problems. Testing against a known-good example tells you immediately whether the model is inventing terms.
What are the honest trade-offs?
The second trade-off is drift. If your facts file goes stale, every document generated after that point is stale too, and at speed. A quarterly review of the reference document is not optional maintenance, it is the thing that keeps the whole setup safe.
The third is that fluent writing can mask thin thinking. An AI-written job description reads well whether or not anyone thought hard about the role. That is a management problem the tool cannot solve, and small teams that treat generation as a substitute for deciding what they actually need tend to hire badly, faster.
Used with those limits in mind, the split is clear. AI handles the drafting, the structuring and the checking against a list. You handle the facts, the judgement and the statutory checks. That combination removes most of the Sunday evening paperwork without moving any of the responsibility.
AI HR for Small Teams — FAQ
Is it legal to use AI to write an employment contract in the UK?
Yes. Nothing in UK employment law says a contract must be typed by a human. What the law cares about is the content and the timing. GOV.UK's guidance on employment contracts says the principal statement must be given on the first day of employment, and it lists the items that statement has to cover, including pay, hours, holiday, work location, probation terms and any obligatory training. A wider written statement covering pensions, collective agreements and disciplinary and grievance procedures is due within two months. An AI draft that hits every one of those items is legally identical to a human draft that does. The risk is not the tool, it is issuing a document nobody checked against the statutory list before it went out.
Can AI write a job advert without creating a discrimination risk?
It can draft one, but it should not be the last reader. Acas warns that phrases such as recent graduate or highly experienced can discriminate on age, and that words like barmaid or handyman imply you want a particular sex. Acas also suggests describing the work rather than the person, using German-speaking sales rep instead of German sales rep. Language models are fluent enough to produce all of those phrases without noticing, because those phrases appear in millions of real adverts. The workable pattern is to give the model an explicit banned-phrase list drawn from the Acas recruitment guidance, ask it to flag anything that describes the applicant rather than the job, then read the output yourself before it goes live.
What HR tasks should stay entirely human?
Anything where the output is a decision about a specific person rather than a document. Shortlisting, interview scoring, performance conclusions, disciplinary outcomes and dismissal reasoning all belong to a named human who can explain their reasoning at a tribunal. Right to work verification is another one. GOV.UK is clear that a correctly conducted check is what protects an employer from a civil penalty of up to £60,000 per illegal worker, and that check involves inspecting an acceptable document or using the Home Office online service, not asking a chatbot. Use AI for the writing, the summarising and the chasing. Keep the judgement and the statutory checks with a person who can account for them.
How long does it take to set up an AI HR workflow for a five-person team?
An afternoon for the core of it, if you keep the scope tight. The bulk of the work is not prompting, it is gathering the raw material: your actual pay bands, your holiday year, your probation terms, your notice periods and one job description you were happy with. Once those sit in a single reference document, generating an advert, a principal statement and a first-week onboarding plan is a matter of minutes each. Expect to spend more time on the review checklist than on the generation itself. Teams that skip the reference document end up correcting the same invented detail in every draft, which costs far more than writing the document once.
Does AI replace an HR consultant or an employment solicitor?
No, and framing it that way leads to bad outcomes. AI is strong at first drafts, structure, plain-English rewrites and consistency across documents. It is weak at anything requiring current case law, sector-specific regulation or a judgement call with legal consequences. A sensible split for a small UK team is to use AI for volume work, adverts, onboarding packs, policy rewrites and internal comms, then pay a professional to review your contract template once and to advise on anything contested. That is usually far less billable time than having a professional draft everything from scratch, which is where the real saving sits.