AI for HR: Where It Helps a Small Team, and Where It Bites

The most useful number I found while writing this comes from a CIPD survey of just over a thousand UK HR professionals: around four fifths of SMEs use no AI or machine learning at all in their recruitment process. Whatever LinkedIn is telling you this week, you are not late.

The owners who ask me about AI for HR almost never have an HR department. They have eighteen staff and a handbook sitting in a Google Drive folder that nobody has opened since 2023. Their Sunday evening goes on writing a contract for a new starter who begins on Wednesday. What they want is not a transformation of the people function. They want two hours of their week back without accidentally buying a tribunal claim.

So this is that version. Which HR jobs are genuinely safe to hand over, which ones AI should draft and you should sign, the two I would not automate at any size, what actually changed in UK law in February 2026, and real prices in pounds from the vendors’ own pages. Where I could not verify something, I say so.

Where does AI actually help with HR?

On the jobs where being wrong costs you a rewrite rather than a tribunal. That one test sorts nearly every use case you will be pitched.

Drafting an onboarding pack, answering the same holiday carry over question for the ninth time this quarter, summarising an exit interview, building a checklist for a new starter. Get any of those wrong and you notice within a day and fix it in ten minutes. Now hold that next to software that ranks applicants and sends rejections while nobody reads them. Same underlying technology. Completely different downside.

The timing argument is stronger than it was two years ago, and not because the models got clever. It is because UK employers picked up more admin. From 6 April 2026 statutory sick pay runs from the first day of illness and the lower earnings limit is gone, paternity and ordinary parental leave became rights from day one, and employers must keep annual leave and holiday pay records for at least six years. In October 2026 the tribunal time limit doubled from three months to six, and the duty to prevent sexual harassment strengthened to taking all reasonable steps, with liability for third party harassment alongside it.

Read that list as an owner rather than a lawyer. More records, kept longer, defensible for twice as long, in a business where the person responsible for all of it is also the person doing sales.

What does the AI in HR software actually do?

There are two completely different products being sold under the same word, and no vendor page I read makes the difference obvious.

The first kind answers from your documents. You upload your own handbook and contracts, and staff ask it questions. Sense HR calls this Counsel and says it reads your live contracts, policies and rota and comes back in about three seconds. BambooHR’s version, Ask BambooHR, covers handbook and benefits documents but only from its Pro tier at $17 per employee per month upward, while the $10 Core tier answers holiday questions and nothing more. Breathe has an AI Assistant that answers from your own company policies, though as of August 2026 it sits in early access and does not appear on the pricing page, so I cannot tell you what it costs.

The second kind answers from the vendor’s content. BrightHR’s BrAInbox will tell you how statutory sick pay works or how to read a tax code, drawing on Bright’s own expert answer bank rather than on anything you uploaded. That is a useful product. It is also a different product, and at £16.67 per employee per month before tax, on a fixed term of 24, 36 or 60 months, you should know which one you are buying.

Here is the buying question that follows from that, and it takes ten seconds to ask on a demo call. If I upload my own handbook, will this answer questions from it? If the answer involves the word “roadmap”, you are buying the second kind.

For completeness: I found no AI features at all on the CharlieHR or Homerun pricing pages, and none on Zelt’s. Cheap and honest beats an AI badge on a page.

Which HR jobs can you hand over now?

The ones where the output is a draft you will read anyway. Six that earn their keep in a small team:

1. Onboarding packs per new starter. Same template, different name, role, manager, start date and equipment list. This is a merge job with a bit of judgement, which is exactly the shape AI is good at.
2. Answering repeat policy questions. How much carry over am I allowed, what is the notice period for holiday, who approves unpaid leave. If four people ask the same thing a month, that question belongs to a tool that reads your handbook.
3. Structured interview scorecards. Ask for the five criteria that actually predict success in the role, then score every candidate against the same ones. Consistency is both a better hiring signal and a better defence.
4. Exit interview summaries. Turn a rambling forty minute conversation into themes you can act on, then compare six of them for the pattern you have been avoiding.
5. Chasing. Missing timesheets, unsigned contracts, outstanding right to work documents, expiring visas. Nothing here needs intelligence. It needs something that does not forget.
6. Payroll preparation. Collating hours, expenses and absences into the shape your payroll provider wants. AI assembles it. A human still checks it before submission, because errors here reach staff bank accounts and HMRC.

One caution on the drafting work, and it is the best cautionary tale in this whole field. Forbes reported in May 2024 that an HR consultant asked a New York client for its employee handbook during a live harassment investigation. The client admitted, sheepishly, that ChatGPT had written it. There was no anti harassment policy in it at all. Nobody had read the thing they had adopted. In the UK from 30 October 2026 that gap is not just embarrassing, it is the exact evidence a tribunal looks for when testing whether you took all reasonable steps.

Which HR jobs should AI draft and you sign?

Anything that ends in a decision about a named person. AI produces the draft, a human owns the outcome, and the human involvement has to be real rather than a rubber stamp.

CV sifting sits here. So do performance review notes, disciplinary paperwork, redundancy selection scoring and contract variations. The tempting move is to let the tool do the first cut on 200 applicants and only look at the top twenty. The problem is that the cut is the decision, and you have just made it without seeing 180 people.

Amazon is the case to remember, because it is not a story about carelessness. Amazon built an automated CV screener trained on ten years of its own applications, most of them from men. The model learned to downgrade CVs containing the word “women’s”, as in women’s chess club captain, and to mark down graduates of two all women’s colleges. Its engineers neutralised those specific terms, then scrapped the project in 2018 anyway, because they could no longer be confident it was not discriminating in some other way they had not thought to test for. Amazon had a machine learning team and still could not audit its own tool. You have a laptop and a busy Thursday.

The UK marker is Pa Edrissa Manjang, a Black courier who delivered for Uber Eats in Oxfordshire. Facial recognition identity checks repeatedly failed to recognise him although nothing about his appearance had changed, and Microsoft has acknowledged that its face software performed less well for ethnic minorities. He was given no explanation of the process that suspended him and no real way to challenge it. He filed a tribunal claim in October 2021 for indirect race discrimination, harassment and victimisation, backed by the Equality and Human Rights Commission and his union. Uber settled in March 2024 before a hearing listed for seventeen days, so nobody ever established what went wrong. The employer wore the liability, and the settlement bought silence rather than an answer.

What should you never automate in HR?

Two things, and I would hold this line at any headcount.

The first is a rejection nobody reads. Sending an automated no, based on a score, with no human looking at the pile, is the one move that converts a time saving into a legal exposure. It is also unnecessary. A person scanning a ranked list for twenty minutes keeps you outside the automated decision rules completely.

The second is inferring how someone feels. AI video interviews that score sentiment or emotion, and biometric attendance systems, are a different category of risk from a badly worded job advert. In February 2024 the ICO ordered Serco Leisure and associated trusts to stop using facial recognition and fingerprint scanning to monitor attendance for more than 2,000 employees across 38 sites, and to destroy the data. The reasoning was simple: staff were not proactively offered an alternative, and the power imbalance between employer and worker meant consent was unlikely to be freely given. If you hire anyone based in the EU, note that the AI Act’s ban on inferring emotions in workplace settings has applied since 2 February 2025.

Right to work belongs in its own bracket. AI cannot perform the check and cannot give you a statutory excuse, and the penalties rose on 13 February 2024 to a maximum of £45,000 per worker for a first breach. Let it run the checklist and chase the documents. The check itself is yours.

What does UK law say about AI in hiring now?

It changed, and in the direction almost nobody expects. Section 80 of the Data (Use and Access) Act 2025 replaced Article 22 of the UK GDPR with new Articles 22A to 22D, and that section commenced in February 2026.

The old rule treated solely automated decisions with significant effects as prohibited unless an exception applied. The new Article 22C permits them for ordinary personal data provided you put safeguards around them: give the person information about the decision, let them make representations, offer human intervention, and let them contest it. Special category data is still restricted under Article 22B, which matters because health data and ethnicity live there.

The definition worth memorising is in Article 22A. A decision is “based solely on automated processing” where “there is no meaningful human involvement in the taking of the decision”. That phrase is your escape hatch and your trap. Real review, outside the rules. A manager approving a ranked list without opening it, still inside them.

Before you buy any tool that touches applicants, the ICO published exactly the checklist you need. Its audits of AI recruitment providers, run from August 2023 to May 2024 and reported in November 2024, produced 296 recommendations and 42 advisory notes. Auditors found tools inferring gender and ethnicity from candidates’ names, and tools that let recruiters filter out applicants by protected characteristic.

Ask a vendor those six on a demo call. The good ones have answers ready. The rest send you a case study.

Do you need HR software, and what does it cost in the UK?

Below roughly five staff, no. A contract template and a shared calendar cover it, and you will spend longer configuring a platform than it ever saves you. Somewhere between five and fifteen people the holiday requests and the document chaos start eating real hours, and at that point the UK flat fee tools are close to trivially cheap.

The pricing shape matters more than the headline number. Some vendors charge per business, so ten staff and fourteen staff cost the same. Others charge per employee per month, which is fine at twelve people and less fun at sixty.

Watch the minimums, because they break the headline maths. Employment Hero starts at £4 per employee a month but requires ten users. Zelt’s £5 needs twenty seats. BambooHR’s floor means a ten person UK firm pays the same $250 as a twenty five person one. And BrightHR’s own FAQ mentions prices starting from £3.00 per person while no purchasable plan on that page comes close, which tells you something about how these deals are actually sold.

What could you build yourself instead?

For the single highest value job on the list, answering staff policy questions, quite a lot, and possibly for nothing.

If you already pay for Google Workspace, NotebookLM has been a core service with enterprise data protection since February 2025, meaning your uploads and queries are not used to train models and do not leave your organisation’s boundary. Drop your handbook, your contracts, your policy documents and last year’s offer letters into a notebook, share it with the team, and you have the searchable staff Q and A that vendors charge per employee per month for. On Microsoft, the Copilot add on is £13.80 per user per month on an annual commitment as a promotional rate, £16.10 regularly, before VAT.

If you want the job done properly rather than approximately, this is where Claude Code earns its £16 or so a month. A working example from my own patch: a script that takes a new starter’s details and generates the whole pack from your reviewed templates, then files it and diarises the probation review. Another that reads a folder of signed contracts and flags every one missing a term you now include as standard. You own both outright, they cost nothing per employee, and when you hire your fortieth person the price does not move. If you want a fast read on which of your own processes would pay back first, the AI Opportunity Scorecard takes about two minutes.

The honest catch is the same as everywhere else on this site. You own what you build, including the Monday morning it stops working.

Where I would start on Monday

If I ran an eighteen person firm, I would not buy anything this month. I would take every policy document I own, put the lot somewhere a tool can read it, and answer the next fortnight of staff questions from there. It costs nothing, it exposes how out of date your documents are, and that discovery is worth more than any software you could buy on top of them.

Then I would automate the chasing, because it is pure friction with no judgement in it, and leave every decision about a named person exactly where it is now. The reason I push back on candidate ranking so hard is not squeamishness about the technology. It is that the time saving is small and the exposure is not, and I have never met an owner who wanted to explain an algorithm to a tribunal.

If you would rather have someone go through your actual processes and tell you which two are worth automating first, that is what the AI Opportunity Audit is for. It is a fixed fee 60 minute session after a free 15 minute intro call, we find the highest impact places to apply AI automation in your business, build one live in Claude Code together, and you leave with a prioritised roadmap and the method. If you just want to know how I work before any of that, it is on my about page.