AI shouldn't decide anything about people's jobs, sign off numbers that can't be corrected later, handle the conversations where a relationship is won or lost, or become the only record of why a decision was made. Outside those areas, most of what it's good at is low risk, because a person reads the output before it goes anywhere.

We get asked "what can AI do for us?" more than anything else, and it tends to produce a list of capabilities. The more useful question asks about consequences: where in this business would a confident wrong answer be expensive, hard to reverse and difficult to pin on a named person? For most businesses the answer is a short list. This piece is ours.

How do you sort a task?

We put each candidate task through three questions. They aren't legal tests, just a quick way to split a long list into two piles.

Can it be undone? A draft someone edits costs nothing to get wrong. A payment that's left the account, a message a client has already read, a decision that changes someone's employment: none of those come back. Reversibility is the best single guide we've found.

Can a person check it in the time they really have? If checking takes as long as doing the task, you've saved nothing and added a way to fail. If it takes longer, you've made the job worse while feeling productive.

Will a named person stand behind the result? If the truthful answer is "the system decided", nobody can defend it when someone asks why.

A task that fails two of the three goes on the list below, whatever the demo looked like.

Should AI make decisions about people?

No. This is the clearest line and the one most often crossed by accident.

AI shouldn't make, or effectively make, decisions on hiring, firing, promotion, pay, performance ratings or redundancy selection. It can help write the job description, structure interview notes, draft the difficult conversation and summarise a pile of applications for a person to read. It shouldn't produce the ranking that decides who gets the job.

The law points the same way. The EU AI Act classes AI used in recruitment and worker management as high risk, which brings duties including effective human oversight under Article 14; after the simplification package agreed this year, those rules for stand-alone high-risk systems apply from 2 December 2027 (European Commission). Under the EU GDPR, people have a right not to be subject to a decision based solely on automated processing with legal or similarly significant effects, subject to exceptions (Article 22). The UK replaced its version from 5 February 2026 with Articles 22A to 22D, which allow such decisions more widely but still require safeguards, including the right to human intervention and to contest the decision (Data (Use and Access) Act 2025, section 80).

Then there's the practical side. People decisions get contested, and the person affected will want to know why. "The model scored you lower" isn't an answer an organisation can defend or a manager should have to give.

What about numbers that only go out once?

Payroll, tax returns, statutory accounts, dosages, structural calculations, customs declarations: anything where the figure leaves the building and can't be corrected a week later belongs here.

AI is often useful in the preparation: reading the messy input, spotting the inconsistency, producing the first pass of a reconciliation. It can't replace the qualified person who signs, and that signature is what the client is paying for. If your process ends in a signature, the AI's work has to be finished and checked before it.

Advice that reaches a client directly belongs here too. A solicitor can use AI to draft; the solicitor's client shouldn't be getting AI advice with no solicitor in between. The same goes for accountancy, tax, medicine and financial advice. Professional bodies' rules usually say so, and where they haven't caught up, clients' expectations already have.

Which conversations should stay human?

Some of the costliest mistakes are about tone rather than fact. Don't let AI deliver bad news, negotiate, apologise to a client you've let down, or answer the complaint from someone who's waited three weeks. Use it to draft if that helps, then rewrite it in your own voice and send it yourself.

A model can write a perfectly good apology. The trouble is that these moments repair or end a relationship, and repair depends on someone visibly taking responsibility. A polished message from nobody in particular makes things worse, because it confirms the client was dealing with a process.

Where should the reasoning live?

Organisations need to be able to explain their decisions, sometimes years later and sometimes to a regulator or tribunal. If the thinking behind a decision exists only in a conversation with a tool, you can't reconstruct it. The output survives and the reasoning doesn't.

So the person who decides writes the reason down, in their own words, in your system of record. AI can help draft that note, but the note has to exist somewhere you control.

The same goes for an assistant's memory. Remembering your preferences is useful. Becoming the only place a project's history is kept turns it into a single point of failure with a friendly interface. Treat its memory as a convenience, keep the archive elsewhere, and see what happens to your data for where that memory sits.

Is a human reviewer enough?

Often it isn't, and critics of "human in the loop" are mostly right about that. Anyone who's watched a busy reviewer approve two hundred items in an hour knows oversight can shrink to a click. A systematic review of automation bias found people over-rely on automated advice, including clinicians accepting incorrect decision support, with time pressure and the system's apparent reliability making it worse (Goddard, Roudsari and Wyatt, 2012). A confident draft on your screen is exactly that kind of advice. If the safeguard is a manager with ninety seconds per item and no incentive to disagree, you've added a signature and no safety.

Critics make a second fair point: keeping AI out of high-stakes work has costs too. A clinician with a good second opinion may make fewer errors. A small practice that never uses a drafting tool may just be slower, with no gain in accuracy. Excess caution lands on the same people it was meant to protect.

That moves our line rather than erasing it. AI can sit close to important decisions when the check is real: a named reviewer, enough time to disagree, access to the underlying material and a written record of the reasoning. Where you can't build that, keep the task human. A quick way to find out which you have is to take a document your reviewer has already approved, plant a plausible error in it, and see whether it gets caught. The safety guide covers what a working check looks like.

Writing your own list

Put it on one page: the things your business won't let AI decide, deliver or sign. Five items that people read beat a twenty-page policy nobody opens. For everything in the grey zone, write "draft" or "decide" next to it; most of the anxiety goes away at this point, because most tasks turn out to be drafts. Then, beside every regulated or irreversible output, write the job title of whoever signs. If no title fits, you've found your first problem.

If you're still working out where AI belongs in your organisation at all, the employer check is a better place to start than a tool comparison.

Common questions

Can AI sit on a hiring panel?

It can support one, but it shouldn't score candidates in a way that decides who progresses unless a person independently reviews the reasoning and can override it. Recruitment counts as high risk under the EU AI Act, and candidates have rights over solely automated decisions with significant effects.

For getting oriented and drafting, often yes; for advice you'll act on, no. Answers are right most of the time and confidently wrong occasionally, and you won't know which until after you've relied on one. Take the decision to a qualified person.

Can managers use AI to write performance reviews?

For drafting the repetitive parts, yes. The rating, the examples and the judgement about the person should stay with the manager, who knows the work and has to hold the conversation. A review built from generic phrasing reads as indifference, and employees notice.

Should we ban AI in high-risk areas?

A blanket ban is easy to write and hard to keep, because people use it anyway where nobody can see. A rule that AI may draft and a qualified person must decide, with the reasoning recorded, is easier to enforce and closer to what most professional bodies are moving towards.