7 decisions you shouldn't hand over to AI

An unhappy customer wants to know a person has heard them. AI can sort complaints by urgency, pull the job history and draft a reply, which saves time. But the decision about what you'll do, and the first real response, should come from someone who can apologise and mean it.

Say you run an estate agency and a seller complains that a viewing went ahead without them being told. An assistant can find the booking, the confirmation emails and who arranged it in seconds. Deciding whether that was your mistake, and what you'll offer, is a conversation for the branch manager.

Screening CVs with AI is tempting, and it's where things go wrong quietly. When the ICO audited AI recruitment tools in 2024, it found some let recruiters filter candidates by protected characteristics, some inferred gender and ethnicity from names, and some collected far more personal data than they needed. It made almost 300 recommendations.

There's a legal angle too. Since February 2026, the Data (Use and Access) Act 2025 has allowed more decisions to be made purely by automated means, but rejecting a job applicant is a significant decision, so you'd still have to tell them, let them make their case, offer a human review and let them challenge it. Use AI to organise applications and draft interview questions. Let a person decide who moves forward.

Promises on delivery dates, fixes or extras are commitments your business has to keep. An AI voice agent or chatbot shouldn't make them. In a 2024 Canadian case, Air Canada was held responsible after its website chatbot gave a customer wrong information about a bereavement fare. Customers reasonably treat what your bot says as what your business says.

Let the assistant capture the request and the context, then have a person confirm what you'll commit to. Our guide on what a voice agent should capture from the call covers where that line sits.

Deciding who to work with shapes your workload, your margins and your team's sanity. AI can summarise an enquiry, check it against your criteria and flag anything missing. It can't weigh up the gut feeling from a first call, or whether this client fits where you want the business to go.

Picture a small accountancy practice in its busy season. An assistant can check a new enquiry for the basics: what services they need, their year end, whether their records are in software the practice supports. Whether to take on a client with three years of late filings in January is a partner's call.

Refunds and credits are money and goodwill, and the right answer often depends on the history. An assistant can find the order, check your policy and draft a response. A person should decide, especially when the case falls outside the written policy, which is when it matters most.

If you want speed, set a clear rule: the assistant can prepare refunds that are fully within policy, and a person approves each one with a click. Anything outside policy goes to a named person straight away.

Sometimes the right call is to pause a job, step away from a client or tell someone you can't help. That involves safety, legal risk, ethics or a relationship, and it needs a person who can explain the reasons and take responsibility for them.

AI can help you spot the warning signs: unpaid invoices piling up, scope creeping, tone changing in emails. But stopping work is a judgement call, and it belongs with whoever is accountable for the client.

None of this means keeping AI out. It means giving it the right jobs. A knowledge assistant is good at finding the policy, summarising the history, drafting the reply and checking a request against your rules, so the person making the decision has everything in front of them.

Before you set one up, check your processes and answers are written down. Our guides on whether your business is ready for AI and how to test an AI agent before it talks to a customer cover the groundwork, and there's more in our guides.