AI customer service: what to hand to AI and what to keep human

Search for AI customer service and nearly everything you find is written by a company selling an AI agent. This guide is for a UK business that answers its own customer email: what to hand to AI, what to keep human, what the helpdesks' agents cost, when to build, and which rules apply.

We build AI automation that sorts customer email and drafts replies, tested on real messages and with a person checking.

What AI does well in a customer inbox

Most inboxes are less varied than they feel: where's the parcel, can the booking move, how do returns work, can you resend the invoice. That is where AI email automation earns its keep, in four jobs:

• Triage. Saying what each message is (an order query, a return, a complaint, a sales lead), how urgent it is and which queue it belongs in, with complaints and anything mentioning a solicitor flagged for a person. Customers describe the same problem in dozens of ways, and language models cope with that far better than keyword rules.
• Pulling out the facts. Order numbers, booking references, postcodes and dates picked out of free text, so the order can be looked up before anyone opens the email.
• Drafting replies. A first draft written from your own policies and the customer's actual order or booking, for a person to check and send.
• Summarising. A long thread turned into three lines at the top of the ticket, so nobody reads twenty emails first.

AI doesn't know your business unless you give it your policies and data, and it doesn't know when it's wrong. A draft a person approves keeps most of the speed and little of the risk, because the time goes on looking things up and starting from blank, not on pressing send.

What it should not do alone

Some messages need a person to decide, even if AI writes the first draft:

• Money that needs judgement. Refunds, credits and goodwill payments for damaged goods, a late delivery or a partly used service. A refund that follows a fixed rule is different, as we explain below.
• Complaints. A fluent but wrong reply makes a complaint worse.
• Anything with legal weight. Contract terms, cancellation rights, warranty claims, data protection requests, and anything that mentions a solicitor.
• Vulnerable customers. Someone who mentions illness, bereavement, money trouble or distress needs a person. Their message may contain health information, which data protection law treats with extra care.

In Moffatt v. Air Canada, decided by British Columbia's Civil Resolution Tribunal on 14 February 2024, the airline's website chatbot told a bereaved customer they could apply for a bereavement fare retroactively, which the airline did not allow. Air Canada argued it could not be held liable for what its chatbot said. The tribunal called that "a remarkable submission", found the chatbot was "still just a part of Air Canada's website", and ordered the airline to pay 650.88 Canadian dollars in damages, plus interest and fees. It binds no UK court, but the lesson is plain: customers will treat what your AI says as what you said.

The helpdesks' own AI agents

If you use a helpdesk, it probably sells an AI agent that answers customers directly. Here is what three say on their own pages, with prices as of October 2026.

Read the billing definitions closely. Fin counts a resolution when "no further help is requested after Fin's last answer", and a customer who gives up and phones you instead can look the same. Before you trust any resolution rate, read a sample of "resolved" conversations yourself.

When buying one is the right answer

Buy, and don't build, when most questions can be answered from published information such as delivery times and returns policy, when your orders or bookings live in a platform the helpdesk already connects to, and when the price per resolution is less than what that answer costs you now. If you're on one already, try its agent on a few message types, read what it sends, and widen it before paying anyone to build anything.

Narrow actions with fixed rules are fine to automate. Gorgias's cancel action only runs on an order that hasn't been fulfilled, and its documentation says the agent can't tell whether a new address changes taxes or shipping costs, so it won't refund or charge the difference. That is the shape to look for: a fixed rule, with anything outside it left to a person.

When to connect AI to your own systems

Many businesses keep their facts where no helpdesk reaches: an order system built years ago, a booking system, a stock spreadsheet. Then the AI drafts from your policies but not the customer's order, and a draft promising delivery in three to five days for a parcel that left yesterday is worse than none.

Building the connection means software identifies the customer, looks up their orders, bookings or stock, and hands those facts to the AI with your policies. The draft then gives the real dispatch date and tracking number, and the checker can see where each fact came from. Fin and Zendesk both say their agents can connect to other systems, so this can plug into the helpdesk you already pay for.

The hard part is usually identifying the customer, not writing the reply. Match on the sender's email address or an order number, never on a name alone. We made the same call in Add to CRM, the Chrome extension we built and run that adds people from LinkedIn, Gmail and Outlook to a team's CRM. When a profile and a CRM record share only a name, it shows a possible match with Approve and Reject, and writes nothing to that record until someone on the team approves. We turned down linking on a name because one wrong link lets a later update overwrite another person's record. In a customer inbox, the same mistake puts one customer's order details in a reply to another.

For the same approach on invoices and other documents, see AI automation in the back office.

The UK rules that apply

The ICO on AI and personal data

Customer emails are full of personal data. The ICO's guidance on AI and data protection says, in its chapter on accountability, that "in the vast majority of cases" using AI will involve processing likely to result in a high risk to people, and so will legally require a data protection impact assessment. The ICO says the guidance is under review after the Data (Use and Access) Act 2025. In practice, know which AI provider each message goes to, on what terms, whether it can train their models, and how long it's kept.

Automated decisions: Articles 22A to 22D

The Data (Use and Access) Act 2025 replaced Article 22 of the UK GDPR with new Articles 22A to 22D, in force from 5 February 2026. A significant decision, one with a legal or similarly significant effect on a person, taken with no meaningful human involvement needs safeguards: you must tell the person about it and let them make representations, obtain human intervention and contest it. Decisions based on special category data, such as health information, are restricted further. An AI that refuses a refund or closes an account by itself may be taking that kind of decision. A person who reads the draft and genuinely decides keeps you outside these rules; one who approves everything unread is hard to call meaningful involvement.

The FCA's Consumer Duty

If you're regulated by the FCA and serve consumers, read the Consumer Duty first. One of its four outcomes is consumer support: ongoing service should be "as easy to access as sales", without "barriers to switching, cancelling, or getting help". An AI agent that makes it hard to reach a person could be one of those barriers.

How to start

1. Sort a few weeks of recent messages by type, noting the systems and decisions each needs. Our guide to process mapping shows how.
2. Pick two or three types that are frequent, factual and low risk, such as order tracking.
3. Have AI draft replies for those, with a person approving every one. Count how often a draft goes out unchanged.
4. Keep a set of real messages with known good answers, and rerun it whenever a prompt, model or policy changes.
5. Let it send on its own only where drafts have been right for weeks, and keep a clear route to a person in every reply.

If your helpdesk's own agent covers what you need, use it. If your facts live in systems it can't reach, building that connection is work we do. Start a project and tell us what lands in your inbox.