AI for Customer Service: What to Automate & What It Costs

Yes, you can use AI for customer service. It works best on six jobs: sorting messages, drafting replies for approval, answering from your own help content, summarising threads, handing over with context and checking quality. Complaints, vulnerable customers, and refunds or exceptions should stay with a person.

Most of the work lies in the choices around those jobs. A first AI agent that does one of them inside your existing tools starts, with us, from £5,000. A helpdesk add-on may be cheaper if all your answers live in one help centre. The rest of this guide covers what to automate, what to keep human, buying versus building, data protection, and how to roll out and measure it.

Key takeaways

• Automate the sorting, drafting and lookup work around a conversation first. Keep the decision on complaints, vulnerable customers, refunds and exceptions with a person.
• Drafting replies for approval is the safest first step on email, because a person sends every message until the drafts have earned trust.
• A helpdesk vendor’s AI is the right buy when your answers live in your help content. A custom agent earns its cost when it must read or change data in your other systems.
• UK GDPR rules on solely automated decisions changed in February 2026, and the ICO’s updated guidance is still in draft. Give every customer a route to a person.
• Pilot on one queue, record your current numbers first, and widen only when a weekly sample review comes back clean.

Can you use AI for customer service?

Yes, for the work around a conversation more than for the decisions inside it. AI reads free text well, writes a plausible first draft quickly and looks things up without tiring. It is unreliable at deciding what the business owes a customer in an unusual case.

The table shows the six jobs we would start with and who makes the final call in each.

Job	What AI does	Who decides
Triage and routing	Labels and prioritises	Rules you set
Drafting replies	Writes a first reply	A person approves
Self-service answers	Answers from help content	The content you publish
Ticket summaries	Condenses long threads	The next person
Handoff	Passes on the context	A person takes over
Quality checks	Flags errors and tone	A reviewer

An AI agent here is software that completes a multi-step job, such as reading a request, looking up the order and drafting the reply. Our guide to what an AI agent is covers the definition.

Examples of AI in customer service: six jobs that work

Triage and routing

Support work starts with sorting. An agent reads each email or message, labels the topic, spots urgency and sends it to the right queue. This suits AI because the input is free text that fixed rules struggle to classify.

Keep the routing rules in your own hands: which topics go to which team, which customers get priority and what triggers an escalation. Our AI agent examples include a support triage agent that drafts first responses for approval.

Drafting replies for approval

The agent writes a first reply from the customer’s message, the order or account record and your policies. A person reads it, edits it and sends it. This is the safest first step on email, because nothing reaches a customer unchecked.

We used the same pattern for SBU, a UK energy broker. Its sales automation classifies email replies and drafts responses for approval, with an 85% confidence threshold deciding what runs automatically and what waits for a person. That job is sales rather than support. The pattern carries over: a threshold you can change, and a person on everything below it.

Answers from your own help content

Many questions are already answered on your site: delivery times, how to reset a login, what the returns policy says. An agent can find the right page and write the answer in the customer’s words.

The ceiling is your content. An out-of-date policy page produces a confident, out-of-date answer. Restrict the agent to approved sources, and let it say it does not know and hand over.

Ticket summaries

Long threads waste time when a case changes hands. An agent can summarise the history, what the customer is asking for and what has been tried, so the next person does not reread twelve messages.

This is low risk because the summary informs a person and the original thread is still there.

Handoff to a person

Handoff is a job in itself. When the agent cannot answer, or the customer asks for a person, it passes the conversation on. The handover should carry the context: who the customer is, what they asked, what was tried and what the agent suggests.

Asking the customer to repeat everything after a handoff undoes the time the agent saved. The route to a person should never be hidden.

Quality checks

A second pass can read outgoing replies, whether a person or an agent wrote them, against your policy and tone guide. It flags replies that promise something you do not offer or leave out a line you require.

Treat it as a reviewer’s assistant: it flags, and a person decides.

What to keep with a person

Complaints, vulnerable customers, and refunds or exceptions stay with a person. AI can write a polite message. These cases turn on judgement and discretion, and on what the business is prepared to do for this customer.

• Complaints. A complaint tells you something went wrong, usually with emotion and with facts still to establish. Let the agent summarise the history and pull the order record for whoever handles it. The reply and the remedy are theirs.
• Vulnerable customers. The FCA describes a vulnerable customer as someone who, because of their personal circumstances, is especially susceptible to harm. Its guidance for firms says: “Deliver appropriate customer service that responds flexibly to the needs of vulnerable consumers” (FCA). That guidance is written for FCA-regulated firms. Other businesses can still borrow the idea. An agent can watch for signs such as a bereavement or difficulty paying and route those cases to a person. It should not decide the treatment.
• Refunds and exceptions. Start with the agent preparing the case: order found, policy matched, amount worked out. A person approves it. Simple in-policy refunds can move to rules later, once your data shows the approvals are routine. Anything outside policy, above a value you set or contested stays with someone who has the authority.

Every customer should be able to reach a person on every channel. That is a design rule, and the data protection section below explains why it can also be a legal one.

Email, web chat or WhatsApp: which channel first?

Start with email. A reply can wait for a person to approve it, the thread carries the context, and there is no pressure to answer in seconds.

Channel	Pace	Start with
Email	Can wait for approval	Drafts a person approves
Web chat	Live	Help-content answers, instant handoff
WhatsApp	Live, with time rules	Only after email works

Web chat is live, so approval before sending does not fit. The agent must stay inside approved content and hand over quickly, and you need to know who is online to take the handoff and what happens out of hours.

WhatsApp adds platform rules. Meta’s developer documentation says “When a WhatsApp user messages you or calls you, a 24-hour timer called a customer service window starts.” It adds: “When the window closes, you can only send pre-approved template messages.” (Meta for Developers, checked 20 October 2026). If a draft waits overnight for approval, the window may close before anyone replies, so plan alerts and cover hours around it.

Helpdesk AI add-on or a custom agent?

If your answers live in a help centre and your tickets live in one helpdesk, the AI add-on from that vendor is probably the right buy. It already sits where the conversations are, and the vendor maintains it. For what these add-ons cost and how they are billed, see our chatbot cost guide.

A custom agent earns its cost in three situations.

• The answer is in another system. Order status, a booking, an account balance or a stock level sit in systems the helpdesk does not reach. Connecting them is the main work, and our systems integration service starts from £5,000 for two systems connected.
• The agent has to act. Changing an address, rebooking a visit or raising a credit note means writing to your systems, with limits and approval rules you control.
• The rules are yours. A threshold, a policy table or an approval chain that does not match the vendor’s settings is easier to build than to bend.

Check what your vendor’s tool can actually reach before you decide, and ask the same of any agency, including us.

AI agents for e-commerce support

Most e-commerce support is a handful of repeats: where is my order, can I return this, can you change my address. An agent that answers them needs the order system, the carrier’s tracking and your returns policy, so it is more a connection job than an AI job.

Lookups only read data, so they are the low-risk start. Address changes, cancellations and refunds are actions, so they wait for approval until sample reviews show they are routine.

Not every enquiry is a support ticket. A question like “how much would this cost?” is a quoting job. We built ChatQuote, an AI quoting assistant for service businesses that drafts a quote from a website enquiry for the team to review.

Data protection checks before you switch it on

Customer messages are personal data, so UK GDPR applies. Four checks come first. This is general information, not legal advice, so ask your data protection adviser about your own case.

1. Do a DPIA. The ICO says a data protection impact assessment is needed where processing is likely to result in a high risk to people’s rights and freedoms. It lists processing that uses innovative technologies, including AI, among the types of processing that can require one (ICO). The page carries a notice that it is under review because of the Data (Use and Access) Act.

2. Get the vendor contract right. If a vendor processes customer messages for you, it is a processor. Article 28 of the UK GDPR says you must use only processors that give sufficient guarantees, and their contract must have them act only on your documented instructions (legislation.gov.uk). Ask where messages are stored, how long they are kept and whether they are used to train models.

3. Know the rule on solely automated decisions. The Data (Use and Access) Act 2025 replaced Article 22 of the UK GDPR with Articles 22A to 22D. A decision counts as solely automated when “there is no meaningful human involvement in the taking of the decision”. A significant decision is one with a legal or similarly significant effect on the person. For those, the controller must provide safeguards, including a way to obtain human intervention and to contest the decision (legislation.gov.uk, section 80). The section has been in force since 5 February 2026. At the time of writing, the ICO’s updated guidance on automated decision-making is a draft out for consultation (ICO).

Not every support decision will meet that threshold. A refusal of a refund, an account closure or a fraud flag might, so check with your adviser. An agent that drafts while a person genuinely reviews and can change the outcome is a different thing from a person who clicks approve without reading.

4. Read the AI guidance. The ICO’s guidance on AI and data protection covers accountability, transparency, fairness, accuracy and security (ICO). It carries the same under-review notice, so check the current version.

How to roll out AI customer service and measure it

1. Pick one queue. Choose a high-volume, low-stakes one, such as order status or opening hours.
2. Record your current numbers. Note first reply time, time to resolution, how often cases are reopened, how often they escalate and how satisfied customers say they are.
3. Run in shadow. The agent drafts, a person writes the real reply, and nothing reaches a customer. Compare the two for a few weeks.
4. Move to approval. A person edits and sends each draft. Log every edit. How often a draft is changed tells you more than any demo.
5. Set the threshold. Begin with everything waiting for a person. Lower the bar only for categories whose drafts are sent unchanged week after week, as in the SBU example above.
6. Review a sample weekly. Read a random set of the agent’s work, including anything sent without review. Log the failures and fix the content or the rules behind them.
7. Widen slowly. Add the next queue or channel only after the sample review is clean.

Measure cases the agent touched separately from those it did not, so a good average cannot hide a bad pocket. Track three things: speed (first reply and resolution time), quality (reopen rate, edit rate, sample scores) and outcome (customer satisfaction and escalation rate).

What it costs and where to start

The route sets the cost. Our chatbot cost guide, linked above, compares how helpdesk add-ons are billed, with dated figures.

A custom agent is a fixed-price build. A first agent does one named job inside your existing tools, with clear limits on what it decides alone and a person signing off where it matters. It starts from £5,000. An agent rollout across several workflows starts from £12,000. Scope and price are fixed and agreed before work starts, and you pay by milestone as each stage is delivered. Because these are builds, you can cancel within the first 21 days and get your money back in full. All code and IP are yours once the work is delivered and paid for. Running costs sit on top: model usage, hosting, and maintenance and support from £400 a month.

If you are not sure which queue to start with, our one-week AI readiness assessment is £3,000. It maps your data, workflows and team skills, and ends with a proposal of estimated returns and an honest verdict where AI is not worth it.

The next step is small. Pick the one queue you would pilot, then write down its monthly volume, the five most common question types and what a person does with each. That page is the brief for a first agent. If it points to a job inside your existing tools, our AI agents and automation service starts from £5,000. We will tell you if a helpdesk add-on would do the job for less.