What an AI customer service chatbot costs in the UK, and what separates a useful one from an annoying one.

Because they answer from general knowledge rather than from the firm’s systems, so they cannot say where an order is; because they will not admit they do not know; because there is no way to reach a person; and because they pretend to be one. Every one of those is a design choice, and every one can be made the other way. A customer who gets an accurate answer in ten seconds, or a person with the context already attached, does not mind that an AI did the first part.

Grounding: the assistant reads from your order system, your policies and your knowledge base, and answers only from them, with sources where useful. Escalation: the moment it is unsure, or the question involves money, a complaint, a faulty product or a distressed person, it hands to a human with the conversation and the record attached, and the customer never repeats themselves. Honesty: it says it is an AI, in the first message, and it says “I do not know” rather than guessing. A threshold: an acceptance rate agreed as a number, measured on real conversations before go-live and sampled weekly after, so the firm knows whether it is working rather than hoping.

A staff-facing or limited assistant inside Claude Team or ChatGPT Business, with the policies and standard answers attached, costs only the seats and suits internal use or a small, non-integrated public FAQ. A customer-facing assistant on the website that answers from your documents but not your live systems is a smaller build. The full version, on the website, in WhatsApp and messaging, and on the phone, integrated with orders, bookings and accounts, logged and auditable, is a custom AI system at £8,000 to £25,000 fixed at UK rates, with the running cost stated per conversation before go-live. Published UK market ranges for a bespoke chatbot start in the low thousands for a templated one and run higher for integrated builds.

Prices from our published ladder. Messaging and voice channels are priced per message and per minute by the carrier; see the running-cost guide.

Web chat is the cheapest and the least used by customers who are frustrated. Messaging, WhatsApp and SMS, is where customers already are, and each message has a carrier cost that usually exceeds the model cost behind it. The phone is where the return is largest for a business that misses calls, and it needs the voice-specific design: short exchanges, an honest AI, an escalation path, and a fallback for the majority of outbound calls that do not connect first time. The same grounded assistant can sit behind all three, which is the integrated build; the channel decision is about where your customers are, not about the technology.

Consumer law requires that what the assistant says about returns, delivery, cancellation and prices is accurate, so it reads from the policy and the order rather than improvising, and a person handles disputes. UK GDPR applies to the customer data it reads and the conversations it stores: business plan or API under a data processing agreement, retention set deliberately, a DPIA where the volume or the data warrants it, and the ICO’s expectation that a customer knows they are dealing with an automated system. Prompt injection applies to anything a customer can type, so the assistant has the least access that does the job and cannot send money, change access or delete.

Before: enquiries per day by channel, time to first response, the share answered within an hour, and hours spent by the people answering. After: the share answered by the assistant without escalation, the share escalated with context, customer ratings where you collect them, and the same hours. Sample fifty conversations a week and score them against the threshold. A useful assistant answers most routine questions, escalates the rest cleanly, and returns hours that the team spends on the customers who need a person.