Almost every founder who asks us about this expects a single figure, and the honest answer is that the cost to add an AI chatbot splits cleanly into two parts that behave very differently. There is the one-off build, which for most apps runs from roughly £6,000 for a narrow assistant that answers questions from a fixed set of documents up to £40,000 for a chatbot that reads your live data, takes actions, and is safe enough to put in front of paying customers. Then there is the model usage, a bill that arrives every month and grows with conversations rather than with users. Budget only the first and you will be fine until the thing succeeds, which is a strange kind of failure. Here is how each part actually works.
What does the cost to add an AI chatbot actually come to?
The cost to add an AI chatbot is a one-off build of roughly £6,000 to £40,000, plus a monthly model bill that scales with how many conversations happen. A scoped assistant answering from your own documents sits at the lower end. A chatbot that reads live customer data, takes actions, and is properly guard-railed sits at the top.
That shape is the whole thing in miniature, and it is why quotes for the same brief can differ by a factor of five. Two studios can both be quoting honestly and be describing completely different products, because “add a chatbot” covers everything from a tidy help widget to a small piece of infrastructure that touches your database and speaks on your behalf.
Part one: the build
The build is the engineering that turns a model into a feature your users can actually rely on. At minimum that means the chat interface itself, a connection to a model provider, the prompt and system instructions that give the assistant its job and its limits, conversation history so a thread survives a refresh, and error handling for the moments when the model is slow, unavailable, or confidently wrong. A narrow assistant of that shape, bolted onto an app that already exists, commonly lands between £6,000 and £12,000.
Costs climb the moment the chatbot needs to know things that were not in its training data. That is where retrieval comes in: your documents, help articles, product catalogue, or policies get indexed so the model can quote from them rather than invent an answer. Retrieval is where most of the real engineering lives, because someone has to build the ingestion, keep the index fresh when the source changes, and tune what gets pulled in so the answers are grounded. A retrieval-backed chatbot usually runs £15,000 to £25,000.
The top of the range is a chatbot that does things rather than just says things: looking up a specific customer’s order, changing a booking, raising a ticket, issuing a refund. Every action is a permissioned integration with its own failure cases, and every one needs a guardrail so the assistant cannot act outside its remit. That work, with the testing and safety review it demands, is what takes a build toward £40,000. These figures are the chatbot slice on its own, and they sit inside the wider picture in our breakdown of how much an app costs to build in the UK in 2026.
Part two: the running cost
Separate from the build, you pay a model provider for the words going in and out. This is the part founders consistently forget, because it behaves unlike any other line in a software budget. It does not scale with registered users, it scales with conversations, and a small group of enthusiastic users can cost more than a large group of quiet ones.
Three things move that bill. The first is which model you use, because the frontier models cost meaningfully more per conversation than the smaller, faster ones, and a great many support chatbots do not need the frontier. The second is how much context you send with each message: a retrieval chatbot that stuffs ten pages of documentation into every request is paying for those ten pages every single time. The third is conversation length, since each turn carries the history before it.
Published model pricing changes often enough that any figure printed in a blog post is a liability, so model your own numbers from the current rates on the OpenAI pricing page and the Anthropic pricing page rather than from an estimate you read second-hand. What matters at budgeting stage is the shape: a low per-conversation cost, multiplied by a conversation count you should forecast honestly, with a ceiling you set deliberately rather than discover on an invoice.
The three types, and why they cost so differently
| Type | What it does | Typical build | Running cost |
|---|---|---|---|
| Scoped assistant | Answers from a fixed prompt and a small set of instructions | £6,000 to £12,000 | Low, short conversations |
| Retrieval chatbot | Answers from your documents, catalogue, or help centre | £15,000 to £25,000 | Moderate, context-heavy requests |
| Agentic chatbot | Reads live data and takes actions on a user’s behalf | £25,000 to £40,000 | Variable, longer multi-turn threads |
Most founders arrive describing the third and need the first. That is not a criticism, it is just what the demos have taught everyone to picture. The useful question is not which type is most impressive, it is which type removes the most work from your team per pound spent, and a scoped assistant that deflects the twenty questions your support inbox gets every day will often beat an agentic build that impresses in a pitch and then needs constant supervision.
What drives the cost to add an AI chatbot up
Four things push the number, and only one of them is the model. The first is accuracy expectations. A chatbot that is allowed to say “I am not sure, here is a human” is dramatically cheaper than one expected to be right every time, because certainty is bought with evaluation work, test sets, and iteration, not with a better prompt.
The second is data access. Every system the chatbot needs to read, your CRM, your booking engine, your order history, is an integration with authentication, rate limits, and edge cases, and integrations are where estimates quietly double. The third is safety and compliance. If the assistant handles personal data you are inside UK GDPR, which means being deliberate about what is sent to a model provider, what is retained, and what users are told. The ICO guidance on AI and data protection is the source worth reading before you scope, not after, and our guide to handling GDPR in a UK consumer app covers the groundwork.
The fourth is evaluation and monitoring. A chatbot is the only feature you will ship that can get worse without anyone changing a line of code, because the model updates, your documents drift, and user questions move on. Budgeting a small ongoing amount for reviewing real conversations is the difference between a feature that improves and one that quietly embarrasses you.
A worked example: a UK services business
Picture a company with a booking app and a support inbox that fills up with the same questions about opening hours, cancellation terms, and what is included in each package. The cheap, correct answer is a retrieval chatbot pointed at the existing help content, with a clear handover to a human the moment it is unsure. Build: somewhere around £15,000 to £20,000. Running cost: modest, because the conversations are short and the deflection is worth far more than the model bill.
Now stretch the brief. The same business wants customers to move a booking in the chat, see their own history, and get a refund processed without a phone call. That is three permissioned integrations, an authentication path, an audit trail, and a guardrail on every action, which pushes the build toward the top of the range and turns a four-week project into a three-month one. Both versions are defensible. Only one of them is the right first version, and it is almost always the one that ships in a month and earns the case for the second.
What to do next
The cost to add an AI chatbot comes down to two honest numbers: a one-off build of roughly £6,000 to £40,000 depending on whether you need a scoped assistant, a retrieval chatbot, or something that takes actions, and a monthly model bill that scales with conversations rather than users. Scope the narrowest version that removes real work, model the running cost at the volume you are actually aiming for, and put a spending ceiling in place before launch rather than after the first surprising invoice.
If you want a straight figure for your own app, including an honest view on whether a chatbot is the right answer at all, send us a project brief and we will scope it properly. If you are weighing this against a human-to-human messaging feature, our breakdown of the cost to add in-app chat uses the same two-part method, and choosing an AI app development company in the UK covers what to ask before you commit to anyone.
.webp)
.webp)
