AI chatbot development cost in the UK: from £0 SaaS to bespoke builds

An AI chatbot costs a UK SME anything from £0 on a helpdesk's free tier to over £20,000 for a bespoke build that answers from your own documents and hands over to a person. We scope that bespoke build at 20–40 days of senior work: roughly £11,000–£22,000 at September 2026 contract rates. Running costs are usually the smaller number.

This post is about the build end. If you only need a chatbot on your website answering common questions, the off-the-shelf tools are priced in our guide to AI for customer service in small businesses, and that is where most SMEs should start.

What are the three ways to get an AI chatbot?

There are three routes, and they differ more in who does the work than in the technology underneath.

Buy a SaaS tool. A helpdesk vendor's AI agent, switched on and loaded with your answers. Setup is measured in hours and pricing is per conversation or per resolution. You get the vendor's features and nothing else.

Build on a platform. A low-code agent builder such as Microsoft Copilot Studio, configured against your SharePoint, website or knowledge base. Someone still has to design the conversation, the sources and the handover, but the hosting, the model and the chat window come with the platform.

Build it bespoke. Your own application calling a model API, with a retrieval layer over your documents, your own guardrails and whatever integrations the job needs. This is what "AI chatbot development" usually means when a developer quotes for it, and it is the only route where the cost is mostly people's time.

How much does a bespoke chatbot build cost?

The market rate for the people who build these is public. In the six months to 30 September 2026, UK contract roles citing large language model (LLM) skills had a median day rate of £550, with the tenth percentile at £405 and the ninetieth at £800, from 669 quoted rates (IT Jobs Watch).

The day counts below are our own planning assumptions, not survey data. They are how we would scope this work, and they move with the state of your documents and systems. The costs are the day counts multiplied by that £550 median, so you can sanity-check a quote.

What you are building	What it does	Planning band	Indicative cost at £550/day (Sept 2026)
Configured SaaS agent	Vendor tool loaded with your answers and handover rules	1–3 days	Roughly £550–£1,650, plus the subscription
Platform build (e.g. Copilot Studio)	Answers from your SharePoint or site, with a designed handover	5–15 days	Roughly £2,750–£8,250, plus platform credits
Bespoke grounded assistant	Answers from private documents via retrieval, with escalation and logging	20–40 days	Roughly £11,000–£22,000
Bespoke agent that takes actions	As above, plus bookings, CRM updates or order lookups	30–60 days	Roughly £16,500–£33,000

These bands agree with the "grounded assistant or chatbot" line in our AI consulting cost guide, which uses the broader AI-skills median. A fixed-price quote from a specialist can land below the arithmetic, because the scope is closed and the pattern is familiar. A quote far above the top of a band should come with a reason you can point at.

What goes into those days?

Less of it is "AI" than buyers expect. In our view (an opinion, not a finding), the days split into five jobs:

1. Getting the source content fit to answer from. Deduplicating policies, removing out-of-date versions, and writing the answers that only exist in people's heads. This is the job most often underestimated.
2. Retrieval. Splitting documents into passages, indexing them, and tuning the search so the model is shown the right paragraph. When chatbots give confident wrong answers, weak retrieval is a common cause.
3. Behaviour and guardrails. What the bot may answer, what it must refuse, how it says "I don't know", and when it hands over.
4. Integration and handover. Passing the conversation to a person with its history, logging it, and, for an agent, calling your other systems safely.
5. Testing and compliance. A set of real questions with agreed right answers, run before launch and after every change, plus the data protection work. For most customer-facing chatbots that includes a DPIA; our GDPR and AI checklist covers what to settle.

If a proposal spends most of its days on items 3 and 4 and almost none on 1 and 5, ask why.

What does it cost to run each month?

Model fees are metered per token, roughly three-quarters of a word. To make the numbers concrete, here is a worked example on our own assumptions: 1,000 conversations a month, four exchanges each, with every exchange sending about 4,000 input tokens (instructions, retrieved passages and conversation history) and returning about 300 output tokens. Prices are the providers' published list prices in US dollars as of September 2026, before any caching or batch discounts.

Model	Price per million tokens, input / output	Cost per conversation	1,000 conversations a month
OpenAI gpt-5.4-mini	$0.75 / $4.50	about $0.017	about $17
Anthropic Claude Haiku 4.5	$1 / $5	about $0.022	about $22
Anthropic Claude Sonnet 5.5	$2 / $10	about $0.044	about $44
OpenAI gpt-5.4	$2.50 / $15	about $0.058	about $58

Sources: OpenAI API pricing and Anthropic pricing, both fetched September 2026. Both providers discount repeated input: Anthropic lists cache reads at 0.1× the base input price for most models, and OpenAI lists cached input for gpt-5.4 at $0.25 per million tokens. A chatbot that sends the same long instructions every time can cut the input bill substantially.

On a platform, the unit is different. Microsoft lists a Copilot Studio pack of 25,000 Copilot Credits at £153.80 a month as of September 2026, and its billing rates charge 2 credits for a generative answer and 5 for an agent action. On the same four-exchange example, with every reply a generative answer, one pack covers about 3,125 conversations a month. Agent actions and other features use credits faster, so model your own mix before you commit.

The pattern in both tables: at SME volumes, the model fee is rarely what makes a chatbot expensive. What costs money every month is:

• Hosting and the retrieval index, usually modest at SME scale but not zero.
• Monitoring, so someone sees wrong answers before customers do.
• Content upkeep. Every time a price, policy or product changes, the source documents must change with it, or the bot answers from the old version.
• Model changes. Providers retire and replace models; each change means re-running your test questions.

When is a bespoke build worth it?

Our rule of thumb, and it is an opinion: build bespoke only when at least one of these is true.

• The answers live in private documents a SaaS tool cannot reach or should not hold: contracts, technical manuals, internal procedures.
• The bot must act, not just answer: check an order in your own system, book an appointment, update a record.
• Volume changes the arithmetic. Per-resolution pricing that looks cheap at 200 conversations a month can overtake a bespoke build's running cost at 20,000.
• You need control over where data is processed, how long logs are kept, or which model is used.

If none of those apply, a SaaS tier plus a fortnight of writing good answers will beat a bespoke build on cost and time to value. We would rather tell you that at the scoping stage than after the invoice.

How do you keep the build cost down?

Fix the content first. A day spent agreeing the canonical refund policy saves several days of tuning a bot that cannot tell two versions apart.

Start narrow. One channel, one audience, the ten questions that make up most of the volume. Widen scope once the handover rate and answer accuracy are known.

Write the test set before the build. Fifty real questions with agreed answers turn "does it work?" from an opinion into a score, and stop the project drifting.

Fix the price for defined scope. A grounded assistant with a known document set is describable, so it should be quotable. Time and materials suits genuine unknowns, such as integrations nobody has documented.

Cap the metered costs. Set a monthly spending limit with the model provider on day one.

Our custom AI development service covers what a production build involves, and the AI practice overview shows how it fits with AI strategy work when you are not yet sure a chatbot is the right answer.

Frequently asked questions

How much does it cost to build a custom AI chatbot in the UK?

We scope a bespoke chatbot that answers from your own documents and hands over to a person at 20–40 days of senior work. At the median UK contract rate for LLM skills (£550 a day in the six months to 30 September 2026, IT Jobs Watch), that is roughly £11,000–£22,000. A chatbot that also takes actions in other systems costs more.

What does an AI chatbot cost to run each month?

Less than most people expect in model fees. On our worked example of 1,000 four-exchange conversations a month, published API prices as of September 2026 come to roughly $17 to $58 depending on the model. Hosting, monitoring and someone's time to review answers usually cost more than the tokens.

Is Copilot Studio cheaper than a bespoke chatbot?

Usually to build, sometimes not to run. As of September 2026 Microsoft lists a 25,000-credit Copilot Studio pack at £153.80 a month, and a generative answer uses 2 credits. It suits businesses already on Microsoft 365 with modest volumes; high volumes or unusual integrations tend to favour a bespoke build.

Should a small business build its own chatbot at all?

Only when an off-the-shelf tool cannot do the job: answers must come from private documents a SaaS tool cannot reach, the bot must act in your own systems, or volumes make per-resolution pricing expensive. Otherwise start with a SaaS tier and spend the money on writing good answers.

---

Easy Insight is a UK consultancy for AI, web, apps and data — senior specialists only, no juniors.