What Custom AI Actually Costs in 2026

An AI subscription costs $20 a month. A custom AI system costs five figures, often six. Both are true, and the gap between them is where most small businesses get confused — or quietly overcharged.

Six development firms published 2026 pricing guides. I read all six, converted them into one currency, and set them side by side. Every firm named here is linked, in the prose, in the tables and in the source list at the end. Check any figure against the page it came from. What follows is a price ladder, what those prices actually buy, and one line worth keeping in your head the next time a proposal lands.

A chatbot and an AI system are not the same product

This is the distinction the whole market rests on, and almost every confused pricing conversation starts by missing it.

A chatbot is cheap. A widget on a website that answers from a list of pre-written responses is a small piece of software. Sharp Code prices a single-channel FAQ assistant at £3,000–£6,000, built in three to five weeks. If that is what a business needs, it should not pay more than that.

An AI system is a different product entirely. It knows the company, and to do that it has to:

• Retrieve your documents, prices and policies accurately.
• Follow your rules and your tone of voice.
• Connect to the tools you already run on.
• Stay correct while the business and its content change underneath it.

Same conversation box on the screen. Different thing behind it, by an order of magnitude in engineering and in price. A business comparing the two on the strength of what the interface looks like will always conclude it is being overcharged.

A quote should make clear which of those two it is for. Most do not, which is exactly how the same box on the screen can be worth £4,000 or $120,000.

The three bands

Read across all six guides, the market resolves into three:

Below the first band sits the FAQ widget at £3,000–£6,000. It is on a different ladder. Above the third, Fulminous and Sharp Code both put full enterprise deployments at £60,000–£250,000+.

Why RAG changes the price

RAG — retrieval-augmented generation — is the step that lets an AI answer from a company's own information instead of from whatever it absorbed in training. It is also the single biggest jump on the price ladder, and the reason is not the model.

You are not paying for ChatGPT. You are paying for:

• Selecting, cleaning and structuring the documents and databases the AI is allowed to see.
• Chunking them, building the embeddings, standing up the index.
• Tuning retrieval until it returns the right passage rather than a plausible one.
• An evaluation harness that proves it does — and keeps proving it as the content changes.
• Guardrails for what the system must never say or invent.
• Authentication, logging and update paths, so it can be changed safely later.

Emvigo prices that layer at £20,000–£55,000 for production — three to seven sources, hybrid retrieval, authentication, evaluation, monitoring. The model subscription inside it costs a few dollars a month.

What you are actually paying for

When a five- or six-figure quote lands, almost none of it is the model. It is the work of turning a general-purpose model into your system — six blocks, and only one of them is the LLM.

Why “just use Claude or GPT” is not enough

Claude and GPT are trained on the public internet. They do not know your business: not your pricing rules, not your contract terms, not your appeal edge cases, not the tone your team uses when something has gone wrong.

Used raw they generalise, they guess, and under pressure they invent. That is survivable in an internal brainstorm. It is not survivable in a system that talks to customers or touches money.

Four things change that, and all four are engineering rather than licensing:

• Retrieval makes the model read your approved content before it answers.
• Guardrails set what it is allowed to say and do.
• Testing turns “it seems good” into a number you can act on.
• Monitoring lets you tighten it when the real world pushes back.

Which is why every serious 2026 guide here is organised around data, retrieval, testing and integration — not around which model was chosen. The model is necessary. It is nowhere near sufficient.

The ladder, rung by rung

Each rung is a real jump in engineering, and each has a published price.

FAQ bot. Pre-written answers, one channel. Sharp Code: £3,000–£6,000.

LLM chatbot. Understands the question instead of matching it. Fulminous: £8,000–£40,000.

RAG. Answers from your documents. Fulminous: £15,000–£60,000. Emvigo, for the retrieval layer alone: £20,000–£55,000.

Agent. It can now do things — create, update, schedule, retrieve, call another system. Fulminous: £30,000–£120,000. Seven Solvers: £25,000–£80,000+.

The pattern is not a UK quirk. Cleveroad prices generative AI assistants at $35,000–$120,000 and transactional, workflow-driven systems at $50,000–$150,000+.

What those prices do not buy

This is the part almost nobody reads, and it is the part that decides whether a quote is expensive or cheap. Emvigo states it plainly in its own guide:

This guide prices RAG specifically — retrieval, embeddings, and vector infrastructure — not general software build costs.

So £20,000–£55,000 buys the retrieval layer. One layer.

It does not buy the website. It does not buy an environment where the site can be changed under supervision — previews, version history, rollback, an approval step. It does not buy the integrations, and it does not buy anyone maintaining it next year.

Those are separate line items in every guide here. Which is why the only honest way to read a price is to ask what fraction of the stack it covers.

Then come the add-ons

Kellton itemises what gets added once a client asks for anything beyond conversation:

• Agentic workflows — tools and multi-step execution: +$20,000–$60,000
• Each major live-system integration: +$10,000–$30,000
• Multimodal input: +$15,000–$40,000
• Multi-language support: +$8,000–$25,000
• Annual maintenance: 15–25% of the build, every year

Read the first line again. The ability for the AI to actually do things — the step from band 2 to band 3 — is priced on its own at up to $60,000.

And it keeps costing after launch

Build price is not the whole number. Kellton puts ongoing operation at $1,000–$15,000 a month, and annual maintenance at 15–25% of the build — $11,250–$18,750 a year on a $75,000 system before anyone writes a new line of code. Sharp Code and Seven Solvers both put an enterprise-tier bot at £2,000–£8,000 a month.

A system that is genuinely maintained — monitored, corrected, retrained on new information, improved as the business changes — carries a running cost everywhere in this market. Any quote without one is either excluding it or hiding it.

How the money is structured in practice

Underneath the ranges, the firms that do this seriously bill roughly the same shape:

• A year-one project fee. Design, build, rollout, and an agreed period of support and tooling — usually six to twelve months. This is the band 1, 2 or 3 number.
• Year-one support and tools, bundled into it. Monitoring, small fixes, and the infrastructure that keeps the system healthy.
• Optional years after that. Active maintenance beyond year one is a separate, scoped annual package, not a quiet assumption.
• A reconnection fee if you go dark. Stop support, come back later, and the system, its data and its guardrails have to be brought back to a current and safe state before anyone is allowed to rely on it again.

None of that is a subscription buried in a footer. It is a defined period of engineering and care around something that touches the real business — which is why an honest quote names the period as clearly as it names the price.

Which band do you actually need?

Most small businesses do not need band 3, and are quoted for it anyway. The useful question is not “how much does AI cost” but three specific ones:

• How much of your own information does it need to know? None, and you are in band 1. Your documents, prices and policies, and you are in band 2.
• Does it need to do things, or only say things? Actions, integrations and live systems are what move a project into band 3 — and Kellton prices that step at $20,000–$60,000 on its own.
• Who keeps it right next year? Every guide here charges for that separately. A build price with no answer to this question is incomplete.

Published 2026 pricing makes one thing unambiguous: professional custom AI is expensive, and it is expensive for reasons that have nothing to do with the model subscription. The engineering is the product.

The one thing you are really paying for

Every range, table and chart above collapses into one sentence. If you keep a single line from this article, keep this one:

You are not paying for access to an AI model. You are paying to turn a general model into a reliable system built around your business.

Hold it against the next proposal that lands. The ladder — and the number at the bottom of it — gets a great deal easier to read.

Sources

Six published 2026 pricing guides. Figures were read off these pages and converted at £1 = $1.27 where the guide quotes sterling. Prices move; if one of these has been revised since, the page will say so.

• Kellton — “Custom AI chatbot development with LLMs and RAG: 2026 cost and CX guide”. Updated 1 June 2026.Used for: $15,000–$300,000+ overall; $75,000–$120,000 mid-complexity; $1,000–$15,000/mo running; add-on and maintenance figures
• Cleveroad — “The In-Depth Chatbot Development Cost Guide for 2026”. Updated 8 January 2026.Used for: $8,000–$25,000+ lead qualification; $25,000–$70,000+ internal knowledge; $35,000–$120,000+ generative assistants; $50,000–$150,000+ workflow-driven
• Sharp Code — “AI Chatbot Development Cost UK 2026”. August 2026.Used for: £3,000–£6,000 FAQ; £8,000–£18,000 custom agent; £18,000–£40,000 multi-source; £60,000–£250,000 enterprise
• Fulminous — “AI Chatbot Cost in 2026: Complete Guide for Businesses in the UK”. Updated 13 August 2026.Used for: £8,000–£40,000 LLM chatbot; £15,000–£60,000 RAG; £30,000–£120,000 agent; £60,000–£250,000+ enterprise
• Emvigo — “RAG Platform Development Cost: 2026 UK Pricing Guide”. 10 August 2026.Used for: £8,000–£18,000 pilot; £20,000–£55,000 production; and the scope disclaimer quoted above
• Seven Solvers — “AI Chatbot Development Cost UK & USA 2026”. 17 April 2026.Used for: £8,000–£25,000 integrated UK; £25,000–£80,000+ enterprise UK; $45,000–$150,000+ USA enterprise agent