How much does an AI agent cost?

First, what you are actually being sold

The word "agent" has been stretched so far it barely means anything, so it is worth knowing what the companies building this technology mean by it, because they are stricter than the people selling it to you.

Anthropic draws the line like this. A workflow is a system where the language model is orchestrated through predefined paths — it does what it is told, in an order somebody set. An agent is a system where the model "dynamically directs its own processes and tool usage, maintaining control over how it accomplishes tasks". Google publishes a three-way version: a bot follows rules, an assistant responds to you and recommends while you decide, an agent plans and acts on its own. Microsoft, for what it is worth, collapses the distinction entirely — for them an agent simply is an assistant. OpenAI publishes no crisp definition at all.

Here is the part that matters to you. A chat box on your website that answers questions from your own information is not an agent by any of those definitions. It is retrieval plus a single model call — the simplest tier there is. That includes the one I am building. If somebody is quoting you tens of thousands for an "AI agent" and what they are describing is a website FAQ answerer, the price and the label are both doing work.

Gartner put a number on this in June 2025. Of the thousands of vendors claiming agentic AI, it estimates only around 130 are genuine — the rest is what it calls "agent washing", rebranded chatbots and automation. It also predicts over 40% of agentic AI projects will be cancelled by the end of 2027, on cost, unclear value and inadequate risk controls.

What it costs to buy one

Published prices, checked on 15 September 2026. Nearly all of these are in dollars even for UK buyers — only Zendesk publishes in pounds — so you are paying an exchange rate on top.

Notice what happened to the pricing model. The industry has moved to charging per outcome — Intercom bills one charge per conversation where the customer did not need further help, and explicitly does not bill when a conversation is handed to your team unresolved. Zendesk says you "pay only for customer requests that were successfully resolved by the AI agent".

That is genuinely fairer if you are a company with a big support queue. For a small business it quietly inverts, because success becomes the thing you pay for. Run a thousand conversations a month at a 50% resolution rate and the same work costs roughly £367 a month on Intercom, £167 on HubSpot, or £41 on Crisp. A ninefold spread, for a FAQ answerer.

What it costs to build one

Nobody publishes this next bit, so here is the arithmetic with the assumptions on the table, and you can argue with them.

Assume a conversation is four back-and-forth turns. Each turn sends the model your instructions plus the relevant chunk of your own website content — call it 3,000 words' worth, growing as the conversation goes on, so around 15,000 tokens of input across the conversation. Each reply is roughly 150 tokens. Embedding your whole website once, so it can be searched, is a one-off of well under a pound.

Longer conversations cost more, obviously — double the turns and the top figure roughly triples. Caching the repeated part of the instructions cuts it by more than half again. Hosting can be free at this size, and the search index a small business needs sits inside the free tiers at both Pinecone and OpenAI.

One counter-intuitive thing worth knowing. If you let the assistant search the live web as well as your own content, that is charged separately at around $10 per thousand searches — which is nearly three times the entire token cost of running a thousand conversations. The cheap part is the AI. The expensive part is letting it wander.

A quick honesty check on the quotes going round: one agency's published guide prices hosting for an AI agent at $500 to $10,000 a month, and then, further down the same page, cites a live deployment running at two cents a conversation. Both numbers cannot be right.

So why would anyone pay £367 a month?

Because the compute is the smallest line on the bill, and pretending otherwise would be dishonest.

What the off-the-shelf price buys you is the chat widget, the dashboard, the hosting, the handover to a human when it gets stuck, the analytics, the data protection paperwork, and somebody to ring when it breaks. Building it yourself means you own all of that. The real cost of a custom build is not tokens — it is the build itself, keeping the approved information accurate as your prices and services change, and actually reading what the thing said to your customers last week.

Anthropic's own engineering guidance says much the same: find the simplest solution possible, and note that agentic systems "often trade latency and cost for better task performance". The complexity is the expense. Not the electricity.

What it will get wrong

Every vendor page says its assistant answers only from your content and does not make things up. Here is what the measurements actually show.

Stanford tested three commercial legal research products built on exactly this grounded-retrieval approach, across 202 queries. They produced incorrect or unsupported answers between 17% and 33% of the time. The researchers' conclusion was that retrieval reduces invention but does not eliminate it, and that errors "remain substantial, wide-ranging, and potentially insidious". In fairness that is a hard adversarial domain and the study is from 2024; a plumber's FAQ is a much easier job and the real rate should be far lower.

So take the easier measure instead. Vectara's leaderboard tests how often a model invents something when it is handed the correct document and asked only to summarise it — the grounding task under ideal conditions. The best models on it drift 1.8% to 3.3% of the time. At a thousand conversations a month, that is roughly 20 to 30 answers a month that are not fully supported by what you actually wrote.

OpenAI has published its own explanation of why: accuracy "will never reach 100%", and standard training rewards guessing over admitting uncertainty. The fix is teaching a model to say it does not know and hand over — which is exactly what a business assistant needs and what most products are not tuned for.

And who is responsible when it does

Four documented cases, because this is the part that decides whether you want one at all.

Air Canada, February 2024. Its chatbot invented a bereavement fare policy. The airline argued in a Canadian tribunal that the chatbot was "a separate legal entity responsible for its own actions". The tribunal rejected that: the chatbot "was just a part of Air Canada's website and Air Canada still bore responsibility for all the information on its website, whether it came from a static page or a chatbot." Canadian, small-claims scale, and not binding here — but it is the case everyone cites.

Cursor, April 2025. An AI company's own support bot, signing itself "Sam", invented a subscription rule limiting customers to one machine. People cancelled over a policy that had never existed. The co-founder's response: "We have no such policy… this is an incorrect response from a front-line AI support bot." If a well-funded AI business cannot stop its own bot inventing its own terms, the risk is not theoretical.

DPD, January 2024. A UK case, and a brand one rather than a legal one. DPD disabled part of its chatbot after it swore at a customer and called DPD "the worst delivery firm in the world".

A Chevrolet dealership, December 2023. Its chatbot was talked into agreeing to sell a car for one dollar, adding "that's a legally binding offer – no takesies backsies." Nobody enforced it. It is still a good illustration of what a bored person can do to a chat box on your website.

For a UK business the relevant law is not an AI law, because there isn't one. The UK has deliberately not passed an EU-style AI act. What applies is ordinary consumer law, and in particular section 50 of the Consumer Rights Act 2015, which treats a service contract as including "anything that is said or written to the consumer, by or on behalf of the trader, about the trader or the service" where the consumer relied on it. Nothing in that wording distinguishes a person, a web page or a chat box. That is my reading rather than a decided case — no UK court has ruled on it — but it is the argument a customer would make, and it is why the Air Canada outcome is the one to plan for.

Two smaller points, since most guides get them wrong. The EU AI Act's transparency rule has applied since 2 August 2026 and was not delayed by the Digital Omnibus, whatever you may have read; it requires telling people they are talking to an AI, which costs nothing and you should do anyway. And using AI does not automatically require a data protection impact assessment — the ICO requires one where innovative technology is combined with another trigger such as systematic monitoring or sensitive data. Plenty of articles claim otherwise.

What I would actually do

If you have a handful of enquiries a week, none of this is your priority. An assistant that answers questions nobody is asking is an expensive ornament, and the thing that will move the needle is being findable in the first place.

If you are fielding the same ten questions every day — capacity, price ranges, areas covered, what is included, how soon you can come out — then the honest first step is not an assistant at all. Write the answers on your website. That fixes it for the customer who never opens a chat box, it is what an assistant would have to read anyway, and it is what search engines and AI tools quote when someone asks them about you. An assistant on top of that is a genuine improvement. An assistant instead of it is a way of hiding your own information behind a widget.

And if you do put one on, label it as AI, keep a log of what it said, and read the log.

Common questions

What is the difference between an AI agent and a chatbot?

Anthropic defines an agent as a system where the model directs its own processes and tool use; a workflow follows paths someone set in advance. Google splits it three ways: a bot follows rules, an assistant responds and recommends while you decide, an agent plans and acts on its own. A chat box that answers questions from your own information is not an agent by any of those definitions.

What does it actually cost to run a website AI assistant?

For a thousand conversations a month, the model itself costs roughly £3 to £14 depending which one you use. Hosting and the search index for a small business site fit inside free tiers. Buying the same capability off the shelf runs about £41 to £367 a month, depending entirely on which product you pick.

Why do agencies quote thousands for one?

Because the compute is the smallest line on the bill. What you are really paying for is the build, the widget and dashboard, the handover to a human, and keeping the approved information accurate as your prices change. That said, the same deliverable is quoted at £0, £3,000, £8,000 and £60,000 by different UK publishers in the same year, so the range is about who is selling rather than what is being built.

Will an AI assistant make things up about my business?

Sometimes, yes, and anyone saying otherwise is selling. Stanford tested three commercial products built on grounded retrieval and found incorrect or unsupported answers 17 to 33 per cent of the time. On the easier task of summarising a document it has been handed, the best models still drift 1.8 to 3.3 per cent of the time — at a thousand conversations a month, roughly 20 to 30 answers not fully supported by what you wrote.

Am I legally responsible for what my chatbot tells a customer?

In the UK there is no AI act and no decided case, but ordinary consumer law applies. Section 50 of the Consumer Rights Act 2015 treats a service contract as including anything said or written to the consumer on the trader's behalf, and nothing in that wording distinguishes a person from a chat box. A Canadian tribunal reached exactly that conclusion about Air Canada's chatbot in 2024.

Do I need to tell people they are talking to an AI?

Yes, and it costs nothing. The EU AI Act's transparency requirement has applied since 2 August 2026 and was not delayed by the Digital Omnibus, whatever you may have read. It only bites if your output is used in the EU, but labelling the assistant is good practice regardless.

Where I stand on this

I am building a website assistant myself, so treat everything above with that in mind. It is not an agent, by the definitions at the top of this page. It answers from information a business has approved, it is designed to say it does not know rather than guess, and it will still occasionally get something wrong, because all of them do. Anyone telling you otherwise is selling.

Prices and sources checked 15 September 2026: OpenAI, Anthropic and Google Gemini published API pricing; Intercom, Fin, Zendesk, HubSpot, Tidio, Crisp and Chatbase pricing pages; Anthropic "Building Effective AI Agents"; Google Cloud and Microsoft Copilot Studio documentation; Gartner press release, 25 June 2025; Stanford RegLab legal RAG study; Vectara hallucination leaderboard; OpenAI "Why Language Models Hallucinate"; BC Civil Resolution Tribunal in Moffatt v Air Canada; The Register; ITV News; Consumer Rights Act 2015 s.50; ICO DPIA guidance; European Commission Article 50 guidance. Sterling figures converted at £1 = $1.35 and are my calculation, not vendors' own pricing. Gemini Flash pricing is scheduled to double on 1 January 2027.