What does an AI agent cost a UK business to build and run?

Buying a ready-made AI agent means paying a usage fee. Intercom charges $0.99 per outcome for its Fin agent, and Microsoft 365 Copilot is £23.10 per user a month on enterprise plans, paid yearly, excluding VAT. Building your own means a one-off project, then model fees, hosting and staff time every month.

There’s no reliable public price list for custom builds. In our experience, an agent can cost as little as £500 to set up, but £5,000 is a realistic floor for a well-run project. Where yours lands depends on the state of your data, what the agent has to do and who checks its work.

This guide sets out what each route costs, what the pricing pages leave out and what to ask before you sign.

First, what are you actually paying for?

An AI agent is software that reads a request, decides what to do, then does it. That might mean answering a customer, updating a record or passing a case to a person. A chatbot mostly talks. An agent also acts, and every action means another system to connect, test and watch.

For budgeting, there are three routes:

• Buy. A vendor’s ready-made agent, paid for by usage or by seat. There’s little or no build cost.
• Configure. Agent tools inside a platform you already pay for, such as Microsoft 365.
• Build. A custom agent on a model provider’s API, the connection that lets your software use their AI, wired into your own systems.

Most teams don’t need a custom agent for their first one. Try the first two routes on a narrow job before paying for the third. Our guide to buying, configuring or building AI goes into that choice in more depth.

The four ways agents are priced

Vendors price agents in four main ways: per seat, per token, per task and by outcome. Each one puts the risk of a bad month in a different place.

A token is a small chunk of text, and model providers bill per million of them. Two quotes on different pricing models can’t be compared on the headline number alone. You have to convert both into the same unit, which we cover further down.

Buying an agent: two real price lists

Intercom Fin: you pay when it closes a conversation

Fin is Intercom’s customer service agent. Its pricing page sets out the terms:

• $0.99 per outcome, charged once per conversation even if Fin takes several actions.
• What counts as an outcome: a resolution, where no further help is requested after Fin’s last answer, or a handoff you’ve set up. Disqualifying a sales prospect, or sending one to self-serve, counts too.
• Sales qualifications cost more, at $9.99 each.
• No charge when a conversation is simply passed to your team without an outcome.
• With your existing helpdesk, such as Salesforce or HubSpot, there’s a minimum of 50 outcomes a month. Intercom says there are no setup, integration or platform fees, and no seat costs.
• With Intercom’s own helpdesk, you add $19 per helpdesk seat a month.
• The Pro add-on, for monitoring quality and Intercom’s Operator agent, starts at $100 a month.
• Billing is in US dollars only, so your cost in pounds moves with the exchange rate.

In practice, 1,000 outcomes in a month is a $990 bill. Your bill rises as Fin succeeds, which is fair. It also rises with volume, which is worth modelling before a busy season.

One thing to ask about: a resolution means nobody asked for more help. A customer who gives up and closes the chat hasn’t asked for more help either. Find out how the vendor tells the two apart.

Microsoft Copilot Studio: you pay per user, or per credit

If your team already lives in Microsoft 365, agents may be closer than you think. Microsoft’s Copilot Studio pricing works like this:

• Microsoft 365 Copilot is £23.10 per user a month on enterprise plans, paid yearly, excluding VAT. You need a qualifying Microsoft 365 plan first.
• Smaller firms can pay less with Microsoft 365 Copilot Business, at £16.10 per user a month, paid yearly, excluding VAT. It’s only for existing customers on an eligible Microsoft 365 Business plan. Bought by 31 December 2026, its first year is £13.80 per user a month.
• Internal agents are included. Licensed users can build and use agents inside Microsoft 365 at no extra cost.
• Outside channels cost extra. Publishing an agent to a website, app or social platform needs a standalone Copilot Studio plan. So does letting unlicensed users use it.
• Standalone plans are credit-based. Packs of 25,000 Copilot Credits cost £153.80 a month, or you can pay as you go.
• A pre-purchase plan lets you buy credits up front and save up to 20%.
• Standalone plans need an Azure subscription before agents can be used.

For a 20-person team on the enterprise licence, Copilot is £462 a month before VAT, or £5,544 a year. On Copilot Business, the same team pays £322 a month, or £3,864 a year. Credits are harder to forecast. Microsoft bills a varying number of credits, depending on what each action or response involves.

Side by side

For a full worked example of an agent handling customer queries, see AI customer service for small business.

Building your own: where the money goes

A custom agent makes sense when the job is specific to you. That might be acting inside your own product, joining up systems no vendor supports, or keeping sensitive data under your control. It costs more, and you get more control in return.

The build

The build cost is set by a handful of things, and the AI model is rarely the main one:

1. How many systems it connects to. Your CRM, booking system, finance tools and product each add work.
2. The state of your data. If the agent answers from documents nobody has tidied since 2019, that work comes first. Our guide to getting data ready for AI explains what’s involved.
3. What it’s allowed to do. Answering questions is cheaper than issuing refunds or changing records, because actions need checks.
4. Testing. It has to be tried against your real, awkward cases before customers see it.
5. Security and data protection. Personal data means checks before launch. Our AI and GDPR governance guide covers the basics.

In our experience, a simple agent can cost as little as £500 to set up. Realistically, £5,000 is the lower bound for a well-managed, organised project. Above that, the price depends mostly on the state of your data and what you want the agent to do. The honest figure comes from looking at your systems.

The running costs

Once it’s live, a custom agent has four monthly costs.

Model fees. Anthropic prices its Claude models per million tokens, and the range is wide. Input costs $0.10 per million tokens for Haiku 5.5 on shorter prompts, and $10 for Fable 5.1. Prices exclude tax, and work that can wait saves 50% through batch processing. We’ve turned those rates into pounds per conversation in how much an AI chatbot costs to run.

Agent extras. Agents often need more than the model. For its Managed Agents service, Anthropic charges $0.08 per session-hour, meaning each hour an agent is actively running, on top of token fees. Web search is $10 per 1,000 searches. Running code is free for 50 hours a day per organisation, then $0.05 an hour per container.

Hosting and monitoring. The agent runs somewhere, and someone needs to see when it goes wrong.

People. Someone reviews the cases it hands off, keeps its knowledge current and fixes connections when other systems change. Neither pricing page above includes this line, so make sure your budget does.

The costs that don’t appear on pricing pages

Run through this before you approve a budget, whichever route you choose:

• Human review. Hours spent checking handoffs and spot-checking answers.
• Knowledge upkeep. Updating the help content and policies the agent answers from.
• Broken connections. Fixing the link when another system changes how it works.
• Minimums and commitments. Fin’s 50-outcome monthly minimum, or Microsoft’s yearly licences.
• Exchange rate. Intercom bills only in dollars, and Anthropic’s API prices are in dollars.
• VAT. Microsoft’s listed prices exclude it.
• Quality add-ons. Monitoring tools are sometimes sold separately, like Fin’s Pro add-on.
• Data protection work. Checks before the agent touches personal data.

A low headline price is lovely until your volume turns up. Model the busy month, not the average one.

Buy, configure or build?

If the last row sounds familiar, try our AI readiness scorecard before spending anything.

Compare quotes on cost per job done well

Prices in seats, tokens, credits and outcomes can’t be compared directly. Put every option into one number: the cost of each job the agent completes properly.

1. Count how many conversations or tasks your team handles a month now.
2. Estimate how many the agent could take on.
3. Add up every monthly cost: fees, seats, credits, hosting and staff time.
4. Spread any build cost over a period you choose, such as a year.
5. Divide the total by the jobs it actually completes well.
6. Compare that with what the same job costs your team today.

If the agent doesn’t beat today’s cost per job, it isn’t saving money. It might still be worth it for speed or out-of-hours cover, but decide that on purpose.

What to ask your team or supplier

• What exactly counts as a billable outcome, task or credit, and how is it measured?
• What’s the minimum commitment, and is it monthly or yearly?
• What currency are we billed in, and is VAT included?
• Who reviews what the agent hands off, and how many hours a month is that?
• Which of our systems will it connect to, and who fixes those connections when they change?
• If you build it, who owns the code, the accounts and the instructions it runs on?
• Where is our data processed, and is it used to train anyone’s model?
• How will we know by month three whether it’s paying for itself?

A supplier who can’t answer these in plain English is telling you something.

Where Contic fits

If you’re weighing up an agent for an existing business, start with a free first chat. You’ll get guidance and advice on what we’d do, from people who do this all the time. If an off-the-shelf tool will do the job, that’s worth hearing before you spend anything. A paid discovery doesn’t have to follow. When it does, it can end in a strategy you take away and run with yourself, or in a scoped plan and a price.

Our AI team is AI-augmented, and senior engineers review and own everything that ships. You own the code and IP, and our contract assigns it to your company in writing. Whoever holds the accounts during the build, you get admin access to them.

Two examples from our work. For Everyone’s Invited, AI anonymisation of submissions saved about 12,500 hours of review, and 52,794 testimonies were published. For The Home Help Company, automating billing and payroll gave the CEO 30 hours a month back.

We’re registered with the ICO and hold Cyber Essentials Plus. We can also get you ready for GDPR and ISO 27001 audits. If data is your first worry, here’s how we keep your data safe when we use AI. To talk through a specific job, get in touch.