What sets how much an AI agent costs
The cost to build an AI agent depends on two things: how many of your systems it touches, and how much it may do without asking a person. The cost to run it has four lines: model usage, hosting, staff review time and upkeep. Model usage is usually the smallest. Review time and upkeep are where budgets go wrong.
We don't publish prices, because every agent we build is quoted against a written scope. This guide explains what moves the number instead. It is written for owners and operations leads at UK and EU firms of 10 to 200 people. You have been told to "do something with AI", and you want to know what you would be paying for.
In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027. The reasons it gave were escalating costs, unclear business value and inadequate risk controls. Each of those can be checked before anything is built, and the rest of this guide shows how.
What sets the cost of building an AI agent
Build cost follows the number of systems the agent touches and how much it may do alone. The choice of model barely moves it. These are the three sizes of AI agent we build, from the least work to the most.
| Size | What it does | Example | Where the build effort goes |
|---|---|---|---|
| Drafting assistant | Reads, looks things up and prepares the work. A person sends or saves it. | Drafts replies to supplier emails from your order system | Read access to one system, and a screen where a person approves each draft |
| Agent that acts in one system | Acts alone inside set limits and asks a person above them | Updates customer records and books repair visits. Any job above a set budget goes to a manager. | Write access, limits, an approval queue, logs and tests on past cases |
| Agent across several systems | Works across your inbox, your documents and two or more business systems | Matches supplier invoices to purchase orders and chases missing paperwork by email | All of the above for each system, plus rules for when the systems disagree |
The jump between sizes comes from permissions and testing. A drafting assistant can be wrong cheaply, because a person reads everything it produces. An agent that acts needs limits, a review queue and logs before anyone should trust it.
An agent that only talks is a chatbot. So every agent we build has the same five parts, whatever its size:
- A connection to your live systems, not a copy of them.
- A written scope of what it may do alone.
- An approval step wherever a mistake is expensive.
- Escalation with the context attached, so the person picking it up is not starting cold.
- A log of what it did and why.
What the build budget pays for
Most of an agent's build cost is ordinary software work. The table shows where the time goes and what makes each part grow.
| Part of the build | What it covers | What makes it grow |
|---|---|---|
| Mapping the job | The steps, rules and exceptions, written down with the people who do the work today | Rules nobody has written down, and exceptions only one person knows |
| Connections | Read and write access to your email, CRM, accounting or booking system | Old systems with no API; systems that charge for a test account |
| Readable inputs | Turning scans, PDFs and attachments into text the model can read | Handwriting, photos of paper, several languages in one document |
| Limits and approvals | What it may do alone, what needs a person, and a review screen for the rest | More kinds of action, and more people who approve |
| Logs | Every decision stored with its reason | Audit or regulatory requirements |
| Testing | Running it on real past cases before it touches live work | Rare mistakes that are expensive, each needing its own test cases |
| Handover | Documentation, with every account in your company's name | Nothing, if it is planned from the first day |
Two of these lines are easy to underprice. The first is readable inputs: ask an agent to judge something it cannot read, such as a scanned invoice with no text layer, and it will decline or guess. A decline looks like bad judgement, when the real fault is a missing step. So we convert inputs when they arrive, before the agent sees them.
The second is logs. Asking a model to explain itself in prose works until the output gets long. Then the explanation goes missing, and nothing tells you it has gone.
We store each decision as a required field instead, with a short list of allowed values and a reason. A missing field is easy to detect. A missing sentence is not.
Demos skip all seven lines, which is why they are cheap. Generating something that looks finished has never been easier. The work is in the second month.
Two public failures that show where the budget goes
Limits and testing take a large share of any serious build. Two documented cases show why.
In 2025 Anthropic let a Claude model run a small shop in its office. Staff talked it into discount codes, it gave some items away, and it sold metal cubes for less than it had paid for them. Anthropic's write-up of the experiment concludes that it would not hire the agent to run the shop.
Each of those mistakes is a missing limit: a floor on prices, a rule for discounts, an approval above a set amount. Writing and testing limits like those is build work.
The second case is the agent that deleted a live database during a code freeze, which we walk through in AI agents vs workflows. The lesson carries over to business agents. A limit written in the instructions is a request; a limit in the permissions holds.
Anthropic's guidance on building agents says the same in general terms: autonomy brings higher costs and the risk of compounding errors, so test extensively in sandboxed environments, with guardrails. That testing is a line in the build. A quote without it is pricing a demo.
What an AI agent costs to run each month
Running costs have four lines. Only the first appears on the AI provider's bill.
| Line | What it covers | What sets it |
|---|---|---|
| Model usage | The text you send to the model provider and get back, billed by the token | Volume, the model, and how much context each task sends |
| Hosting | The server or functions that run the agent, plus a database for logs and the review queue | How often it runs and how long you keep the logs |
| Review time | Staff time on the cases the agent hands to a person | How narrow the job is and where the limits sit |
| Upkeep | Monitoring, fixes when a connected system changes, re-testing when a model changes | How often your systems and the models change, and whether the build left test cases behind |
A token is a piece of text, about three-quarters of an English word by Anthropic's rule of thumb. Agents use far more of them than chat does. In Anthropic's own data, agents used about 4 times as many tokens as chat interactions. What your team spends on chat subscriptions is no guide to what an agent will use.
For the price of a token, go to the source. Anthropic, OpenAI and Google each publish theirs. We don't copy the numbers here because they move: Google's page already lists price rises for several of its Flash models from 1 January 2027. Anthropic and OpenAI also list discounts for batch processing and for input the model has seen before, which suit an agent that sends the same instructions on every call.
Upkeep is not optional, because the models underneath change. Anthropic's model deprecations page lists six Claude models retired in 2026 so far, and requests to a retired model fail.
Each retirement means moving to a newer model and testing again. If the test cases from the build still exist, that is a round of checks. If they don't, it is close to a rebuild.
A worked example: 2,000 emails a month
Say you run a 25-person property management company. About 2,000 tenant and contractor emails arrive each month, and each takes someone about 3 minutes to read, look up and deal with. That is 100 hours of staff time a month.
The agent reads each email and looks up the tenant and property in your management system. Then it files the email, drafts a reply for a person to send, or books a repair within a set budget.
Assume three model calls per email. Each call sends about 8,000 tokens (the instructions, the tool definitions, the email thread and the records it looked up) and gets back about 500.
| Line | Volume | How it was sized |
|---|---|---|
| Emails handled | 2,000 a month | Your own count |
| Model calls | 6,000 a month | Three per email |
| Tokens sent and received | 48 million in, 3 million out, each month | Price these on the provider's page |
| Emails handed to a person, 1 in 10 | 200 a month, about 13 hours | 4 minutes each, as these are the harder cases |
| Emails handed to a person, 3 in 10 | 600 a month, about 40 hours | 4 minutes each |
| First weeks, a person checks every action | 2,000 checks, about 33 hours | 1 minute each, until the error rate is known |
Token counts differ between models for the same text: Anthropic says its newer tokenizer produces about 30% more tokens. Test a sample of your own emails before you trust any estimate.
Now compare the lines. At UK office pay, 13 hours of review cost more than a month of model usage on any mid-priced model listed on those pages in September 2026.
The review line is also the one that moves. If the limits send 3 emails in 10 to a person, review triples to 40 hours. Switching to the most expensive model changes the model line and leaves the hours where they were. So a cheaper model that hands more cases to a person can cost more overall.
The cheapest agent has the narrowest job and the clearest limits; model choice comes second. In a good month, your team's time on this job falls from 100 hours to about 13. That saving is what to weigh against the quote for the build and the running lines.
UK and EU costs most price guides leave out
If your agent reads personal data or talks to customers, three costs apply that most price guides skip. Each costs less when it is planned from the start.
- A data protection impact assessment. The ICO's guidance on AI and data protection says that in the vast majority of cases, using AI triggers the legal requirement for one. You also need a data processing agreement with the model provider, as with any other processor.
- Telling people they are talking to an AI. Since 2 August 2026, Article 50 of the EU AI Act has required AI systems that talk to people to tell them so, unless it is obvious. For a customer-facing agent, that is a line of copy and a design decision.
- Keeping data in a region. Some providers charge more when requests must be processed in a set region; OpenAI's pricing page lists an uplift for data residency, for example. Ask where each request is processed before you sign.
An agent that works only on internal documents, with no personal data, avoids most of this.
Renting an agent instead of building one
For some jobs you can rent an agent. Fin, a customer service agent, charges per resolved outcome with a monthly minimum, works with several helpdesks and needs no build. Renting wins when the job is standard and the product already connects to your tools. Answering customer questions from a help centre is the obvious case.
Building wins when the agent must act inside your own systems or follow rules only your firm has. To compare on volume, multiply the per-outcome price on the vendor's page by your monthly count. Then set that beside the running lines of a quote for a built agent.
A rented agent's rules and logs live in someone else's product. Ask how you would leave before you start.
When you don't need an AI agent
Anthropic's guidance for developers starts from the simplest approach that works, which it says "might mean not building agentic systems at all". Gartner says many use cases sold as agentic don't need an agent. Skip the agent in these cases.
- The job follows fixed rules. If every case goes the same way, a plain workflow with no model does it for less and never guesses. Our guide to AI agents vs workflows explains why most business agents should be plain automations.
- The volume is low. At 3 minutes each, thirty cases a month is an hour and a half of someone's time. The build will not pay for itself.
- Nobody has written the process down. Automating a broken process makes it break faster. Map it first, delete the steps nobody needs, then automate what is left.
- A mistake is expensive and hard to spot. Payments, legal wording and medical advice belong with a person, with an assistant that drafts at most. Our checklist of what not to automate with AI goes through these cases one by one.
- A rented product already does the job well enough.
Saying so is part of the job. When a client asks us to review a plan before anyone builds, the result is a written recommendation they own. "You don't need this yet" is one of the answers it can give, and the recommendation stays useful if they take the work elsewhere.
What a proper quote for an AI agent contains
A quote you can compare with others names every line in this guide. Look for these, and ask for any that are missing.
- A written scope: the job, what the agent may do alone, and where a person decides.
- The build, split into its parts: mapping, connections, inputs, limits and approvals, logs, testing and handover.
- Model usage estimated from your own volumes, with the provider and model named, so you can check it against the pricing page.
- The expected review hours at the planned hand-off rate, and in the first weeks.
- Hosting and upkeep, and what upkeep covers when a model is retired or a connected system changes.
- How the agent will be tested before it touches live work, and on which past cases.
- Who owns the code, the accounts, the logs and the documentation. The answer should be you, with everything in your company's name.
A quote that prices only the build and the model is pricing the demo. So is one that can't say how many cases will reach a person.
Five facts to write down before you ask for a quote
Any honest quote starts from five facts. Write them down before you talk to anyone, including us.
- The job in one sentence, with a verb: "match supplier invoices to purchase orders".
- How many times a month it happens, and how long it takes a person today.
- Which systems the agent must read, and which it must change.
- What it may do without asking, and the limit above which a person decides.
- What one mistake costs, in money or in trust with a customer. In 2024 a Canadian tribunal held Air Canada liable after its website chatbot wrongly promised a customer a bereavement refund (Moffatt v Air Canada).
With those five answers, a builder can quote the build and estimate the running lines in one conversation. Without them, any number you hear is a guess about the demo.
Start with the second fact this week. Count how many times the job happens and how long each one takes. If the total is under an hour a month, you don't need an agent for it.
Questions
Why don't you publish prices for AI agents?
Because two things sold as an "AI agent" can be very different jobs. One drafts replies for a person to send. Another acts in three of your systems, with approvals and an audit log. A published range would be too wide to plan with, or wrong for your job. We quote each project against a written scope, and the quote names every running line, so you can compare it with anyone else's.
Can we use ChatGPT or Copilot instead of a custom agent?
For drafting inside one person's own work, often yes, and it costs a monthly seat rather than a build. A custom agent makes sense when the work has to happen inside your systems, on its own schedule, with limits and logs your firm controls. Start with the seat licence and build when you hit its edges.
Who owns the agent once it is built?
You do. The code, the hosting, the repository and the documentation sit in your company's own accounts and move to you at handover. If you stop working with us, the agent keeps running and another team can take it over. Ask any builder the same question before you sign.
How we wrote this: drafted with AI, then checked and edited by the people who build our projects. Figures come from our own work or from the linked sources.