Insights
How much does custom AI agent development cost in the UK?
By Rare Agentics engineering teamLast updated First published 9 minute read
There is no honest single price: the cost of a custom AI agent depends on the task, the systems it connects to, the state of your data and how much it is allowed to do on its own. With Rare Agentics, an agent for one defined task takes 3 to 6 weeks at a fixed price agreed in writing, and the smallest paid starting point, a proof sprint, is from £3,000. Budget separately for running costs: model usage, hosting and someone to tune it.
What actually drives the cost of a custom AI agent?
An agent is not priced by the model behind it. The model is usually the cheapest part. What you pay for is the engineering around it: connecting it to your systems, getting your data into a shape it can use, deciding what it may do on its own, and proving it behaves before anyone relies on it.
Five things move the price more than anything else.
The task. One defined task with a clear owner is the cheapest agent to build and the easiest to prove. "Read supplier invoices, match each to its purchase order and draft the payment run for a person to approve" is a task. "Help the finance team" is not, and an agent scoped that loosely costs more and delivers less.
The systems it touches. Every system the agent reads from or writes to is integration work: authentication, permissions, error handling, and what happens when that system is slow or down. One inbox and one spreadsheet is a small job. A CRM, an accounting package and a document store that disagree with each other is a larger one.
The state of your data. If the answers live in clean, current documents, the agent can work from them. If they live in old PDFs, duplicated folders and people's heads, part of the budget goes on organising the material before the agent is any use. If the agent will handle personal data, allow time for a data protection impact assessment, which UK GDPR often requires for new automated processing.
How much it is allowed to do. An agent that suggests is cheaper than one that acts. Once it can change records or send messages, you need approvals, logging and tests for the rules that must never break. That is money well spent, but it is real work, and a quote that leaves it out is a quote for a demo.
How it is checked. An agent has to be run against real cases, and someone has to read the results. The more varied the cases and the more exceptions your process has, the longer this takes.
| Cost driver | Pushes the cost down | Pushes the cost up |
|---|---|---|
| The task | One rule-bound task with a clear owner | Several loosely defined jobs at once |
| Systems | One or two, with good APIs | Several, some with no API at all |
| Data | Clean, current, in one place | Scattered, duplicated or out of date |
| Autonomy | Observes or suggests | Acts on records, with approvals and logging |
| Checking | Narrow, predictable cases | Wide variety and many exceptions |
| Sensitivity | Internal, non-personal data | Personal, health or financial data |
If a supplier gives you a price without asking about most of the rows in that table, they are guessing, and you will pay for the guess later.
How much does a first agent cost with Rare Agentics?
We build a custom agent for one defined task in 3 to 6 weeks, at a fixed price agreed in writing before work starts. We do not print a single figure for it, because the drivers above vary too much between businesses, and a number that ignored them would be wrong for most of the people reading this.
What we do print: the smallest paid way to work with us is a proof sprint, from £3,000. It delivers one bounded feature, workflow or working prototype, and that can be a first agent on one task, built and running in your own accounts in 1 to 2 weeks. It proves the idea and the working relationship before either of us commits to more.
If you do not yet know which task should come first, the Agent Readiness Review is a two-week, fixed-fee review of where agents can safely take on work across your business. It ends with a ranked shortlist, the controls each task would need and a costed roadmap you can act on with us, with another partner or on your own.
Either way, the price comes before the work. Within two working days you get a first view, and within a week a written plan, a fixed price and the risks. Any change to scope is re-scoped and re-priced in writing before it is built.
What does an AI agent cost to run after launch?
The build is one line in the budget. Running it is another, and a supplier who does not raise it is doing you no favours.
- Model usage. Providers bill per token, roughly a few characters of text, for everything sent in and everything generated, and output usually costs several times more than input. Your bill depends on volume, how long the documents are and which model does the work. Ask for an estimate at your expected monthly volume, and ask for a spending cap to be set on the account.
- Hosting and storage. The agent, its logs and any search index it uses have to run somewhere. They should run in your own cloud accounts, so the bill is visible to you and the system stays yours.
- Monitoring and tuning. Someone has to read what the agent escalates, adjust its sources and update its rules when your process changes. Our Run and improve plan covers hosting, monitoring, agent tuning and support with agreed response times, and it is quoted to scope.
- Your own people's time. Payments, deletions and changes to customer records should always wait for a person. That is the right design, but approvals take minutes, so make sure they are quick to give.
One lever is worth knowing about. Not every step needs the largest model. Sorting, tagging and extraction often run well on a smaller, cheaper one, with the larger model kept for the steps that need it. Two more levers help: prompt caching, which charges far less for instructions and documents the agent reuses on every call, and batch processing, which is cheaper for work that does not need an instant answer. We use Anthropic, OpenAI or open models, the best one for each job, and we check what each provider keeps before it touches your data.
How do you work out whether an agent is worth building?
Start with hours, not with enthusiasm.
Take an illustrative task that takes 20 minutes and happens 300 times a month is 100 hours of work a month. If an agent prepares each case and a person spends two minutes checking and approving it, the approvals take 10 hours, and roughly 90 hours a month come back. Put your own cost on those hours, then compare a year of savings with the build price plus a year of running costs.
That arithmetic is deliberately plain. Count the softer gains too, such as faster turnaround and fewer keying errors, but do not build a business case on them alone.
The same sum also tells you which tasks to leave alone.
| Good first agent | Poor first agent |
|---|---|
| Happens many times a week | Happens a few times a year |
| Follows rules a person could write down | Needs fresh judgement every time |
| Easy to check before it lands | A mistake is only found after it causes harm |
| One owner who knows the task | Nobody quite owns it |
| Sources already exist in your systems | The knowledge lives only in someone's head |
Which pricing model protects you: fixed price, day rate or subscription?
There are three common ways suppliers charge for agent work, and each puts the risk in a different place.
Day rate, or time and materials. You pay for the hours worked. If the work overruns, you pay for the overrun. This suits genuinely open-ended research, but cap it, and ask for a written estimate with the assumptions spelled out.
Fixed price. The supplier carries the risk of overrun, so they need a tight scope before they start. That discipline is the point: if a supplier cannot write down what they are building, they cannot price it, and you should not start. This is how we work.
Platform subscription. Some vendors charge a monthly fee, often per seat or per action, for an agent that runs on their platform. It is cheap to start, but you are renting. Your rules and prompts live in their system, the cost grows with your team, and leaving can mean starting again. Ask what you keep if you cancel.
Whichever model you choose, ask what the price includes. Testing, documentation, handover and a warranty are often the first things cut from a low quote. Ours is in writing: any defect against the agreed acceptance criteria, found in the 30 days after release, is fixed at no cost.
What should you ask an AI agent supplier before you sign?
These questions separate a supplier who has shipped agents into real businesses from one who has built demos.
- What exactly will the agent be allowed to do on its own, and what will always wait for a person? Can we have that in writing?
- Whose accounts will the code, the data and the model keys live in?
- Can the agent, or any agent used to build it, reach our production systems? (Our view, set out in how to keep AI agents away from production, is that it should be impossible, not merely discouraged.)
- Is every action logged, with the agent that took it, the sources it used and the person who approved it?
- What happens when the agent's sources do not support an answer? It should go to a person instead of being sent.
- Which model provider will process our data, and what does that provider keep?
- What will it cost to run each month at our volume, and whose account pays the model bill?
- What is fixed, and what happens when the scope changes?
- What is covered after launch, and for how long?
Vague answers to the first three are the ones to worry about. The others can be fixed in a contract. Those three describe how the supplier thinks.
Where do agent budgets usually go wrong?
Starting too wide. An agent for "customer service" becomes a programme. An agent that drafts replies to one category of enquiry for a person to send is a project with an end date.
Paying for a demo. A demo works on the happy path. A system has to handle the exceptions, log every action, route unsupported answers to a person and survive the day a connected service changes. Make sure the quote is for the system.
Ignoring the data. If the documents an agent answers from are out of date, it will give out-of-date answers, confidently. Tidying the sources is part of the job, so price it in. The same is true of any first build; see how much an MVP costs in the UK.
No owner. An agent with nobody responsible for reading its escalations drifts. Name the owner before the build starts.
Lock-in by default. If the code, the data and the keys sit in the supplier's accounts, you are paying for the build and for the privilege of leaving later. Ask for your own accounts from the first commit.
How Rare Agentics can help
We build custom AI agents for one defined task in 3 to 6 weeks, with senior engineers leading the work, written bounds on what each agent may do and every action logged. If you want to test the idea first, a proof sprint starts from £3,000 at a fixed price. Tell us the task on our contact page and you will get a first view within two working days.