On this page
- What does each step cost?
- What does the build pay for?
- What does an AI agent cost per month to run?
- What does our price leave out?
- Is it worth the money?
- What should you ask about any vendor's savings claim?
- How does our price compare with other quotes?
- When should you not build a custom AI agent?
- Frequently asked questions
At XY Space, building an AI agent for one business process costs £5,000 fixed to see it working on your own files, usually £10,000 to £30,000 for the full build, and from £1,500 a month to keep it running. On our own internal operations inbox, model fees were a small part of the cost: about $18 a month on the model we used first and about $1.50 after we rebuilt it (model providers bill in US dollars).
Most pages that answer this question give a range and then ask you to book a call. We publish ours because you build most of the business case before you speak to anyone, and a price that only exists on a call is missing from that case.
What does each step cost?
| Step | Price | What you get |
|---|---|---|
| First call | Free | Thirty minutes. Bring one job your team repeats every week, and the spreadsheet or inbox it lives in. |
| One job, working | £5,000, fixed | About two weeks on your own files, so you see it work before you commit to more. If you go ahead, the £5,000 comes off the build. |
| The build | Usually £10,000 to £30,000 | We connect it to the systems you already run. Your team checks its work until you trust it. |
| Live operations | From £1,500 a month | We keep it accurate as your work changes, train the people who use it, and take on the next job. |
The build range is what our builds have actually sold for. Prices are in pounds, as they appear on our homepage.
The fixed first step exists because of how buyers decided in practice. One client took a fixed-price first job on the spot because it ran on their own exports. Another turned down a pitch that started with mapping, calling it "selling the invisible man": they could not see what they would be paying for until the work was done.
Before any of your files reach us, put a data processing agreement in place and agree what the first job is allowed to see. If your data protection impact assessment is not done yet, the first job can run on redacted or sample files.
What does the build pay for?
Most of the build cost is connecting the agent to the systems your work already lives in. The model call itself is a small part of the code, and the work around it is where the weeks go.
Connections to your systems of record. Your existing system stays the source of truth. The agent writes back through that system's own API, reads every write back to check it landed, and logs it. Where a system has no API, the agent operates its screens the way your staff do, under approval. Each extra system it has to read from or write to adds integration and testing time, and that is one of the main things that moves a build from the bottom of the range toward the top.
The review lane. Any message or record the agent sends outside your organization goes to a named person for approval, except standard replies you have approved in advance. Those limits sit in the system connections and the approval steps, because anything written into a prompt can be talked around. In the claims inbox we built for an insurer, a complete claim gets a standard acknowledgement automatically and is written to the CRM. An incomplete one is held out of the CRM and shown to a reviewer beside the source email, with a chase email drafted for that reviewer to approve. When the client replies, extraction runs again. Building that second path takes as much work as the first.
Testing on your own samples. We check the agent's work against your real documents, and re-test it on your samples whenever the model changes.
An audit trail. Every action is recorded on a tamper-evident log you can export for an auditor, a regulator or your insurer. Our safety model explains how it works.
Where it runs. The agent runs in your own cloud account, and the code, the rules and the data are yours. Each model call still goes to a model provider: an enterprise endpoint where your obligations allow it, or an open-weight model served in the UK or Canada. That provider belongs in your data processing agreement as a sub-processor, with its region and its data-use terms written down. Our guide to where AI inference runs compares the providers and regions on offer in Canada and the UK.
What does an AI agent cost per month to run?
Running an agent costs model fees, hosting and people's time, and on our own inbox, the one place we have measured it, the model fees were the smallest of those by a long way.
On our own internal operations inbox, the model we used first cost about $18 a month in model fees. After we rebuilt it on a cheaper decision model, the same work costs about $1.50 a month, roughly a twelfth of the original bill. That is one data point from our own inbox; your volume and documents will differ.
Hosting, and any document reading such as scanning PDFs, runs in your own cloud account and appears on your cloud provider's bill.
The larger monthly cost is keeping the agent accurate. Your forms change, a supplier changes their invoice layout, a new product line arrives, a model version is retired. Someone has to notice the drop, fix it, and re-test. That is what our Live operations line pays for, along with training the people who use it and building the next job when you are ready.
What does our price leave out?
Some costs sit on your side, whoever you hire. Budget for them alongside any quote:
- Your team's time. Someone answers questions during the build, checks the agent's work while trust is built, and keeps reviewing the cases it hands back.
- Running both ways at once. For the first weeks, the old process keeps running beside the new one so you can compare them. That doubles the work for a while.
- Security and supplier checks. Your security review, penetration testing if you require it, your data protection impact assessment, and any supplier onboarding or outsourcing checks your regulator expects. An agent that reads inbound email should be tested against messages written to trick it.
- Cloud and document processing. Billed by your provider, as above.
- People adopting it. The team whose job changes needs time to learn the new way of working.
Is it worth the money?
Work it out on hours your team spends today, measured before anything changes.
Here is a worked example using a real deployment. In the insurance claims inbox, three people each spent two hours a day reading claims email and retyping it into the CRM: 30 hours a week. After the build, one person spends 30 minutes a week on the claims the agent hands back. That frees 29.5 hours a week.
The hourly rate, build price and timing below are assumptions for illustration. Use your own fully loaded rate, and remember that freed hours only save money if the time goes to other work or the role changes.
| Year one | Year two | |
|---|---|---|
| Weeks of savings (48 working weeks a year, less 12 weeks of build and running both ways in year one) | 36 | 48 |
| Hours freed at 29.5 a week | 1,062 | 1,416 |
| Value at an assumed £30 an hour | £31,860 | £42,480 |
| Build, including the £5,000 first job (assumed £20,000) | £20,000 | £0 |
| Live operations at the £1,500 a month minimum | £18,000 | £18,000 |
| Total cost | £38,000 | £18,000 |
On those assumptions, year one is about £6,000 short and year two returns about £24,000, so the build pays back during the second year. Your own costs from the list above come on top.
A quick test before you call anyone: at £30 an hour and 48 working weeks, an agent has to free about 12.5 hours a week just to cover Live operations at its minimum. If the job your team repeats takes three hours a week, the arithmetic will not work, and we will say so on the first call.
What should you ask about any vendor's savings claim?
| Question | What we would want to hear |
|---|---|
| What was the baseline? | Hours or errors counted on the old process, over a stated period |
| Who measured it, and when? | Someone on your side, while the old way was still running |
| Does the running system report its own numbers? | Yes, from its own logs, without a consultant compiling them |
| Is the number able to go down? | Yes, and you would see it if it did |
The second question matters most. If the before-number came from the vendor's own model rather than from someone measuring the work while the old way was still running, the saving is a quote that nobody measured. It might still be right, but neither of you can check it.
How does our price compare with other quotes?
Search for this question and you will find development firms quoting ranges such as $20K to $60K for most businesses. That range is in US dollars, and converted to pounds it overlaps our build range, so the useful comparison is what each quote includes:
- What is being built. A new software product with its own interface, user accounts and app release is a bigger job than connecting an agent to the inbox and CRM you already have.
- Whether running it is included. A build-only quote leaves the cost of keeping it accurate for you to find out later.
- Who owns it. Ask whether you get the code or rent a platform, and what happens if you leave.
When should you not build a custom AI agent?
If your team already has Microsoft 365 Copilot or ChatGPT Enterprise and the job lives inside email and documents, try that first. Those tools can read your organization's mail and files, and Copilot Studio can build simple agents with connectors to common systems. A custom build earns its place when the job writes to a system of record, needs approvals and an audit trail around every action, or has to run inside your own cloud account.
If the job is mostly moving data between two SaaS tools with a small AI step in the middle, a workflow tool such as Zapier, Make or n8n will cost less and ship faster. If it is a fixed sequence of screens with no judgement, robotic process automation or an off-the-shelf document-processing product may be cheaper. We compare the trade-offs in n8n vs custom AI agents and on our comparison page.
Frequently asked questions
What happens to the £5,000 if we go ahead?
It comes off the price of the build.
Do we own the AI agent once it is built?
The code, the rules and the data are yours, and the agent runs in your own cloud account. Ask any vendor, us included, for the licence terms in writing before you sign.
Why not just use Copilot or ChatGPT?
If the job stays inside email and documents, you may not need anything else. When it has to write to a system of record under approval, with a log of every action, a seat on its own will not do the whole job.
If you have a job your team repeats every week, book the free 30-minute call and bring the spreadsheet or inbox it lives in.
Written by
Mahmoud Halat
Principal AI Solutions Architect, XY Space
Principal AI Solutions Architect at XY Space. A decade of regulated health-tech deployments across 100+ hospitals and 10M+ patient journeys, with a co-authored patent on patient journey reconstruction.
More from Mahmoud HalatShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project