How much does it cost to build an AI agent? For a UK mid-market business, agencies quote anywhere from around £6,000 for a configured workflow to £120,000 for a custom agent wired into core systems, plus a few hundred to a couple of thousand pounds a month to run it. That range is honest but useless on its own. The useful number is the full cost of ownership over two or three years, set against the cost of the work the agent removes.
This piece itemises that cost line by line, works through current model prices with real token counts, compares buying, configuring and building, and shows where our own prices sit. The model bill is almost always the smallest line on the page.
The model bill is almost always the smallest line on the page.
What agencies quote for an AI agent in the UK
Published agency and freelancer quotes for AI agent builds in the UK run from about £6,000 to £120,000, with monthly running fees, where they're stated at all, typically quoted at £400 to £3,500. The bottom of that range is usually a workflow on an automation platform with a model call or two inside it. The top is a custom agent integrated with several business systems, with approval steps, audit logging and an evaluation set.
For transparency, here are ours. A paid discovery week is £8,000 to £12,000 and is credited against what follows. An embedded sprint of two to three weeks is £20,000 to £35,000 and ends with the top one or two processes automated and running. A build of tooling, automation or agents is £30,000 to £120,000, fixed against an agreed outcome, over four to eight weeks. The running costs are yours and are stated up front, because they're your accounts.
Why the spread? Almost none of it is the AI. It's integration, controls and how much uncertainty the supplier is pricing in. Bain's 2026 CEO survey, published in September, found 82% of 100 CEOs said their AI transformation was falling short of ambition. A lot of that gap opens up between the build quote and what it actually takes to keep the thing working.
The full cost of ownership, line by line
Here is every line I'd put in a business case. Ranges are illustrative for a single-process agent in a business of £20m to £200m revenue.
| Cost line | What it covers | Illustrative range |
|---|---|---|
| Discovery | Mapping the process, timing the baseline, agreeing scope and success measures | £0 to £12,000 (sometimes folded into the build) |
| Build | Designing and coding the workflow, prompts, rules, exception handling and approval steps | £5,000 to £100,000 |
| Integration | Connecting to the ERP, CRM, inbox or document store; often the biggest single piece | 30% to 50% of the build |
| Evaluation and testing | Building a test set of real historical cases, measuring accuracy, parallel running | 10% to 20% of the build |
| Internal time | Your process owner, IT and the team: requirements, testing, training | 5 to 20 days at loaded cost |
| Model usage | API charges for the model calls, per item processed | Tens to low hundreds of pounds a month at mid-market volumes |
| Hosting | Where the agent runs, its database and logs | £50 to £500 a month |
| Licences | Automation platform or SaaS seats, if used | £0 to £2,000+ a month |
| Monitoring | Someone watching the exception queue, cost per task and accuracy | 2 to 8 hours a month |
| Maintenance | Supplier format changes, upstream system updates, prompt and rule changes | 1 to 3 days a month in year one, falling after |
Run costs matter more than the build over any sensible horizon. A cheap build that needs a day a week of babysitting isn't cheap. And internal time always gets left out, which flatters every business case. Cost it at the same loaded rate you use for the benefit.
AI agent running costs: a worked token example
Model prices are published per million tokens (a token is roughly three quarters of an English word). As of 4 October 2026, Anthropic's published API prices are $2 per million input tokens and $10 per million output tokens for Claude Sonnet 5.5, $1 and $5 for Claude Haiku 4.5, and $4 and $20 for Claude Opus 5.5. Cached input, the repeated instructions sent with every call, is billed at a fraction of the input price: $0.20 per million on both Sonnet 5.5 and Opus 5.5. The Batch API halves prices for work that doesn't need an instant answer.
Take an illustrative agent that triages a customer service inbox: reads each email, looks up the account, classifies it, and either drafts a reply for approval or routes it with a summary. Volume is 4,000 emails a month. Per email, across the agent's steps:
- 6,000 tokens of instructions, policies and tool definitions, served from cache.
- 6,000 tokens of fresh input: the email thread and the account data the agent pulls in.
- 1,000 tokens of output: the classification, the summary and the draft.
On Sonnet 5.5 that's 6,000 × $2 per million ($0.012), plus 6,000 × $0.20 per million ($0.0012), plus 1,000 × $10 per million ($0.010). About $0.023 per email, or roughly $93 a month for 4,000. At around $1.32 to the pound, call it £70. Add 30% for retries, cache writes and rerunning the evaluation set, and you're near £90 a month. The same workload on Opus 5.5 comes to roughly £180 a month on the same basis; routing the simple classification step to Haiku would bring Sonnet's figure down further.
Two caveats. The pricing page notes that newer Claude models use a tokenizer that produces about 30% more tokens for the same text, so measure real token counts on your own documents rather than estimating from word counts. And agents that loop (call a tool, read the result, decide, call another) multiply input tokens with every step. That's the main reason a badly designed agent costs ten times a well designed one. But even at ten times, this agent's model bill is smaller than one day a month of someone maintaining it.
AI agent cost vs salary
The search people actually run is "AI agent cost vs salary", so here's the honest version. The ONS puts median gross annual earnings for full-time UK employees at £39,039 in April 2025. Add employer's National Insurance and pension and a median role costs roughly £45,000 a year before desk, equipment and management time.
A well-scoped agent might return 150 to 300 hours a month, which is one to two people's worth of time. On a year-one cost of £30,000 to £60,000 including the build, that looks like an easy win, and in years two and three it looks better still.
The comparison is still misleading. An agent removes tasks, not people. The hours it returns are scattered across a team, and they only become cash if you redeploy them, don't backfill a leaver, or absorb growth without hiring. I've set out how to turn hours into a number finance will sign in how to measure AI ROI. Use the salary comparison as a sense check on scale, never as the business case.
Build vs buy AI: SaaS, platform or custom
There are three ways to get an agent into a process, and each wins in different conditions.
| Buy SaaS | Configure a platform | Custom build | |
|---|---|---|---|
| What it is | A product with AI built in: helpdesk, AP, expense or credit control tool | A workflow or agent platform, configured by you or a partner, calling a model | Code written for your process, in your accounts and repos |
| Upfront | Low: setup and internal time | Medium | Highest |
| Monthly | Highest: seats or per-task fees that scale with volume | Medium: platform licence plus model usage | Lowest per item: hosting and model usage, plus maintenance |
| Wins when | The job is common to every business and the product fits your systems | The process is yours but the systems and integrations are standard | Volume is high, the process is distinctive, or integrations are awkward |
| Loses when | Your process doesn't fit its assumptions, or per-task fees grow with volume | Logic outgrows the platform and becomes unmaintainable | Volume is low, or nobody will own it after the builders leave |
Whichever you choose, insist on owning your data, your prompts and your evaluation set. Those are what let you switch later.
A worked example in pounds
Back to the service inbox: 4,000 emails a month, currently about four minutes each to read, look up, answer or route. That's roughly 267 hours a month. All figures below are illustrative, with hours valued at £24 (a £40,000 loaded cost over about 1,680 productive hours).
| Over three years | Buy SaaS | Configure a platform | Custom build |
|---|---|---|---|
| Upfront (including internal time) | £5,000 | £20,000 | £51,000 |
| Monthly run cost | £1,200 | £700 | £1,400 |
| Hours returned per month | 100 | 130 | 160 |
| Monthly value of hours | £2,400 | £3,120 | £3,840 |
| Payback | 4 months | 8 months | 21 months |
| Three-year net value | £38,200 | £67,120 | £36,840 |
At this volume, the configured platform wins comfortably. The SaaS tool is quickest to pay back but its per-task fees eat the value, and it returns fewer hours because it doesn't know your account structure. The custom build returns the most hours but carries too much upfront cost for 4,000 emails a month.
Triple the volume and the picture changes. SaaS fees scale with it; the custom build's costs barely move. On the same assumptions at 12,000 emails a month, custom edges ahead of the platform and leaves SaaS well behind. Which is the general rule: custom wins on volume, distinctiveness and awkward integration, and almost nowhere else.
Red flags in an AI agent quote
Most bad AI spend is visible in the quote, if you know where to look.
- A price before anyone has seen the process. No discovery, no baseline, no timed sample. They're pricing a demo, not your work.
- No running cost line. If the quote doesn't state model, hosting and maintenance costs, you'll find out at month three.
- Time and materials with no defined outcome. Open-ended day rates move all the risk to you.
- No evaluation set or acceptance criteria. "It works" isn't a test. Ask how accuracy will be measured before go-live and after every change.
- They own it. Code, prompts and accounts in the supplier's tenancy means you're renting, not buying.
- "Digital worker" pricing. A monthly fee per agent pegged to a salary is priced on your saving, not their cost.
- Fully autonomous, no approvals. No human approval before irreversible actions and no exception queue is a control failure waiting for a date.
- Integration hand-waved. "We'll connect to your ERP" is usually where half the cost hides.
Gartner has already predicted, as reported by MarTech, that more than 40% of agentic AI projects will be cancelled by the end of 2027, and estimated that only about 130 of the thousands of vendors claiming agentic AI offer the real thing. A sceptical read of the quote is cheaper than becoming part of that statistic. For where agents genuinely fit, see AI agents for business operations; for a finance-specific list of where they pay back, see AI for finance teams.
What a sensible budget looks like
For a first agent in a mid-market business: a discovery or baseline phase, a fixed-price build sized to one process, a stated monthly run cost, an owner in the business, and a payback you'd still accept at half the expected benefit. If the numbers only work at full capture, scope smaller. If they work at half, build it.
That's how we price every embedded AI delivery engagement: fixed against an agreed outcome, running costs stated before we start, and everything built in your accounts so nothing depends on us afterwards.
Questions
How much does it cost to build an AI agent in the UK?
Agencies quote anywhere from around £6,000 for a configured workflow to £120,000 for a custom agent integrated with core systems. Our own fixed-price band for tooling, automation and agents is £30,000 to £120,000 over four to eight weeks. Add running costs, typically a few hundred to a couple of thousand pounds a month once hosting, monitoring and maintenance are counted, not just the model.
What are the monthly running costs of an AI agent?
Running costs cover model usage, hosting, any platform or SaaS licence, monitoring and maintenance time. For a mid-market process handling a few thousand items a month, model usage is often under £150 a month at current API prices. The larger lines are usually licences and the hours someone spends reviewing exceptions, updating rules and fixing integrations when upstream systems change.
Is an AI agent cheaper than hiring someone?
Sometimes, but it's the wrong comparison. An agent removes tasks, not people. A well-built agent can return the hours of roughly one full-time person for less than that person's annual cost, but the saving only becomes cash if you redeploy the time, don't backfill a leaver, or absorb growth without hiring. Compare the agent's full cost of ownership with the value of the hours it removes.
Should we build or buy an AI agent?
Buy when the job is common to every business, such as helpdesk triage, expense checks or invoice capture, and a product already does it well in your systems. Configure a platform when the process is yours but the integrations are standard. Build custom when volume is high, the process is distinctive, or the integrations are awkward enough that off-the-shelf tools can't reach the data.
