Build costs run £6,000 for a feasibility sprint to £120,000 for a multi-agent system. The number most quotes leave out is the running cost, which scales with every task the agent performs — and that is the one that kills projects after launch.
Here is the full arithmetic, including the parts suppliers prefer not to itemise.
Build costs
| Scope | What it includes | Typical range |
|---|---|---|
| Feasibility sprint | 2–3 weeks on your real data. Measured accuracy. A go/no-go you can act on. | £6,000 – £12,000 |
| Single production agent | One defined process, live, monitored, audit trail, human escalation | £18,000 – £45,000 |
| Multi-agent workflow | Several agents, orchestration, integration with existing systems | £45,000 – £120,000 |
| AI operations retainer | Evaluation runs, prompt and policy updates, model-change testing | From £1,500/month |
These assume UK delivery. ITJobsWatch reports a median advertised UK contractor day rate of £525 for software developers in the six months to 1 October 2026, London 15–30% higher. Agentic work typically carries a premium because it needs both software engineering and the evaluation discipline to prove the thing actually works.
The cost nobody quotes: model usage
Every step an agent takes is a paid call to a model. This is an operating cost and it scales with volume.
Why pilots mislead
A pilot processes 50 documents and costs £40. Production processes 4,000 a month. People assume the cost scales linearly to £3,200 — but it is usually worse, because production tasks are messier, trigger more retries and take more steps.
Rough arithmetic
Work it out like this:
Steps per task × Average cost per step × Retry factor × Tasks per monthWorked example — invoice processing:
| Steps per invoice (read, extract, look up PO, compare, decide) | 6 |
| Average cost per step | £0.015 |
| Retry factor (failures, ambiguity, re-reads) | 1.4 |
| Invoices per month | 3,000 |
| Monthly model cost | £378 |
That is manageable. Now change two assumptions: a more capable model at £0.06 per step and a task needing 14 steps because the documents are harder.
| Monthly model cost | £3,528 |
Same process. Nine times the cost. This is why the number has to be modelled before you commit, not discovered in month three.
Ask every supplier: “What is the expected monthly run cost at our volume, in writing?” If they have not calculated it, they have not finished designing the system.
Why quotes vary so widely for the same brief
Three suppliers quoting £18,000, £45,000 and £95,000 for the same description is normal, and the spread is informative rather than random.
They assumed different accuracy targets. Reaching 85% is quick; 97% can cost as much again, because the remaining errors are each unusual in a different way. A cheap quote often assumes the lower figure without saying so. Ask what accuracy is being quoted to.
They assumed different integration depth. “Connects to your accounting system” can mean a nightly export or a live two-way sync with error handling. Order-of-magnitude difference.
Some included governance and some did not. Audit logging, authority boundaries, approval workflows and compliance documentation are real engineering. A quote without them is a quote for something that will be decommissioned after its first incident.
Some included the evaluation suite and some did not. Without one, nobody can tell you whether accuracy has degraded.
Some are quoting a prototype. It will demonstrate well and need rebuilding.
How to make quotes comparable: state the accuracy target, the integration depth, and that governance and an evaluation suite are required deliverables. The spread narrows dramatically once everyone is pricing the same thing.
Everything else in the five-year picture
| Cost | Typical | Notes |
|---|---|---|
| Build | £18k – £45k | One-off, single agent |
| Model usage | £100 – £3,000/month | Scales with volume. The variable one. |
| Hosting and infrastructure | £30 – £400/month | Usually modest |
| Monitoring and logging | £0 – £150/month | Often bundled |
| AI operations retainer | £1,500/month+ | Optional but recommended |
| Model migration | £3k – £15k every 18–24 months | The one nobody warns you about |
Model migration — the hidden recurring cost
Model providers deprecate models. When yours is retired, the system must be moved to a successor, and the successor behaves slightly differently — which means re-testing everything.
Budget £3,000–£15,000 every 18–24 months for this, or absorb it in a retainer.
You reduce it by building against an abstraction layer so the model can be swapped, and by maintaining an evaluation suite so the impact of a change is a measured number rather than a guess. Both are design decisions made at the start. Ask whether your supplier does them.
What drives the build cost
Integration complexity. Connecting to a modern documented API is routine. Connecting to a fifteen-year-old system with no API can cost more than the agent itself. This is the largest single variable.
Data quality. An agent reading clean, consistent records is straightforward. One reading inconsistent records where the same entity appears several ways needs far more handling and far more testing.
Accuracy required. Getting to 85% is quick. Getting from 85% to 97% can take as long again, because the remaining errors are each unusual in a different way. Decide what accuracy you actually need — if a human reviews everything anyway, 85% with good escalation may be worth more than 97% at twice the cost.
Governance depth. Audit logging, authority boundaries, approval workflows and compliance documentation add real engineering. They are also what keeps the system running after an incident, which is the specific failure Gartner expects to cause 40% of enterprises to decommission agents by 2027.
Number of exception types. Each distinct exception needs handling and testing. Ten exception types is a different project from three.
How to spend less
Start with feasibility, not a build. £6,000–£12,000 tells you whether the thing works on your real data. Discovering it does not at feasibility stage is a cheap outcome.
Narrow the scope hard. One process, done properly. Scope expands after something works, never before.
Use the cheapest model that passes. Capability costs money per call. Many tasks run perfectly well on a smaller model. A supplier defaulting to the most expensive model for everything is not optimising for you.
Accept lower accuracy with good escalation. If a human checks the output anyway, paying for the last few percentage points may be pointless.
Automate the deterministic steps. Only the genuinely varying step needs an agent. Wrapping the whole process in an agent is simpler to build and more expensive to run.
Cache aggressively. Repeated identical calls should not be repeated paid calls.
What suppliers leave out of quotes
Six omissions worth asking about directly, because each one lands after you have signed.
The monthly run cost. The most common omission. Ask for the expected figure at your volume, in writing, with the calculation shown.
Model migration. Providers deprecate models every 18–24 months. Moving to a successor means re-testing everything. Budget £3,000–£15,000 per cycle or hold a retainer.
The evaluation suite. If nobody built one, nobody can tell you whether accuracy has degraded. Ask whether it is included and whether you keep it at handover.
Exception handling for the long tail. The first 80% of cases is quick. The last 15% each fail differently. A quote assuming uniform difficulty is a quote that will be revised.
Monitoring and alerting. Someone must notice when the exception rate climbs or the cost per task doubles. This is engineering work, not a dashboard that appears for free.
Your own time. Building the test set, reviewing shadow-run output, deciding edge cases. Realistically several hours a week during the project.
How to get a cheaper build without getting a worse one
Narrow to one process. Not “invoices and purchase orders and delivery notes” — invoices. Scope expands after something works.
Accept a higher escalation rate initially. Getting from 85% to 97% accuracy can cost as much as reaching 85% did. If a human reviews output anyway, launch at 85% and improve with real data.
Automate the deterministic steps separately. Only the step that genuinely requires judgement needs an agent. Wrapping the whole process in one is simpler to build and more expensive to run.
Use the cheapest model that passes your evaluation. Capability costs money per call, and many tasks run perfectly well on a smaller model. A supplier defaulting to the most expensive model for every step is not optimising for you.
Clean the data first. Yourself, in the source system. It reduces build cost, improves accuracy, and has standalone value.
Run feasibility before committing. £6,000–£12,000 to find out whether the thing works at all is the cheapest decision on this page.
Is it worth it?
Run the arithmetic before you commission anything.
Annual cost of the process now:
hours per week × 52 × fully loaded hourly cost
A process consuming 12 hours a week at £22/hour fully loaded costs £13,728 a year.
Against: a £30,000 build plus £500/month running cost = £36,000 in year one, £6,000 a year after.
Payback: roughly 2.5 years on labour alone. Add error reduction, faster turnaround and the capacity freed, and it may be faster — but on labour alone, 2.5 years.
If that is not acceptable, the honest answers are a bigger process, a cheaper build, or not doing it.
A process consuming four hours a week does not justify a five-figure build. That is the calculation most people skip.
Where the build cost actually goes
A £30,000 single-agent build, broken down. Useful for judging whether a quote is shaped sensibly.
| Component | Share | What it is |
|---|---|---|
| Integration | 30–40% | Connecting to your systems, handling their failure modes. Almost always the largest piece. |
| Evaluation and testing | 20–25% | Building the test set, measuring accuracy, tuning until it passes. The part that distinguishes a production agent from a demo. |
| Governance | 15–20% | Audit logging, authority boundaries, escalation paths, approval workflow. |
| The agent itself | 10–15% | Prompts, tool definitions, the loop. The smallest piece, and the one people assume is the whole job. |
| Deployment and monitoring | 10% | Running it reliably, alerting when it misbehaves. |
If a quote is mostly “the agent” with little for integration, evaluation or governance, it is a quote for a demo. Demos do not become production systems by extension — they get rebuilt, and you pay twice.
Worked payback, three real shapes
Invoice processing, 3,000/month
| Current: 4 min/invoice × 3,000 × 12 | 2,400 hours/year |
| At £19/hour fully loaded | £45,600/year |
| Agent handles 78%, rest escalates | saves ~£35,500/year |
| Build £32,000 + £450/month running | £37,400 year one |
| Payback | ~13 months |
Email triage, 250/week
| Current: 6 min/email × 250 × 52 | 1,300 hours/year |
| At £19/hour | £24,700/year |
| Agent drafts 70%, human approves | saves ~£14,800/year |
| Build £26,000 + £280/month | £29,400 year one |
| Payback | ~24 months |
Compliance document checking, 120/month
| Current: 12 min each × 120 × 12 | 288 hours/year |
| At £22/hour | £6,336/year |
| Agent handles 85% | saves ~£5,400/year |
| Build £20,000 | |
| Payback | ~44 months — do not build this |
The third one is the point. The use case is sound, the technology works, and the volume does not justify it. Run this calculation before committing, because enthusiasm is not a substitute for it.
Questions to ask before signing
Six, each one closing a common gap.
“What is the expected monthly run cost at our volume, in writing?” Steps per task, cost per step, retry factor, volume. If it cannot be produced, the design is not finished.
“What model will this use, and have you tested a cheaper one?” Many tasks run well on smaller models at a fraction of the per-call cost. Defaulting to the most capable model for every step is not optimising for you.
“What are the step and spend caps?” They should be in the code, not in a policy document.
“Do we keep the evaluation suite?” The test set and measurement harness are the most durable output of the project and should transfer to you.
“What happens when the model is deprecated, and who pays?” Providers retire models on their own schedule. Establish whether migration is covered by a retainer or billed separately.
“Which parts of this are agent and which are ordinary code?” The deterministic steps should be code. Wrapping the whole process in an agent is simpler to build and more expensive to run for ever.
Frequently asked questions
How much does it cost to build an AI agent?
A feasibility sprint runs £6,000–£12,000, a single production agent £18,000–£45,000, and a multi-agent workflow £45,000–£120,000. Model usage is a separate monthly cost that scales with volume.
What are the ongoing costs of an AI agent?
Model usage (£100–£3,000+/month depending on volume and model), hosting (£30–£400/month), optional AI operations retainer (from £1,500/month), and model migration every 18–24 months at £3,000–£15,000.
Why is agentic AI more expensive than normal automation?
It needs both software engineering and evaluation discipline — proving accuracy on real data, building governance and escalation, and maintaining a test suite as models change. It also has a per-task running cost that automation does not.
How do I estimate model usage costs?
Steps per task × cost per step × retry factor × tasks per month. Get your supplier to show this calculation with your real volumes before you commit.
Can I use a cheaper model?
Often yes. Many tasks run well on smaller, cheaper models. The right approach is to test the cheapest model first and only move up when accuracy requires it.
What is the payback period?
Calculate the annual labour cost of the process and compare it with build plus running costs. Two to three years on labour alone is typical for a mid-size process. Faster if error reduction or capacity has real value to you.
What happens when the model gets deprecated?
The system moves to a successor and everything is re-tested, because behaviour differs. Budget £3,000–£15,000 every 18–24 months, or hold a retainer. Building against an abstraction layer and keeping an evaluation suite reduces this considerably.
Written by Jas Singh, Founder. Last updated 8 October 2026. Prices reviewed quarterly.
Day rate figures from ITJobsWatch UK contract data for the six months to 1 October 2026. Model usage figures are worked examples using indicative per-call costs — actual costs depend on provider, model and task.
Related:
Agentic AI systems · Where agentic AI fails · Our full pricing
