What separates an agent from a chatbot
A chatbot receives a message and returns text. An agent receives a goal and works towards it: it plans a sequence of steps, calls tools and APIs, reads the results, decides whether it succeeded, and retries or escalates when it didn't. That autonomy is the entire value proposition and the entire cost problem.
The moment software can take actions in your systems rather than just describe them, you inherit a set of requirements that chatbot projects never face — permissions, identity, approval gates, rollback, audit logs, and a way to answer "why did it do that?" three weeks later. Teams that budget for an agent as though it were a chatbot with extra steps typically discover this at around week six.
Quick cost ranges
Based on what UK projects actually price at in 2026, there are three broad tiers. A single-workflow agent — one bounded task, one or two systems, clear success criteria — costs £10,000–£30,000. A multi-step agent with several integrations, human review, and error handling costs £30,000–£80,000. An enterprise multi-agent system with orchestration, governance, and full audit trails runs £80,000–£200,000 and upwards.
| Agent type | Typical build cost | Timeline |
|---|---|---|
| Single-workflow agent (invoice triage, lead qualification) | £10,000 – £30,000 | 5–9 weeks |
| Multi-step agent with integrations and human review | £30,000 – £80,000 | 10–18 weeks |
| Enterprise multi-agent system with governance | £80,000 – £200,000+ | 5–9 months |
| Readiness assessment / feasibility spike only | £3,500 – £8,000 | 2–3 weeks |
Running costs for a mid-sized SME deployment typically land between £150 and £1,600 per month in model and infrastructure spend, before observability tooling and human oversight time. That range is wide because agent economics depend on task volume and task complexity, not on headcount.
Why token costs behave differently for agents
This is the single most common budgeting mistake, so it's worth being precise about. A chat assistant handles one turn: one prompt in, one answer out. An agent completing a single task may run ten or more model calls — plan, call a tool, read the response, re-plan, call another tool, verify, summarise. Industry measurements put a single agentic workflow execution somewhere in the range of 15,000–80,000 tokens, an order of magnitude above a comparable chat interaction.
The practical consequence: stop pricing in tokens and start pricing per successfully completed task. Take your expected monthly task volume, measure the real token cost of ten representative tasks end to end, and add a failure multiplier — because failed and retried attempts consume tokens too, and early-stage agents fail a lot. A workflow that looks like pennies per call can be a pound or more per completed task once retries, verification steps, and long tool outputs are counted.
Model routing is the main lever. Most agent steps — classification, extraction, formatting, deciding which tool to call — do not need your most capable model. Routing the routine steps to a cheaper model and reserving the frontier model for genuine reasoning typically cuts running costs substantially without a measurable drop in task success rate.
The costs that don't appear in the quote
Build cost is the visible number. For agents, a large share of total cost of ownership sits in the operational layer that makes the agent trustworthy enough to leave running.
| Component | Typical cost | Why it's needed |
|---|---|---|
| Observability and tracing stack | £6,000 – £16,000 setup | Without step-level traces you cannot debug why a task failed |
| Human-in-the-loop review interface | £9,000 – £20,000 | Someone has to approve, correct, and unblock edge cases |
| Evaluation harness and regression tests | £5,000 – £15,000 | Detects silent quality drift when prompts, models, or data change |
| Audit trail for regulated use | £12,000 – £35,000 | Required where decisions affect customers, staff, or money |
None of this is optional in the way it sounds. An agent without traces is a system nobody can debug; an agent without an evaluation harness degrades quietly every time a model version or an upstream API changes. These are cheap to build in at design time and painful to retrofit after the first incident.
Why most agent pilots never reach production
The industry data here is sobering and consistent. Gartner has warned that over 40% of agentic AI projects could be cancelled by 2027 on the grounds of escalating cost, unclear value, and inadequate governance. Broader analysis of enterprise pilots puts the share reaching production scale in the low teens — the large majority stall somewhere between a working demo and a system anyone will actually depend on.
In our experience the failures are rarely about model capability. They are about the surrounding engineering:
What actually kills agent projects:
Tool calling and the MCP question
Model Context Protocol has become the default way to connect agents to systems, and adoption in 2026 has been rapid. It is a genuine improvement over hand-rolled integration glue, but it moves risk rather than removing it. Two things deserve attention before you standardise on it.
First, tool descriptions are prompt input. A malicious or compromised tool definition can influence what your agent does — the class of problem usually described as tool poisoning. Treat third-party MCP servers with the same scrutiny you'd apply to any dependency with production credentials, and prefer servers you host or have reviewed.
Second, connecting more tools makes agents worse before it makes them better. Every additional tool expands the selection space the model has to reason over. Agents with six well-described tools reliably outperform agents with thirty vaguely-described ones. Add tools deliberately, and measure task success rate before and after each addition.
Where UK businesses are actually getting returns
Adoption across UK SMEs sits at roughly a third of businesses, with cost and regulatory uncertainty the two most-cited barriers. The organisations seeing real returns have generally picked narrow, high-volume, tolerant-of-review workflows rather than showpiece autonomy.
The pattern that works: a task performed dozens or hundreds of times a week, with a clear correct outcome, where a wrong answer is caught cheaply. Invoice and purchase-order matching. Support ticket triage and routing. Lead enrichment and qualification. Document chasing and compliance checklist completion. Internal data lookups that currently cost someone twenty minutes of clicking between systems.
The pattern that doesn't: anything where the agent makes an irreversible decision with no review step, anything where the correct answer is genuinely contested, and anything where the underlying process is undocumented. An agent cannot automate a process that no human can describe.
How to scope a first agent sensibly
Pick one workflow. Instrument it so you know the current cost per completion in staff time. Build the agent with a human approval gate on every action from day one, then remove gates selectively as measured accuracy earns it. Define success as a percentage of tasks completed without human correction, measured weekly, and agree the threshold before you start.
That approach keeps the first project inside the £10,000–£30,000 band, produces a number you can take to a board, and surfaces the data and permissions problems you'd otherwise hit at ten times the cost on a larger build.
AyTech note: The agents that survive contact with real users are the boring ones — narrow scope, heavy instrumentation, an obvious human owner. We'd rather ship one workflow that runs unattended for six months than five that need supervising forever.
Thinking about your first AI agent?
AyTech can assess which of your workflows are genuinely agent-shaped, price the build honestly, and tell you where automation won't pay back.
AI integration services