Building an AI agent typically costs between £3,000 for a single task bot bolted onto an existing system and £200,000 or more for a multi agent enterprise deployment, with most mid complexity builds landing somewhere between £15,000 and £45,000. Businesses asking this question are usually weighing up a specific use case, not shopping for “AI” in the abstract, so the real answer is less about a single price tag and more about where that use case sits on the scope and autonomy spectrum below.
TL;DR: The final number depends far less on “AI” as a line item and more on scope: how much autonomy the agent needs, how many systems it has to talk to, how complex its retrieval or reasoning layer is, and how much ongoing model usage it will rack up once live. A simple, single purpose agent wired into one existing tool sits at the low end of custom software pricing; a multi agent system that plans, delegates and executes across several business systems sits at the top. On top of the build, expect to budget separately for API usage, hosting and maintenance, which typically run at 15 to 20% of the build cost per year. Below we break down pricing by project type, by region, by build versus run cost, and walk through a worked example so you can budget with real numbers rather than vague reassurances.
Cost by Project Type
AI agents are not a single product category, so a single headline price is close to meaningless. The clearest way to budget is to map agent scope onto Emvigo’s own published web application development tiers, since an agent is, underneath the reasoning layer, a web application with added orchestration, memory and tool calling. These figures come from Emvigo’s web application development cost guide and are general UK software development ranges applied to agent build scope, not an agent specific quote.
| Scope | What it covers | Typical cost | Typical timeline |
|---|---|---|---|
| Simple agent | Single task automation on one data source, one integration, rule-based or lightly autonomous. | £3,000–£12,000 | 3–6 weeks |
| Mid complexity agent | Multi-step reasoning, two to four integrations, RAG over a defined knowledge base, basic memory. | £12,000–£40,000 | 8–16 weeks |
| Multi-agent enterprise system | Several specialised agents coordinating via an orchestrator, deep integration across core systems, governance and audit logging. | £40,000–£200,000+ | 4–9 months |
A single task agent, say one that triages support tickets and drafts first responses, sits comfortably in the simple tier. A mid complexity agent that pulls from a CRM, a knowledge base and an email system, and makes a judgement call on next steps, climbs into the second tier. A multi agent system coordinating sales, support and operations agents against shared context, with proper audit trails for compliance, belongs in the third.
What Actually Drives the Cost
Four variables move an AI agent’s price more than anything else, and none of them is “which model you use.”
Scope and autonomy. A fully scripted agent that follows a fixed decision tree costs far less to build and test than one making genuinely open ended judgement calls. The more autonomy you hand the agent, the more edge cases, guardrails and human in the loop checkpoints have to be engineered, tested and monitored.
Integrations. Each system an agent needs to read from or act on, a CRM, a ticketing tool, an ERP, a payment gateway, adds authentication work, error handling and testing. Three integrations cost meaningfully more than one, not because the code triples but because the failure modes multiply.
RAG complexity. If the agent needs to retrieve information from your own documents or data rather than reason from general knowledge, you are building a retrieval augmented generation pipeline: document ingestion, chunking, embedding, vector storage and relevance tuning. A narrow, well structured knowledge base is cheap to stand up. A sprawling, messy or constantly changing document set is not.
Model and API usage. Choice of underlying model affects both build cost (some models need more prompt engineering and evaluation work to behave reliably) and ongoing run cost, covered below. Heavier reasoning models cost more per call than lighter ones, and multi agent systems that make several model calls per task multiply that spend quickly. Anthropic’s own API pricing page is a useful reference point here, since per token cost varies several times over between its lightest and most capable models, and that gap is exactly what shows up in an agent’s monthly run cost.
It is also worth sizing the broader trend: Gartner forecasts worldwide AI spending to grow 47% in 2026, with agentic AI software spending rising sharply as a share of that total. Agent projects are being funded at scale, which also means the cost conversation, build and run combined, is one more businesses need a clear answer to before committing budget.
Hourly Rates by Region
Agent development is typically priced either as a fixed scope quote or on an hourly basis, and rates vary sharply by where the team sits. The table below uses the same source as the cost tiers above, Emvigo’s web application development cost guide, for consistency.
| Delivery model | Hourly rate |
|---|---|
| UK based agency | £75–£175/hr |
| Offshore development team | £25–£60/hr |
A UK agency commands a premium for time zone overlap, direct access to senior engineers and easier compliance conversations, which matters for agents touching regulated data. Offshore teams can deliver the same scope at a lower hourly rate, though the total project cost gap narrows once you factor in management overhead and the extra testing cycles that distributed teams sometimes need for anything judgement heavy, like an autonomous agent’s edge case behaviour.
For context on what a UK based developer actually costs to employ directly, ITJobsWatch’s software developer salary data puts the median UK salary in the mid fifties of thousands of pounds, before employer National Insurance, pension contributions and overhead, which is part of why an agency’s hourly rate can still work out cheaper than building and retaining an in house AI engineering team for a single project.
Build Cost vs Ongoing Run Cost
These are two separate budgets, and conflating them is the single most common AI agent costing mistake businesses make.
Build cost
This is the one off figure in the tier table above: discovery, prompt and workflow design, integration development, RAG pipeline setup if needed, testing and deployment. It is a project cost, paid once (or across milestones), and it is what most quotes refer to.
Run cost
Once live, an agent keeps costing money every month it runs. Three components make up run cost:
-
- API or model usage. Every call to a model API has a per token cost. A high volume agent making hundreds of calls a day will accumulate meaningful spend even at a few cents per call.
- Hosting and infrastructure. Vector databases, orchestration layers and logging all need somewhere to run, whether that is a managed cloud service or self hosted infrastructure.
- Maintenance. Models drift, source documents change, integrations break when a third party API updates. Emvigo’s software ROI guide puts annual maintenance at 15 to 20% of build cost for custom software generally, and that figure holds reasonably well for AI agents, with the caveat that agents touching fast changing data sources or frequently updated model APIs can sit toward the higher end of that range.
Budgeting for run cost from day one, rather than treating it as an afterthought once the invoice for month one API usage arrives, is what separates agent projects that stay on budget from the ones that quietly become expensive six months in.
A Worked Example
Emvigo’s case study catalogue does not currently include a published AI agent development project with a verifiable cost breakdown, so rather than retrofitting an unrelated case study’s figures onto an agent narrative, here is a clearly labelled illustrative scenario built from the tiers above.
Illustrative scenario: A mid sized UK professional services firm wants an agent to triage incoming client enquiries, check them against existing case records in their CRM, and draft a first response for a human to review and send. Scope: one integration (CRM), a narrow RAG layer over existing case notes, moderate autonomy with a mandatory human approval step before anything is sent. This sits squarely in the mid complexity tier, roughly £18,000 to £28,000 for the build, with ongoing run cost estimated at 15 to 20% of that figure annually, plus API usage that scales with enquiry volume. A firm handling 50 enquiries a day would expect API spend in the low hundreds of pounds per month at current model pricing, though this varies by model choice and prompt length.
Total Cost of Ownership and ROI
The build cost is the number that gets quoted, but the number that determines whether an agent project was worth doing is total cost of ownership set against what it actually saves or earns. Over a two or three year horizon, run cost typically overtakes build cost, which means the ROI case has to hold up on an ongoing basis, not just at launch.
The clearest way to frame ROI is against the task the agent replaces or augments. If an agent handles a volume of first line triage that previously consumed several hours of staff time daily, the maths is straightforward: hours saved multiplied by loaded hourly cost, compared against build cost plus 15 to 20% annual run cost. Agents that augment rather than fully replace a task, catching errors, drafting first passes, surfacing relevant information faster, are harder to quantify precisely but tend to pay back through speed and consistency rather than headcount reduction.
Before committing budget, it is worth running your own numbers through Emvigo’s broader AI development cost guide, which covers budgeting for AI projects beyond agents specifically, and weighing whether an agent is even the right build against a simpler automation, covered in build vs buy for AI.
Conclusion
An AI agent’s price tag is really a reflection of scope, autonomy and how many systems it has to work across, not a fixed “AI premium.” Use the tiers above as a starting point, budget run cost from day one rather than as an afterthought, and get a proper scoping conversation before committing to a number.
Frequently Asked Questions
How much does a simple AI agent cost?
A simple, single task AI agent, such as one that automates a single workflow against one data source, typically costs between £3,000 and £12,000 to build, based on general UK software development pricing applied to agent scope. Timeline is usually three to six weeks.
What is the cheapest way to build an AI agent?
The cheapest route is narrowing scope as far as possible: one integration, one clearly defined task, no custom RAG pipeline if the agent can work from a small, stable document set, and using a lighter weight model for simpler reasoning steps. Offshore delivery teams, typically £25–£60/hr versus £75–£175/hr for a UK based agency, also lower build cost, though the gap narrows for work requiring close compliance or judgement review.
What are the ongoing costs after an AI agent launches?
After launch, expect three ongoing cost categories: API or model usage, which scales with call volume; hosting and infrastructure for any orchestration or vector database layer; and maintenance, generally 15 to 20% of build cost annually, covering model drift, integration updates and data refresh.
Is it better to pay hourly or get a fixed price for AI agent development?
A fixed price works best once scope is well defined, giving budget certainty for a clearly bounded agent. Hourly pricing suits projects where scope is likely to shift as the agent’s behaviour is tested and refined, which is common for anything with meaningful autonomy, since the right level of guardrails often only becomes clear during build.
Does the AI model I choose affect the cost?
Yes. Model choice affects both build cost, since more capable reasoning models can need less prompt engineering for complex tasks but cost more per call, and ongoing run cost, since heavier models are priced higher per token than lighter ones. A multi agent system making several model calls per task will see this cost multiply faster than a single agent handling one call per task.
