How much does a custom AI agent cost in the UK in 2026?

TL;DR A production AI agent built by a UK consultancy in 2026 typically costs between £6,000 and £30,000 depending on type and integration complexity. Voice agents start around £6,000 setup. Multi-step agentic systems land between £10,000 and £28,000. Expect 4–10 weeks delivery and a flat monthly retainer of £400–£1,800/mo for monitoring and updates.

Key takeaways

What drives AI agent cost in 2026

AI agent cost in the UK in 2026 is driven by three things: how many tools the agent has to call, how much new UI you need around it, and whether voice is involved. The model itself is rarely the dominant cost.

A simple agent answering questions from a knowledge base sits at the low end. An agent that integrates with Salesforce, Stripe, your support inbox and a custom database — and can take actions in all of them — sits at the high end.

UK pricing benchmarks by agent type

Agent typeSetupMonthlyTimeline
AI Voice Agent£6,000–£18,000£400–£1,5003–6 weeks
AI SDR (productised)£4,500£1,800–£4,5004–6 weeks setup
AI Customer Support Agent£7,000–£20,000Usage-based4–8 weeks
AI Copilot (per dept)£8,000–£22,000Optional retainer5–10 weeks
Agentic AI System£10,000–£28,000Optional retainer6–10 weeks
Custom AI-Powered CRM£12,000–£32,000Optional retainer6–12 weeks

Hidden costs to plan for

  • Model usage (token cost) — typically £30–£500/mo for SMB workloads on Claude or GPT-5.5
  • Voice provider fees (Vapi, Retell, Twilio) — £0.10–£0.30 per minute of conversation
  • Vector database hosting (Pinecone, Weaviate) — £30–£200/mo at SMB scale
  • Monitoring and evaluation tooling (Langfuse, Sentry) — £0–£100/mo
  • Domain rotation and inbox warming for AI SDR — £100–£300/mo per seat

What you should not pay for

Avoid consultancies that quote 'AI strategy' deliverables ahead of any working software. In 2026 the cost of building a working prototype has collapsed — a competent consultancy will ship a working agent inside 2 weeks rather than spending 4 weeks on a slide deck.

Avoid platform lock-in. If the proposal includes 'our proprietary AI platform' as a per-seat licence, walk away. The market standard is open: MCP for tool calling, your choice of frontier model, source code in your private repo.

The costs quotes leave out

Build price is the number everyone compares. It is rarely the number that decides whether the project pays back.

CostTypical shapeWhy it gets missed
Token spendScales with usage, not with buildEasy to model badly. A support agent at 10k conversations/month costs an order of magnitude more than a pilot at 500.
Integration workOften 30-50% of total buildYour CRM's API being awkward is not visible at quoting stage.
Data cleanupHighly variable, occasionally the largest lineThe agent is only as good as what it reads. Messy knowledge bases surface late.
Evaluation harnessOngoingWithout it you cannot tell whether a model or prompt change made things worse.
Human oversightOngoing, decreasingReal for the first months and rarely budgeted.
Model migrationEvery 6-12 monthsThe lineup moves. Re-testing against a new model is routine maintenance, not a surprise.

Why token cost is the line that surprises people

Frontier-tier models sit around $5 input and $25 output per million tokens, and the most capable tier reaches $50 output. Workhorse models run several times cheaper, and open-weight options can be two orders of magnitude below frontier output pricing.

The consequence is that architecture decisions dominate running cost far more than vendor choice. An agent that sends the full conversation history on every turn will cost several times one that summarises, and a system that routes routine work down a tier typically cuts spend by more than any negotiated discount would.

Model a realistic monthly volume before signing anything. A pilot's token bill tells you almost nothing about production.

What actually drives the quote up

  • Number of systems to integrate — each one adds auth, error handling, rate limits and a failure mode
  • How clean your data is — the single most common cause of an over-running project
  • Whether the workflow is genuinely defined, or is being defined during the build
  • Compliance requirements — audit logging, retention and data residency are real engineering
  • How much autonomy the agent needs — every step it takes unsupervised needs a guardrail and a rollback
  • Whether anyone internally owns it after handover

Frequently asked

Why is there such a wide price range?

The range reflects integration scope. A voice agent that books meetings into one calendar costs £6,000. A voice agent that also updates your CRM, qualifies against custom rules and triggers a finance workflow costs £18,000. Get a fixed-scope quote — if the consultancy can't fix scope, they don't understand your problem yet.

Are monthly retainers necessary?

For voice agents, AI SDRs and customer support agents — yes, because there's ongoing operation cost (telephony, prospect data, model usage). For one-off agentic systems and copilots — optional. Most clients take a £400–£800/mo retainer for monitoring, evaluation and model updates rather than building ML-ops in-house.

Can I get something built for under £5,000?

Realistically, no — not for a production system you can rely on. Under £5,000 you can buy seat licences for off-the-shelf tools (Tidio, Intercom Fin, ChatGPT Enterprise) which is often the right answer for very small teams. Custom development starts to make sense from £6,000.

What's typical project payback?

Most clients recover the project cost within 90 days through saved labour, recovered missed-call revenue or additional booked meetings. Voice agents and AI SDRs have the fastest payback — often 30–60 days because the saved revenue is so direct.

Why do AI agent quotes vary so much?

Mostly integration surface and data quality, not model choice. Two agents doing superficially similar work can differ several times over in cost if one reads a clean documented API and the other has to reconcile three systems that disagree about what a customer is.

What will it cost to run per month?

Token spend scales with usage, so model a realistic production volume rather than extrapolating from a pilot. Architecture matters more than vendor: summarising conversation history instead of resending it, and routing routine work to a cheaper tier, typically save more than any discount you could negotiate.

Is there an ongoing cost after the build?

Yes — token spend, monitoring, and periodic re-testing when the model lineup moves, which currently happens every few months. Budget for maintenance rather than treating the build price as the total.