A-AI // Answer

How much does an AI agency in the UK cost?

UK AI engagements in 2026 fall into four price brackets (all + VAT): discovery + audit £3k to £15k, proof-of-concept LLM apps £15k to £60k, production AI features inside existing software £40k to £200k, and ongoing AI engineering retainers £8k to £40k per month. Nuvenar quotes in writing per scope, typical project bracket £25k to £150k + VAT, retainer bracket £8k to £25k per month + VAT for AI-only.

Published 2026-05-31 // Updated 2026-05-31 // By NUVENAR LTD

The short answer

UK AI engagements in 2026 fall into four pricing brackets. Each bracket corresponds to a different scope and a different buyer intent. Knowing which one you are buying saves the discovery call.

The four pricing brackets

BracketTypical priceWhat you getRight for
Discovery + audit£3k to £15k + VATWritten assessment of where AI fits in your business, model-cost projections, build-or-buy recommendation, prioritised roadmap.Operators who suspect AI helps but want a senior view before committing.
Proof of concept£15k to £60k + VATA working prototype: chatbot, document Q&A, internal assistant, lead scoring agent. Usually 4 to 8 weeks. Deployed in staging, not yet wired into production.Teams that want to prove ROI before a full build.
Production AI features£40k to £200k + VATAI inside your existing software: RAG over your CRM, an AI agent that takes real actions, fine-tuned models, evals, observability, admin dashboard. 8 to 20 weeks.SaaS scaleups productising AI, mid-market businesses automating real workflows.
Ongoing AI retainer£8k to £40k / month + VATContinuous prompt engineering, eval maintenance, model upgrades, incident response, new feature shipping.Operators who already have an AI feature live and need it to keep improving.

What drives the price

Model layer

Calling GPT-5 or Claude Opus via API is cheap to prototype and expensive to operate at scale. A 10k-request-a-day customer support agent on Claude Sonnet 4.6 might cost £1.5k a month in inference. Fine-tuning a smaller model on your data is more expensive upfront (£20k to £80k) but creates an asset with predictable unit economics. Self-hosted models are cheaper at huge scale (100k+ requests a day) and otherwise rarely worth the operational cost.

Integration surface

Hooking AI into a single form is small. Hooking it into a CRM with multi-tenant isolation, audit logging, webhook signing, OAuth, and a permissions model is large. As a rule of thumb: integration is 70 percent of the cost, prompt engineering is 20 percent, and the AI model itself is 10 percent.

Safety surface

A customer-facing AI in a regulated industry (finance, health, legal) needs evals, content filters, refusal policies, audit trails, and approval workflows. These triple the build cost compared to an internal-tool use case. If your AI talks to patients, customers, or end consumers, budget at the top of the bracket.

Model upgrades

Anthropic and OpenAI ship new models every 4 to 12 months. Every upgrade requires a re-eval, prompt tuning, and sometimes regression testing against your existing prompts. A retainer budget that does not include upgrade cycles will go over by 20 to 40 percent in year one.

How Nuvenar prices AI engagements

  • Fixed-scope, fixed-price: every engagement starts with a written Statement of Work and ends with a fixed-price quote. No hourly invoices.
  • Typical project bracket: £25k to £150k + VAT for AI-only scope. Larger if combined with software development and DevOps.
  • Typical retainer bracket: £8k to £25k a month + VAT for AI-only ongoing work. Larger if the retainer also covers application development.
  • API costs pass through at cost. If we manage the OpenAI or Anthropic account on your behalf, you see the receipts and there is no markup.
  • Senior-only team: every engineer who touches your AI work is senior. No junior layer behind a project manager.
  • Production from day one: Nuvenar runs NuvenarHub in production, which means the patterns applied to your build are battle-tested (evals, observability, fallback strategies, prompt versioning, model routing).

How to estimate your project

  1. Write down the user (who interacts with the AI).
  2. Write down the action (what the AI does).
  3. Write down the inputs (what data it has access to) and outputs (what it produces).
  4. Decide if it is internal or customer-facing.
  5. Decide if the AI takes real actions or just suggests.
  6. Decide if you need fine-tuning or off-the-shelf models are fine.

Each of those decisions moves the price within the bracket. The first conversation Nuvenar has with a new AI buyer is exactly this list, written down together in 45 minutes. Book a discovery call and we will scope in real time.

Free tier: NuvenarHub as a starting point

If your AI use case is a WhatsApp auto-reply, AI assistant over your CRM data, call summarisation, or lead scoring, NuvenarHub Pro already includes those at £480 / mo + VAT with no bespoke engineering required. Try the 7-day Pro trial before scoping a custom build; in 30 percent of cases the bundled AI does the job and the bespoke project is not needed.