AI integration feels expensive because the discourse is all about training frontier models. The reality for most products is much cheaper — and much more boring. Integrating AI into your product in 2026 is mostly engineering around hosted models. The pricing reflects that. Here's the honest UK breakdown in three tiers, what moves the number, and the running costs you'll only see in invoices.
For context, we've shipped AI integrations across 11+ products, including Lexa — Pakistan's first AI legal chatbot — and the senior team behind My Friend, an AI companion app. These numbers reflect what we actually deliver, not what an agency would quote you for the same brief.
Starter — from £1,500
Who it's for: product teams adding their first AI feature. Drafting, summarisation, classification, semantic search over a small dataset — well-bounded jobs where AI removes obvious friction.
What you get: an LLM-backed feature integrated into your existing product. One use case, one input/output shape. Hosted model (GPT, Claude, or Gemini — your pick), prompts tuned and tested, basic input/output validation, sensible fallback behaviour, and a small cost-control layer so the inference bill can't surprise you. Built into your repository and your stack, deployed alongside your existing application.
Timeline: 1–2 weeks.
What it isn't: retrieval-augmented generation (RAG) over your own data, custom evaluation pipelines, agentic workflows, or any kind of model fine-tuning. Those move you into the next tier.
Standard — £5,000 to £15,000
Who it's for: teams shipping AI as a real product feature, not a checkbox. This is where most of our integration work lives.
What you get: a production AI feature grounded in your own data via RAG — embeddings, vector store, retrieval strategy, citations. Real guardrails: input validation, prompt-injection defences, structured outputs, safe-refusal behaviour for high-stakes questions. An evaluation pipeline that tracks accuracy on representative inputs across every change, so quality is a number you can watch instead of a vibe. Cost monitoring, observability, and a small ops surface so your team can see what users are asking and how the system is responding.
If you want a deeper picture of what this layer looks like in production, we wrote why most AI features fail in production — this tier is exactly the engineering that closes that gap.
Timeline: 6–8 weeks.
Advanced — from £20,000
Who it's for: organisations building multi-step AI agents, agentic workflows, or AI features that involve fine-tuning, custom models, or specialised retrieval at scale.
What you get: multi-step agents that can use tools and APIs, agentic AI orchestration across multiple workflows, advanced RAG (hybrid search, rerankers, query routing), light fine-tuning where prompting genuinely can't deliver the behaviour you need, deeper observability and evaluation, and the production architecture to run all of it sustainably. This is also where serious regulated workloads land — finance, healthcare, legal — because the compliance and audit work moves the bracket.
Timeline: 10–16+ weeks.
Ongoing / monthly AI ops
AI features need monthly attention. Models change, prompts drift, content updates, edge cases surface. Expect:
- From £400/month: essentials — dependency upgrades, monitoring, small prompt tweaks, monthly eval run.
- £800–£1,500/month: production cadence — full eval cadence, retrieval-corpus refresh, model migrations, small feature work, prompt iteration.
- £1,500–£2,000+/month: active AI team — ongoing agent and pipeline development, multi-model routing, custom fine-tuning maintenance.
Separate from this is inference cost — what you pay your LLM provider per call. For most products this is small (single-digit pounds per thousand interactions on hosted models) but it scales linearly with usage, and long-context retrieval can be expensive on enterprise workloads. We model and cap this at delivery time so the invoice never surprises you.
What actually moves the price
In roughly the order they hit the budget:
- Knowledge surface. A feature grounded in 50 pages of content is far cheaper than one grounded in 50,000 pages. Embeddings, chunking strategy, retrieval evaluation — all scale with the corpus.
- Evaluation rigour. A serious eval pipeline (representative inputs, automated scoring, regression detection) is real engineering work, but it's the difference between a feature you can improve and one you can only hope works.
- Guardrails and safety. Public-facing AI features need real safety work — prompt injection, scope enforcement, safe refusal, escalation paths. Internal-only AI tools can lean lighter on this.
- Number of providers and models. A single-provider integration is cheaper than a multi-provider routing layer with failover. Most products start with one and add the second only when it's justified.
- Agents and tool-use. Multi-step agents (taking actions, calling APIs, routing across tasks) are meaningfully more work than a single-shot LLM call.
The hidden costs nobody quotes you
- LLM inference at scale. Cheap per call, but it scales with usage. Worth modelling at your expected volume.
- Embedding storage and re-embedding. Tiny at the Starter tier, real at scale.
- Eval iteration over time. Quality drifts. A small monthly eval cadence is the cheapest insurance against silent regression.
- Prompt iteration. Prompts aren't fixed. They evolve as the product evolves. Cheap individually, real over a year.
- Vendor changes. OpenAI and Anthropic ship new models and deprecate old ones. Migrations are normal product work.
Why simple AI features can be £1.5k while production is £15k
The difference between Starter and Standard isn't model quality — it's scaffolding. A £1.5k Starter feature wraps a hosted model in good engineering. A £15k Standard feature adds retrieval, evaluation, guardrails, observability, and cost control. Both produce useful AI features. One produces a feature you can improve and depend on; the other is a useful first version that proves the value before you invest in the scaffolding.
We've shipped both, deliberately, depending on the stage and the stakes of the product. The right answer is rarely "spend the maximum" or "spend the minimum" — it's "spend exactly what the use case warrants right now, with a clear upgrade path."
How Lexa was built
Lexa — Pakistan's first AI legal chatbot — landed in the Standard tier. The work that mattered: a real RAG pipeline grounded in legal text so answers cite their basis; safe-refusal behaviour so the system declines when it shouldn't guess; conversation design for non-lawyers under stress; and an evaluation rhythm that catches regressions before users see them. We dug into the technique choices in our RAG vs Fine-Tuning post — for most AI integrations, RAG is the right starting point and fine-tuning is the answer only when prompting and retrieval can't deliver the behaviour you need.
FAQ
Why is there such a wide range in AI integration pricing? Because "AI integration" describes everything from wrapping an OpenAI call (£500) to a full agentic platform (£200k+). Scope, guardrails, evaluation rigour, and knowledge surface decide where you land. A clear use case is the cheapest cost-control tool you have.
Can I really get a working AI feature for £1,500? Yes, at the Starter tier — single use case, hosted model, well-bounded scope. We've delivered exactly this. What you won't get for £1,500 is grounding over your own data, a real eval pipeline, or agents — those are Standard-tier work.
Which LLM should I pick? There's no single right answer — we pick per use case on quality, latency, cost, and data residency. GPT, Claude, Gemini, open-source models all have cases where they're the right answer. We build with a provider-agnostic layer so you're not locked in.
Do I need a dedicated AI team? No — most products don't. The right answer for almost every team is senior engineers who have shipped AI before, integrating hosted models well, with the right guardrails. We wrote a longer take on this in how to add AI without hiring an AI team.
Who owns the code, the prompts, and the eval data? You do. Code in your repos, prompts and pipelines documented, eval data in your accounts. No licensing back to us, no platform lock-in. The engagement ends; the AI system stays yours.
Building an AI feature and want a realistic price for your specific brief? Get a written estimate — honest tier recommendation, written estimate, no sales pressure.
