Everyone wants "AI" in their product, and almost nobody knows what it should cost. Quotes for AI development in the UK range from a few thousand pounds to seven figures, and the spread isn't just agency mark-up — "AI project" covers everything from wiring up an existing chatbot API to training a bespoke machine-learning model on proprietary data. This guide breaks down honest UK cost bands by project type, explains what really drives the price (spoiler: it's usually your data, not the model), and walks through the build-vs-buy maths that stops most overspending before it starts. As ever, these are broad market ranges, not quotes.
UK AI development cost bands by project type
AI chatbot (customer-facing): £8,000–£40,000
A chatbot built on existing large language models (via APIs), grounded in your own content — FAQs, support deflection, lead qualification, internal knowledge assistants. The lower end covers a well-configured bot on your docs; the upper end covers custom integrations (CRM, booking systems), brand-tuned behaviour, analytics and proper testing. Most business chatbot projects land £15,000–£30,000.
Document intelligence / RAG systems: £20,000–£80,000
Retrieval-augmented generation: AI that searches your documents, contracts, manuals or knowledge base and answers with citations. Think "ask our 10,000-page policy library anything." Cost scales with document complexity, source-system integrations, access-control requirements (who's allowed to see what), and accuracy testing. Regulated industries sit at the higher end because the guardrails are the product.
AI agents and workflow automation: £15,000–£100,000
AI that doesn't just answer but does: triaging inbound enquiries, drafting and routing work, enriching CRM records, multi-step processes with human checkpoints. Simpler single-workflow automations start around £15,000–£30,000; multi-system agents with complex logic, error handling and audit trails climb toward £100,000. The integration surface — how many systems the agent touches — is the main cost lever.
Custom machine-learning models: £60,000–£200,000+
Bespoke models trained on your data: demand forecasting, churn prediction, fraud detection, image recognition, recommendation engines. This is data-science work — data pipelines, feature engineering, training, validation and deployment — and it's priced like the specialist discipline it is. Fine-tuning an existing model on your data sits at the lower end; training novel architectures from scratch is a different budget entirely.
Ongoing costs: the subscription nobody mentions
AI projects have running costs that traditional software doesn't: inference/API fees that scale with usage (pennies per query add up at volume — model your unit economics), data pipeline maintenance, model monitoring and retraining as data drifts, and evaluation to catch degrading quality. Budget 20–30% of build cost annually as a starting point, more for high-volume consumer-facing systems. An AI feature that's cheap to build and ruinous to run is a failure of arithmetic, not technology.
What actually drives AI development cost
1. Your data (the big one)
The model is rarely the hard part anymore — your data is. AI projects live or die on data quality, and most organisations discover theirs is messier than assumed: scattered across systems, inconsistently labelled, full of gaps, or locked in PDFs nobody can query. Data cleaning, labelling, pipeline-building and governance routinely consume 40–60% of a custom AI project's effort. If a vendor quotes you without asking hard questions about your data, they're guessing.
2. Accuracy requirements
A chatbot suggesting blog topics can be wrong sometimes; a system summarising contracts or triaging medical enquiries cannot. Every "nine" of reliability costs real money — in better data, retrieval architecture, testing, human-in-the-loop design and monitoring. Be honest about the cost of being wrong in your use case, and engineer (and budget) accordingly.
3. Integration surface
An AI demo in isolation is cheap. An AI wired into your CRM, ticketing, telephony, document stores and reporting — with single sign-on, permissions respected and audit logs — is a systems-integration project that happens to include AI. Count the systems, multiply accordingly.
4. Compliance and governance
Regulated sectors (finance, insurance, health, legal) need guardrails, logging, bias testing, human oversight and documentation — see our guide to AI in insurance marketing for how this plays out under FCA scrutiny. The EU AI Act also has UK-relevant implications for companies operating across Europe. Governance isn't overhead; for high-stakes AI it's the product.
Build vs buy: the decision that saves (or wastes) the most money
Before building anything custom, ask what's already for sale. The honest framework:
- Buy (SaaS/API) when the capability is commoditised: general chatbots, transcription, standard document Q&A, off-the-shelf automation. Faster, cheaper, maintained by someone else. Your differentiation isn't the chatbot — it's what the chatbot knows about your business.
- Build on APIs when you need custom behaviour on commodity models: grounded assistants, domain-specific agents, branded experiences. This is the sweet spot for most UK businesses — custom value without custom model costs.
- Fine-tune when you need consistent domain tone/behaviour and have quality training data. Moderate cost, real differentiation.
- Build custom models only when your data is the moat and no existing model does the job — forecasting on proprietary signals, specialised vision tasks, novel research. Expensive, slow, and justified only by genuine competitive advantage.
Most "we need to build our own AI" conversations end at option two once the maths is done. Building what you could buy is the most common way AI budgets die. For the general principles behind this kind of decision, our custom vs off-the-shelf software comparison covers the same logic beyond AI.
How to keep AI costs under control
- Start with a pilot on one workflow — prove value and learn your true data costs before committing to a platform.
- Model the running costs per query/user/month before signing off the build. If the unit economics don't work at scale, redesign now.
- Fix the data first — or at least price the data work honestly. It's the iceberg under every AI quote.
- Demand evaluation — accuracy benchmarks, adversarial testing, ongoing monitoring. "It seemed fine in the demo" is not QA.
- Plan the human oversight — who's accountable when the AI is wrong, and what the escalation path looks like.
If you're weighing an AI project and want an honest read on which tier it sits in — and whether you should build, buy or do nothing — our AI automation and our application development team team does this assessment regularly for UK businesses. Get a free consultation; we'll tell you if an off-the-shelf tool does the job, because we'd rather earn your trust than sell you a build you don't need.
FAQs
How much does it cost to build an AI chatbot in the UK?
Typically £8,000–£40,000 for a business chatbot built on existing language models and grounded in your content. Simple FAQ bots sit at the lower end; bots integrated with your CRM, booking or support systems with proper testing sit higher. API running costs are on top.
Is it cheaper to build AI in-house or buy off the shelf?
Buying is cheaper for commoditised capabilities (chat, transcription, standard automation). Building on top of AI APIs suits custom behaviour on your data. Fully custom models are only justified when proprietary data creates genuine competitive advantage. Most businesses overspend by building what they could have bought.
What are the ongoing costs of an AI system?
API/inference fees scaling with usage, data pipeline maintenance, model monitoring and retraining, and continuous evaluation. Budget roughly 20–30% of build cost per year as a starting point — more for high-volume systems. Always model per-query unit economics before building.
Why do AI projects go over budget?
Usually the data: cleaning, labelling and piping messy real-world data consumes far more effort than expected — often 40–60% of the project. Other culprits include underestimated accuracy requirements, integration sprawl, and running costs nobody modelled.
How long does an AI development project take?
A focused chatbot or automation pilot: 4–10 weeks. Document-intelligence systems: 2–5 months. Custom ML models: 4–9+ months including data work. Data readiness is the biggest schedule variable — audit it before committing to dates.
