Quick cost ranges
A small internal knowledge assistant usually costs £12,000-£25,000. A customer-facing support assistant with citations and escalation paths costs £25,000-£60,000. A regulated or enterprise RAG system with role-based access, audit logs, and multiple data sources often costs £60,000-£150,000+.
The biggest variable is not the LLM. It is the state of the source knowledge. Clean, current documentation is cheap to index. Scattered PDFs, SharePoint folders, support tickets, CRM notes, and old policy documents need discovery, cleanup, chunking strategy, metadata design, and ongoing governance.
| System type | Typical cost | Timeline |
|---|---|---|
| Internal knowledge assistant (staff-only) | £12,000 – £25,000 | 6–10 weeks |
| Customer-facing support assistant with citations | £25,000 – £60,000 | 10–18 weeks |
| Regulated / enterprise, multi-source, audit logged | £60,000 – £150,000+ | 4–6 months |
RAG vs fine-tuning vs long-context: which one you actually need
These three approaches solve different problems and get conflated constantly. RAG is the right choice when the knowledge changes often and answers need to be traceable to a source document — policies, product docs, contracts, support history. Fine-tuning is the right choice when you need the model to consistently adopt a style, format, or specialised skill rather than recall specific facts, and it's a poor fit for knowledge that changes weekly since retraining is slow and expensive. Long-context (feeding entire documents directly into a large context window) can work for small, stable document sets, but costs scale with every query since the full context gets reprocessed each time, and it doesn't scale to a knowledge base of thousands of documents.
Most production systems that look like "fine-tuning" from the outside are actually RAG with a lightly fine-tuned response style layered on top. Start with RAG unless there's a specific reason the knowledge is static and small enough for a different approach.
What a production RAG system includes
A serious build normally includes source discovery, document parsing, embedding generation, vector database setup, retrieval tuning, prompt design, answer citation, safety instructions, user feedback capture, evaluation tests, analytics, and deployment. If the system answers different users differently, it also needs permissions mapped from your identity provider.
For many businesses, the evaluation framework is the difference between a demo and a production tool. You need test questions, expected answers, confidence thresholds, and monitoring for stale or wrong answers. Without that, the system can silently get worse as documents change.
Main cost drivers
Data quality is the first driver. If the source documents are inconsistent or duplicated, budget for a cleanup phase. Integration depth is the second. Connecting one document store is straightforward; joining SharePoint, Google Drive, Notion, Zendesk, HubSpot, and a custom database takes more engineering. Security is the third. A RAG system that exposes private HR, finance, or client documents needs strict access control and logging.
Usage volume matters after launch. LLM and embedding costs can be low for internal tools, but high-volume customer support assistants need caching, cheaper model routing, and observability to keep costs controlled.
When RAG is worth it
RAG is strongest when the business has a large and changing knowledge base: policies, contracts, manuals, product documentation, technical runbooks, customer support history, or compliance material. It is less useful when the answer space is tiny, stable, or better handled by deterministic rules.
A good first project is usually an internal assistant for staff. It has a lower brand-risk profile than a customer-facing bot, gives fast feedback from real users, and exposes the data gaps you need to fix before external release.
Common RAG failure modes
What actually goes wrong in production RAG systems:
Most of these are cheap to prevent at design time and expensive to retrofit once users have learned not to trust the answers. Build the evaluation and re-indexing pipeline as a first-class part of the project, not a follow-up task.
AyTech note: The safest projects start with a narrow, measurable workflow, then expand after real users prove the value. This keeps budgets controlled and gives Google, buyers, and stakeholders clearer proof of expertise.
Need a practical technical plan?
AyTech can review your requirements, map the risks, and turn the idea into a scoped delivery plan.
AI integration services