RAG Chatbot Cost: From $6k Pilots to Enterprise
A chatbot that answers from your own documents is the most common first AI project we build. The price range is wide because "answer from our docs" hides very different amounts of work. Here is how it splits.
- Published
Cost by tier
| Tier | What you get | Timeline | Typical cost |
|---|---|---|---|
| Pilot | One clean document set, web widget, basic evals | 2–3 weeks | $6k – $12k |
| Production assistant | Several sources, citations, feedback loop, analytics | 3–5 weeks | $12k – $25k |
| Enterprise knowledge | Permissions per user, many connectors, audit, SSO | 6–10 weeks | $30k – $70k |
What moves the price
- Data mess. Clean PDFs and a help centre are cheap. Scanned documents, spreadsheets with meaning in the layout, and wikis full of outdated pages need cleaning and parsing work.
- Accuracy bar. An internal assistant can be right 90% of the time with citations. A customer-facing bot for regulated products needs a higher bar and a much larger eval set.
- Access control. If different users may see different documents, retrieval must respect permissions — this is often the single biggest cost jump.
- Freshness. A nightly re-index is simple; real-time sync with a CRM or ticketing system is not.
Running cost
Running cost has three parts: the LLM per question, embeddings when documents change, and the vector database. For most business assistants this is modest compared with the build, and it scales with usage — a quiet internal bot costs very little; a public bot on a busy site needs caching and a cheaper model for simple questions.
The expensive mistake
The costliest mistake is skipping evaluation. Without a set of real questions and expected answers, nobody can tell whether a change made the bot better or worse, and teams burn weeks tweaking prompts by feel. We build the eval set in week one, before the first prompt.
Frequently asked questions
Can we start with the pilot and grow?
Yes — that is what we recommend. The pilot’s ingestion pipeline and eval set carry straight into the production tier.
Does the data leave our control?
We deploy in your cloud account where required, and use model providers with no-training-on-your-data terms.
How accurate will it be?
We agree a target on your own question set before building, and report against it every week.
