Does your business need a RAG chatbot? A decision guide, not a sales pitch

Most "why your business needs AI" articles skip the part where AI is the wrong tool. This one does not. If you are deciding whether to build a RAG (retrieval-augmented generation) chatbot — one that answers from your own documents instead of guessing — here is how to tell, and roughly what it costs.

We built one for a real client: Maktabah AI, inside the Hikam Shughro app, answering questions strictly from reference texts with citations. This guide is the conversation we have with prospects before that kind of project — including the times we have said RAG was not the right fit.

Signs you actually need RAG

• Answers live in documents, not in three FAQ entries. SOPs, policy manuals, product catalogues, contracts, reference texts — anything too large or too specific for a human to memorise.
• A wrong answer is expensive. Customer support that misquotes a warranty, HR that misstates a policy, sales that misreads a spec sheet — the cost of being wrong outweighs the cost of building retrieval with citations.
• The source material changes. A RAG system re-indexes updated documents; a hand-written FAQ bot needs a rewrite every time.
• You need to show where an answer came from. Regulated industries, B2B sales, or any team that gets asked "are you sure?" benefits from an answer with a citation attached.

Signs you do not need it yet

We would rather tell you this before the invoice than after:

• Your real question set is small and stable. Under ~30 genuinely distinct questions that rarely change is a simple flow-based FAQ bot or even a well-organised help page — a fraction of the cost.
• Your documents are messy or you do not have them yet. Retrieval quality is capped by source quality. Organising the documents is real work, and doing it first is cheaper than doing it twice.
• You need this in days, not weeks. A grounded RAG assistant needs ingestion, retrieval tuning, and testing against a question set before launch — see how we built one. If the timeline does not allow that, say so up front.

What it costs (market context, not a quote)

We do not publish fixed prices — scope drives cost too much for a number on a page to be honest. But published Indonesian market research gives a useful range: a simple flow/FAQ bot typically starts around Rp5–15 million for a one-off build, while an LLM-based chatbot with a real knowledge base (RAG) commonly runs Rp20 million upward, with more complex, high-volume document sets going higher still. Ongoing model and hosting costs over the first year often exceed the initial build cost, so budget for that, not just the launch.

The actual number depends on document volume, how messy the source material is, and what it needs to integrate with. We give a real range after a short discovery conversation — see our engagement models — not before we understand the documents involved.

The question everyone asks: will it make things up?

This is the right question, and it is answerable by design, not by promise. A RAG system that is built correctly answers only from retrieved passages and declines when nothing relevant is found — "I could not find that in your documents" is a feature, not a bug. We cover exactly how in the Hikam Shughro case study: ingestion, retrieval, grounded answers, and the testing pass that catches it before launch.

How to start

Bring us the documents your team or customers keep asking about — even messy ones. We will tell you honestly whether RAG, a simpler search tool, or plain automation is the right fit, and what a realistic range looks like before you commit to anything. Start the conversation.