Answers that point to the source:
not a confident guess
A generic AI assistant answers fluently and sometimes wrongly, which is a problem the moment the answer concerns a policy, a procedure or a number someone will act on. We build a retrieval system that answers only from your actual documents and shows exactly which source it pulled from, so a wrong or outdated document is visible, not hidden inside a confident-sounding answer.
The process today
A team relying on a generic AI assistant for internal questions, what does a policy say, what is in a specific procedure, usually gets a fluent answer that sounds right and sometimes is not, because the model is drawing on its general training rather than your actual current documents. Nobody can tell the difference between a correct answer and a confident-sounding wrong one without going to check the source document anyway, which defeats the point of asking in the first place.
The second cost is documents going stale without anyone noticing. A policy updated last month but never re-indexed anywhere an assistant reads from means the assistant keeps confidently repeating the old version, and the first sign of the problem is usually someone acting on outdated information.
The third is access. A knowledge base that answers from every document regardless of who is asking can leak information a given person should not see, which is a real risk the moment the knowledge base includes anything sensitive, HR policies, contracts, financial data.
What the agent does
We index your actual documents, policies, procedures, product information, internal wikis, into a retrieval system that answers questions by pulling the relevant passage and citing exactly where it came from, rather than generating an answer from general knowledge. If nothing in the indexed documents matches the question well enough, the agent says so instead of guessing, which is the behavior that actually makes the answers trustworthy.
Documents are kept current through an update pipeline, so a changed policy is reflected in the next answer rather than requiring someone to remember to re-upload it. Where different askers should see different information, the agent’s retrieval is scoped by access control, so a question only pulls from documents that asker is allowed to see.
Typical integrations: your document storage, wiki or knowledge management system, and whichever chat interface, Slack, an internal portal, a support widget, your team already uses to ask questions.
What stays with humans
Deciding what counts as an authoritative source document, and keeping that set current, stays with whoever owns the content; the agent retrieves from what it is given, it does not decide what should be treated as policy. Access control tiers, who can ask about what, are set by your team based on your existing permission structure. Any answer that touches a decision with real consequences is expected to be verified against the cited source by the person acting on it, which the citation makes fast to do.
Guards
Every question asked and every source cited is logged, so you can audit what the system has been telling people and trace any wrong answer back to either a bad source document or a retrieval issue. The agent is built to refuse rather than guess when no good match exists, tested specifically for that behavior before launch. Access control is tested against real user roles before go-live, and a kill switch takes the system offline in one message if something looks wrong.
Price and timeline
| Option | Price | What it covers | Timeline |
|---|---|---|---|
| Single automation | from $1,200 | One document set, retrieval and citation, one access channel | 2 to 3 weeks |
| Department package | from $3,500 | Multiple document sets, access control tiers, update pipeline across the organization | 4 to 7 weeks |
Running cost is usually $30 to $120 a month in hosting and model usage depending on document volume and query frequency.
Related
This pairs well with internal knowledge base QA for a simpler FAQ-style front end over the same retrieval system, and with model fine-tuning on company data for tone and style consistency alongside factual accuracy. See the AI agents service page and the automation-everything overview for full package details. For a real build on large-scale indexed retrieval and multi-brand analytics, see the archaeological atlas 1.9M objects case study and the two-brand analytics hub case study.
Tired of confident-sounding answers nobody can verify? Get in touch and we will look at what documents are worth indexing first.
Tired of doing this by hand? We can take the whole routine off your team, not just this step: Routine takeover, from $400 →
FAQ
How much does it cost to build a RAG knowledge base?
From $1,200 for one document set and one access channel, live in 2 to 3 weeks. Larger document sets with access control tiers usually run $2,500 to $4,500.
How long before it is live?
2 to 3 weeks once we have access to your documents and a sense of who should be able to ask what.
What kind of documents work for this?
Policies, procedures, product documentation, internal wikis, contracts, research data, anything structured enough to be split into retrievable passages with a clear source.
What happens if the answer is not in the documents?
The agent says so rather than filling the gap with a generic guess. That refusal is a deliberate design choice, since a wrong answer that sounds confident is worse than an honest 'not found here.'
Is our document data secure?
Documents are indexed and stored in infrastructure you approve, access control limits what each asker's queries can retrieve, and every question and answer is logged for audit.