A company knowledge chatbot that answers from your own policies, manuals and past work, shows where each answer came from, and respects who is allowed to see what. Built as an internal AI assistant first; customer-facing only once it has earned it.
The problem
The answer exists. It is in a PDF on SharePoint, a Notion page from 2023, an email thread and the head of one experienced colleague. New staff ask the same questions in the team channel; experienced staff answer them for the fortieth time; and occasionally someone follows the old version of the policy because it was the first search result.
A general-purpose chatbot does not help, because it does not know your documents and will confidently make something up.
Retrieval, in plain language
The technique is usually called RAG (retrieval-augmented generation). It works in two steps. First, when someone asks a question, the system searches your approved documents and pulls out the few passages most likely to contain the answer. Second, the AI model writes an answer using only those passages and shows them as sources. The model does not “learn” your documents; it reads the relevant ones each time. That is why answers stay current when a document changes, and why a missing document produces “I could not find this” instead of a guess.
Scope
What we build
Internal policy and process assistant
Sales and proposal knowledge
Support agent assist
Document preparation and upkeep
A question in Teams, answered with sources
A demonstration using sample data. A property management company with 60 staff keeps its procedures in SharePoint and answers most “how do we handle X” questions in a Teams channel.
- A staff member asks the assistant in Teams: “What is the process when a tenant reports a water leak outside office hours?”
- The assistant checks who is asking and which document libraries that person can access. Documents outside their permissions are never searched.
- The relevant passages from the current emergency procedure and the contractor list are retrieved. The superseded 2022 version was excluded during document preparation.
- An answer is written from those passages only: the steps, the on-call number, and a link to each source with its last-reviewed date.
- If retrieval confidence is low or the documents disagree, the assistant says so and offers to notify the operations lead instead of answering.
- Every question, the sources used and a thumbs-up/down are logged. Unanswered questions become a monthly list of documentation gaps.
Credit classification: advanced (four or more credits). Fits Team or Business scope. See also property management.
Permissions, citations and limits
- Permissions are inherited, not copied. The assistant searches only what the asking person could open themselves. HR files do not leak into the sales channel.
- Every answer cites its sources. No source, no answer. The person can check in one click.
- “I don’t know” is a feature. Low-confidence questions are handed to a named person and logged as a documentation gap.
- Accuracy is tested against a set of real questions with known answers before launch and after every model or document update.
When not to build one
- The documentation does not exist or is badly out of date. Retrieval makes wrong documents easier to find. Fix the source first; we can help scope that.
- Fewer than a few dozen questions a week. A well-organised wiki and a search box are cheaper and good enough.
- The answers carry legal, medical or financial consequences for individuals. An assistant can point to the policy; it should not be the decision.
- Nobody will own the documents. Without a review owner, the assistant is confidently stale within six months.
What we measure
- Answer accuracy on the test set, per topic
- Share of questions answered with sources vs handed off
- Repeated questions in team channels (before and after)
- Time for a new starter to find standard procedures
- Documentation gaps found and closed per month
Tools we connect: SharePoint, Google Drive, Notion, Confluence, OneDrive, help-centre platforms; delivered in Slack, Teams or a web widget. Names indicate compatibility, not partnership.
Which package fits
Knowledge assistants are advanced scope by definition: document preparation, permission mapping, retrieval tuning and an accuracy test set. A focused internal assistant on one well-maintained document set fits Team (€6,900). Multiple departments, several document stores or a customer-facing assistant are Business scope (from €16,500) and start with the audit, which includes the documentation inventory. Private or self-hosted model options are available for sensitive content. Prices excl. VAT and usage; see pricing.
Questions about knowledge assistants
Does the AI provider train on our documents?
We use business or API tiers whose terms exclude training on your data, and we record the provider, region and terms in your subprocessor inventory. For highly sensitive content we can use EU-hosted or self-hosted models. Your documents stay in your own stores; only retrieved passages are sent with each question.
How do you keep the answers current?
Documents are re-synced on a schedule, so an updated policy is reflected on the next sync without retraining anything. Each document has a review owner and a review date; the assistant shows the date with its sources. Under a care plan we re-run the accuracy test after each sync and each model update.
Can it answer customers on our website?
Yes, but we recommend starting internally. A customer-facing assistant needs a narrower, fully approved document set, stricter hand-off rules and adversarial testing. It is scoped as part of customer support automation once the internal version has proved the documentation is sound.
What is the difference between this and a fine-tuned model?
Fine-tuning changes how a model writes; it is poor at storing facts and cannot be updated by editing a document. Retrieval keeps facts in your documents, where you can read, correct and permission them. For company knowledge, retrieval is almost always the right approach, and we do not sell custom model training.