AI chatbot for company documents: answers with sources

Key points

• Clean, current documents are the foundation of the project.
• Access is checked before information is retrieved.
• Answers and sources are tested with real questions.

When your team keeps asking the same person

A colleague looks for the procedure for a routine situation, finds two files and asks the person who knows which one is right. If that person is away, work stops or continues on an old version. An assistant can help people find information, but it cannot decide on its own which document is official.

Start with a clear area: internal procedures, product sheets or training material. A chatbot for every document in the company is hard to evaluate when sources have different roles and access levels. First choose the repetitive questions that have a verifiable answer.

What to prepare before connecting AI

Inventory the documents, remove superseded versions from the active sources and name an owner for each item. Note the date, the scope and the conditions under which it applies. If two procedures contradict each other, resolve the contradiction before asking the assistant to interpret it.

Formats must be readable correctly. Tables, scans and attachments may need extra processing. Check that the extracted information keeps the units and the relationships between rows. Uploading a PDF does not prove that every page can be used correctly.

What a useful answer should look like

Ask for a short answer tied to the question, with the source and the relevant section. Your colleague should be able to open the document to check. If the information is missing or ambiguous, the assistant must explain the limit and point to who can clarify it.

Decide which questions it may not turn into recommendations or actions. Finding a procedure and carrying out an operation in the ERP are different projects. Changing data requires separate permissions, validation and approvals.

How to protect internal documents

Separate public sources from internal ones and check the user's permissions before searching. It is not enough for the interface to hide a folder if its information can still end up in an answer. Test accounts with different access levels and questions that try to retrieve restricted data.

Choose providers and configuration based on the data being processed. Check how information is retained, accessed and transferred under the terms of the services you use. Do not connect sensitive documents just because a demo works with public files.

• Approved sources and owners responsible for updates.
• Roles and the documents each user is allowed to see.
• Answers with a verifiable source.
• A procedure for missing or contradictory information.

How to test before your team uses it

Collect real questions, including some with incomplete answers, outdated documents or restricted access. Check that the answer is correct and that the source actually supports the claim. A convincing answer with the wrong reference should not count as a success.

Compare the time needed to find information and the number of answers that need correcting. Keep a feedback channel and withdraw sources that produce errors until they are fixed. Do not promise perfect answers. Define the conditions under which the system can be used.

The relevant example and the first project

In MedProtocol, Webmate brought protocols and training together on web, iOS and Android. The example shows information organized in an accessible system. It does not show an AI chatbot in that project. The distinction matters when you choose the proof for a new assistant.

Prepare a set of anonymized documents and questions for the conversation with Webmate. We can evaluate whether an assistant is an improvement over ordinary search and which checks are needed. A pilot project must show useful answers in that area before it is extended.

Frequently asked questions