An AI assistant with company documents lets you ask about a procedure, a product, or an internal rule and get an answer based on the sources you’ve connected. To roll it out, you need to choose which queries it will handle, prepare the current documents, define who can see them, and check the answers before opening it to the team.
Think of that question that always ends up with the same person: “What do we need to register a distributor?”. If the answer is already written down, the assistant can help find it and explain it. If every colleague gives a different answer because the procedure hasn’t been agreed, then the company still has a decision to make.
How it goes from a question to an answer
A common way to build these assistants is RAG, or retrieval-augmented generation. The system looks for relevant information in the available sources and passes it to the model so it can draft the answer. Azure AI Search documentation on RAG describes this pattern and how to prepare content so it can be retrieved. It’s not about retraining the model every time you change a manual.
The flow it’s worth asking for in a demo is simple:
Team question → search in authorized sources → answer with reference → open the document to verify it.
A useful reference lets you locate the file and the section that support what was said. If the assistant mixes two documents up or answers about another product with a similar name, the link should make it easy to spot. A smooth answer still needs checking.
Choose a query that interrupts the team today
Start with a specific task: checking product sheets, finding steps in a procedure, or clearing up onboarding questions. Gather the questions exactly as they come up at work, including abbreviations and informal names.
For each one, note who asks it, which document answers it, and what that person will do next. That way, you can tell the difference between a document lookup and a request that needs live data from another program.
For example, “what’s included in this service?” can be answered with its approved sheet. “How much stock is left right now?” needs to check an up-to-date stock source; a saved catalog is not enough. “Can we make an exception for this customer?” may require a manager’s decision even if the general policy is documented.
If you’re still comparing which task to tackle first, our guide to process automation in a company explains how to narrow down that first scope.
Prepare the version the team should consult
For the initial set, create a list with the document name, its official location, owner, current version, and authorized profiles. Mark which files are out of scope. You don’t need to review the whole historical archive to test one specific query.
Also check how the content reads. Open the text the system retrieves from a scanned PDF or a table: a shifted column can attach a condition to the wrong product. Keep the titles and the context needed to interpret each section.
Agree on what will happen when someone replaces, removes, or moves a source to another folder. Include a full test: change a condition in a test document, wait for the expected update, and ask again. Repeat after removing that document. The result should reflect the change within the agreed timeframe, and sync failures should be visible to whoever maintains the system.
Keep permissions in place when connecting sources
Define what each profile can query before connecting folders. A sales person may need commercial sheets without access to management documents. Check that separation both in the answers and in the links the assistant returns.
The connection method matters. For example, SharePoint integration in Copilot Studio queries content using the authenticated user’s permissions. That behavior belongs to a specific integration and setup: you need to verify how access is preserved in the chosen solution, especially if files are copied to another repository.
Ask to test it with accounts from different profiles, including one without access. Add another check after removing a permission. Don’t assume the separation works just because it works with the account of the person who built the assistant.
Example: registering a new distributor
This example is illustrative. Let’s say the approved procedure requires completing a sales sheet, assigning an owner, and getting operations approval before activating the distributor.
A salesperson asks: “I’ve got a new distributor, what should I do?”. A useful answer would summarize those steps and point to the procedure they come from. If the document includes the template to use, the assistant should provide a link the salesperson can open.
The second question might be: “Can I activate it without approval?”. If the document doesn’t cover exceptions, the assistant should leave that decision to the designated manager. It should not turn a common practice or a comment in the conversation into authorization.
Also prepare for the case where it only finds an old procedure or two conflicting versions. The expected behavior should be defined: state that it can’t confirm the current process and direct the question to the person responsible for maintaining it. That issue helps correct the documentation.
What to check in a demo
Bring your own test questions and documents, with expected answers reviewed by someone who knows the process. This table can serve as an acceptance checklist:
| Test | What the team should be able to verify |
|---|---|
| Frequent question with a documented answer | The explanation matches the current source and lets you open it. |
| Two products with similar names | It asks for the missing detail to identify the right one. |
| Question with no answer in the sources | It explains what information is missing and offers a defined next step. |
| Two contradictory versions | It applies the agreed validity rule or asks for review. |
| User without permission | It does not show restricted content or summarize it. |
| Updated or removed document | It reflects the change within the agreed sync window. |
Also test follow-up questions, because a conversation can carry an incorrect assumption forward. Opening the source and comparing the answer is still necessary: Microsoft warns in its knowledge source documentation that disabling responses without documentary support does not guarantee the model will stop adding general knowledge when combining information.
During the pilot, log resolved queries, answers that need correction, wrong references, and time spent searching and reviewing. Count the maintenance work for the documents too. With that data, you can decide whether to expand the sources or fix the scope first.
At Digency, we develop custom AI and automation solutions. Tell us what question your team keeps repeating and where the answer is: we can start there to define an assistant and a test you can evaluate with your own documents.
Cover: illustration generated with AI.
