Why I Built an AI Assistant That Reads the Handbook.
What building a document-based AI assistant taught me about finding company policies, checking sources, and knowing when to ask a person.
When I worked with dealership support systems, I heard some version of the same sentence more times than I can count: "I know we have a document on that somewhere."
Sometimes it was a procedure. Sometimes a training guide. Sometimes the answer was buried in an email that three people remembered receiving, but nobody could find.
The information existed. Finding it when somebody actually needed it was another matter.
That old problem is one reason I've been experimenting with a different kind of AI assistant. Not one that claims to know everything. One that starts by looking in the right documents.
The handbook is useful. Finding the right page is the problem.
Think about a new employee trying to figure out whether a particular tool is approved, how an expense should be submitted, or what to do when a customer asks for something outside the usual process.
The company might have perfectly good written policies. But those policies may be scattered across a handbook, a shared drive, a procedure manual, and a few updates that came along later.
People naturally ask whoever seems to know the answer. That works, until the knowledgeable person is out for the day or gets asked the same question for the tenth time.
I wanted to see whether AI could make those documents easier to use without turning a policy question into a guessing game.
What I built
For Aegis AI Advisory, I built a public Policy Q&A demo using documents for a fictional professional-services company. The example includes an employee handbook, an AI-use policy, and selected public NIST guidance.
You can ask it a question in ordinary language, such as:
- Can I put client information into ChatGPT?
- What should I do before trying a new AI tool?
- Does somebody need to review AI-generated work before it goes to a client?
Instead of simply producing an answer that sounds plausible, the assistant searches the approved source material and brings back passages you can inspect.
You can try the working Policy Q&A demo here.
It's a demonstration with fictional company documents, not a place to upload real employee records or confidential client information.
There's a name for this: RAG
The approach is called retrieval-augmented generation, or RAG. That's a technical name for a fairly straightforward idea.
First, the system looks through a collection of approved documents for information related to your question. Then the AI uses those passages to help prepare a response. It can show which passages it relied on so you have somewhere to check.
Think of it as the difference between asking somebody to answer from memory and asking them to open the manual first.
It is not magic. The assistant can still misunderstand a question, miss an exception, or select the wrong passage. A citation is a way to verify the answer, not a guarantee that the answer is correct.
But at least you're giving the person asking the question a trail back to the source.
My favorite test is a question it cannot answer
When I test one of these systems, I don't only ask questions I know it can answer. I deliberately ask things the documents don't cover.
Maybe the handbook explains the normal approval process but says nothing about an unusual exception. Or perhaps somebody asks for a decision that belongs to a manager, not a software tool.
That's when I want the assistant to say, in effect, "I can't find that in the material I was given."
I'd much rather see that than a confident paragraph somebody mistakes for company policy.
It also raises a useful question for the business: Who should handle that unanswered request? Good AI doesn't make that responsibility disappear.
The familiar IT work still matters
After 25 years working in IT, I'm convinced the interesting technology is often the easy part to talk about. The harder part is making it dependable in day-to-day use.
For a document assistant, somebody still needs to answer some very ordinary questions. Which version of the handbook is current? Who can change it? Can everyone see every document? What happens when a policy changes? Who checks the answers and handles exceptions?
You wouldn't give every employee access to every personnel file just because searching them became easier. The AI needs the same sensible boundaries.
And I wouldn't judge success by how impressive its first answer sounds. I'd ask whether people find the right guidance faster, whether they can check it, and whether the system knows when to hand something back to a person.
Where I think this could help
An employee handbook is only one example. A similar approach could help a small business search approved training materials, standard operating procedures, onboarding guides, or technical documentation.
Imagine a service technician finding the right procedure without digging through five folders. Or an office manager locating the section that explains a routine process instead of interrupting a colleague.
I see possibilities for professional firms, dealerships, and organizations that have accumulated years of useful knowledge. The real question is whether the information is reliable, appropriately accessible, and worth making easier to find.
I'd start small: pick a few documents, write down ten questions people actually ask, and include a couple the documents cannot answer. Have the people who know the policies evaluate the results before expanding it.
If the assistant makes people more confident and saves them time, you've found something worth improving. If it just produces faster confusion, it isn't ready.
A useful assistant knows its limits
What I like about this project is that it solves a very old problem with a new tool. Businesses have always struggled to get the right information to the right person at the right time.
AI gives us another way to tackle it. But the goal isn't to replace the handbook, the expert, or the person making a decision. It's to make their knowledge easier to find and use.
If you're curious, visit the Aegis AI Advisory demo page and give the assistant a few questions of your own. Try one it should know and one it shouldn't. Both answers can teach you something.
Disclosure: I founded Aegis AI Advisory, where I help businesses find practical AI workflows and put sensible controls around them.
Thanks for reading.
See you in the next story.
