Our employees had the answers somewhere. Now they can get to them.

Every question was an interruption, and one person was the search engine

A small operation runs on knowledge that lives everywhere and nowhere: a part has a formal ERP name nobody uses on the floor, a process answer sits in a manual, a lead time means comparing two systems, and the person who knows where the answer lives is on the phone. Most questions got answered by interrupting that person.

Then our operations lead gave notice, and the risk became immediate. The role turned out to be a bridge across the CRM, the ERP, the scheduler, the inbox, and years of lived context. Losing the person meant losing the bridge.

First we captured the knowledge. Then we made it answerable.

Before the handoff ended, we pulled together roughly a year of calls and Teams conversations, about six months of email, our whole document library, and two weeks of recorded shadowing: voice notes and screen walkthroughs of the work that had been living in one person's head. All of it went through review before the agent was allowed to use it.

When two sources disagreed, I did not want the agent to guess. Every conflict went into a queue and got settled by a person, one at a time, mostly on the drive to work. Only the settled version became something the agent could say.

The agent itself lives inside Microsoft Teams, the tool everyone already has open. An employee asks in plain words. The agent looks up the answer in the right business systems, checks what that employee is allowed to see, and comes back with an answer and its source, and how old the data is. When it cannot answer, it says so and files the gap for us instead of improvising.

Four kinds of work it does every day

Find the document before the trip. A vendor could not match a part name on a purchase order to the physical piece, and a production leader was about to drive over to point at it. Asked from a phone in the back of a car, the agent found the drawing on the right page of a document in our file share in about two minutes. Nobody left the building.

Check a quote before sales promises a date. A salesperson pastes a quote number and asks whether we can build it. The agent reads the quote, checks the materials list against current stock, and looks at material already on order before anyone commits.

Say no to the wrong person. When a request lands outside someone's lane, the agent refuses and offers to route it, because it checks who is asking on every call. Refusing is a feature, not a failure.

Turn "I don't know" into work. When the verified sources cannot answer, it says so plainly and routes the question for human review. A missing answer becomes a gap to close, never a confident guess.

The questions our employees actually ask it

The agent is used across the operating questions that used to interrupt someone: orders and delivery status, inventory and whether a job is buildable, product and technical specs, scheduling and service, finance and purchasing, and the how-do-we-do-this process questions. The mix is what you would expect from a shop: mostly "where is this order" and "can we build this," with a long tail of everything else.

It does not give everyone access to everything

The agent follows the same job-based access our business already uses. Everyone can ask about general knowledge and product information. People who already work in scheduling, production, inventory, or purchasing can reach those lanes. Finance and dealer terms stay with the roles that own them. And the few actions that change a business record preview first, wait for a person to confirm, and read the result back afterward. The permission check happens on the server, not in the conversation, so it cannot be talked around.

The honest part - the first version had real rough edges

Not everyone came back after trying it once. That was useful feedback, not something to explain away. The knowledge base had real holes, and logging a gap helped us improve it but did not solve the employee's problem in that moment. Teams itself needed work: early on the agent lost the context of quoted replies and could not read attached documents, and we fixed both after people ran into them. And read-only can feel like failure. Employees naturally ask an agent to change things, and it had to learn to say no when a task was not ready or not safe.

It was useful, but it was not magic. It made the undocumented parts of the business impossible to ignore.

Start with one question your team asks every week

Here is what we would tell any owner: do not connect an agent to everything on day one. Pick one question that interrupts a person every week, connect the source that actually holds the answer, and make the agent show the source and its age. Route contradictions to a named person instead of averaging them away. Keep everyday questions separate from changes that need approval. And treat a missing answer as a chance to improve the business, not a reason to make something up.

What an internal AI agent still cannot do: replace an employee's judgment, know facts nobody ever wrote down, give every employee access to sensitive information, or safely make every change on its own.