Procurement and commercial staff ask questions in Microsoft Teams and get verified answers from millions of purchasing, receipt, payment and stock rows in about a minute.
How much did we spend with this supplier. How many orders are awaiting approval. Simple questions, asked all the time.
Answering one meant an analyst, a spreadsheet, and days of waiting. Every question became a task in someone else's queue.
Off the shelf chatbots were not an option. They were not trusted with real commercial data, because they cannot show their working. A figure with no trail behind it is a figure nobody will act on.
A governed agent that runs entirely inside the client's own data platform. The agent goes to where the data lives, not the other way around.
It works under two layers of written rules. One layer governs how SQL is written. The other governs what may be claimed in the answer. They are separate on purpose. A correct query can still produce a misleading sentence, so the sentence gets its own rules.
The agent is wired into Microsoft Teams. People ask where they already work, and the answer lands in the same chat.
Every question runs through the same steps, whoever asks it.
The agent's behaviour is tested, not assumed. 379 automated checks run against it, grown from 140 during user testing. Each check pins down something the agent must do, or must refuse to do.
Every figure in every answer traces back to the rows it came from.
The clearest signal came from the users themselves. The client's team adopted the agent unprompted within days of going live.