AI inside your admin screen, not in a separate tab

Somewhere in your company there is someone with two tabs open. One holds your system, the other a public AI tool. They copy a chunk out of an overview, paste it into the chat, ask their question and copy the answer back.

That works, more or less. There are just two holes in it. Your company data now sits somewhere you do not control. And the AI knows nothing about your system, so it works with whichever fragment happened to come along.

The idea: bring the chat inside

The answer is not to ban the AI. You lose that fight, and rightly so, because people use it because it helps.

The answer is to bring the chat inside. On a recent project it now sits in the admin itself, in the messages menu, behind the same login and the same permissions as everything else. Anyone without access to the shared mailbox does not see it. What someone asks stays theirs.

That is the boring half. The interesting half is what happens before the AI is allowed to say anything.

The question before the question

Suppose someone types: which locations in France are still free in September?

An ordinary AI chat will make something of that. It has a language model and no calendar, so it produces something that sounds like an answer. That is more dangerous than no answer, because a guess is indistinguishable from a fact once it sits in a well-formed sentence.

What happens now: the server sees from the question that it is about availability. It works it out itself, using exactly the same check the booking form uses, and hands the AI only that outcome. The instruction alongside it is explicit: here are the facts, invent nothing.

That splits the work the way it should be split. The system does the arithmetic, because it is better at that. The AI turns it into a readable answer, because it is better at that.

After the first round of testing with the administrator, that recognition was widened. It now understands a country, a month and "all locations", instead of only a specific name with a specific date.

Saving works in reverse

One detail I thought about longer than about the chat itself: what happens to a conversation.

The usual approach is to keep everything. That felt wrong. A chat is thinking out loud. Three quarters of it is scratch work, a half-finished question, a badly phrased instruction. You never need to see that again.

So it saves nothing unless you press Save. But you should not lose a conversation by accident either, so it runs like this: an unsaved conversation stays waiting, your browser warns you when you leave the page, and on your next visit it is there again with the question of whether you want to keep it. Yes keeps it, no discards it, and carrying on talking is fine too.

Attachments without a second archive

You want to be able to paste a screenshot and hand over a spreadsheet. That is exactly where a feature like this quietly goes off the rails: within six months there is a second document archive that nobody maintains.

So it has an expiry date. Pasting works with Ctrl+V, uploading is limited to a fixed list of file types, and spreadsheets and documents are unpacked to text on the server so the model never has to open a file. What you upload sits separately per person, under a random file name, in a folder that is closed off from outside. Every legitimate read passes the login and an ownership check. Anything that is no longer attached to a conversation after a week is cleaned up.

One route, and it says so when it is down

There is one connection to the model, and there is no fallback to a second route.

That was a deliberate choice by the client, and I agree with it. A fallback that switches in silently produces answers of a different quality without anyone noticing. You end up relying on something you cannot identify.

If the connection is down, the screen says so out loud. An honest outage beats an invisible detour.

The fuse

Per user there is a limit on the number of questions and uploads per ten minutes.

There are two reasons for it. The first is money: you pay per question, and a script that fires thousands of questions overnight is an invoice. The second is misuse from inside, and that is why the limiter belongs there even if you trust everyone. A fuse is not there because you distrust your house.

What it is and is not for

It is not a replacement for your screens. Filtering a list is quicker in the list.

What it is for: the question you would otherwise have asked a colleague. How are we looking in September. Turn this into a short report. Write a first version of this text. Sketch this idea out so we can talk about it.

Those are questions nobody puts in a menu, and for which you used to open a second tab.

Read also

• AI that reads your own website first, and only then answers the customer, the same principle for your inbox
• What AI really means for your SMB (without the hype), where AI does and does not belong
• View all completed projects, web platforms and automation solutions

Are your colleagues pasting company data into a public AI tool? Book a no-obligation call. A chat inside your own environment is often less work than you would expect.

Bert