An AI Employee Should Show Its Work

Ask a colleague how many customers converted last week and you will usually get more than a number. You will hear where it came from, what was counted, and how fresh it is. That padding is not politeness. It is what makes the number usable. A figure you cannot trace is a figure you cannot act on.

AI assistants are notoriously bad at this. They answer with total confidence whether the number came from a live query, a three-day-old note, or thin air. The number might even be right. You still cannot put it in front of your team, your board, or an investor, because you cannot say what it means or when it was true.

We hit this ourselves. An AI employee quoted a conversion count as current when it was actually remembering an older conversation. Days later a fresh check produced a different number, and the two answers looked like a contradiction when they were really two snapshots taken days apart, counted two different ways. Nobody had done anything wrong except skip the part where you show your work.

What showing your work actually means

So we made it a standing rule for every AI employee on AgentTeams. Not a suggestion in a help doc. A working standard that every employee follows in every conversation, on every connected channel.

Every figure names its source and its definition. Not “438 converted” but “438, from our production database, counting sign-ups from July 20 to 26 who now hold an active paid subscription.” The definition matters as much as the source. Most numbers that look wrong are actually two different definitions wearing the same name.

Current means measured now. An employee asked for a current figure runs the check in that moment. If a value comes from memory instead, it carries its date, so an old snapshot can never dress up as today's truth. And when two of its own answers disagree, the employee reruns the checks and explains the actual cause instead of guessing at one.

Answers arrive in your language. A German question gets a German answer, even when the underlying spreadsheet, ticket, or database is in English, and the other way around. Rules you set still win: a support employee told to answer every ticket in the ticket's language will keep doing exactly that.

You can see it, and so can we

Each reply in the conversation view now carries a small note listing the systems that turn actually consulted, like “checked SQL Database, Google Analytics.” It reads like a receipt. When an answer checked nothing because nothing needed checking, the note simply is not there, and that absence tells you something too.

Behind the scenes, the platform holds employees to the standard. An hourly review looks across real conversations for the two failures that matter most: quoting figures without having checked a system, and answering in the wrong language. Findings go to our team, and they are how we caught and removed the last rough edges within a day of shipping the rules.

None of this required the employees to get smarter. It required the workplace to have standards. Trust in a number is built from its paper trail, and paper trails can be designed.

If your team runs on numbers, this is the difference between an AI employee whose answers you forward and one whose answers you quietly re-check. See how it fits your stack on our pricing page, or read how thinking power per employee keeps the hard questions with your sharpest workers.