AI assistant says a colleague is not in my team
17 per cent of live employees had no manager recorded, so an AI assistant said a real colleague was not in the team. The check, and why names stay out of the index.
If your AI assistant says a colleague is not in your team, the usual cause is not the assistant. It is that the colleague has no manager recorded, so no leader’s team includes them. The first check is to count the people with no manager in your records.
Here is how that looked at one business, and what we found.
You ask the company’s new data assistant how many orders a colleague placed. It answers that nobody of that name is in your team. He is, and he is on the payroll.
We were building this assistant for the sales data of Aries Agro, a crop-nutrition manufacturer, when the owner proposed a shortcut: export the employee names and the distributor names into the assistant’s search index, so it could recognise a first name in a question. We argued against it, with figures from the real data, and in doing so found the gap that makes the colleague “disappear”.
The gap: 218 of 1,273 live employees, 17 per cent, have no manager recorded. They are unreachable from any leader’s team, so any feature that asks “who is under me” cannot find them.
What was actually going on
Seven employees share one common first name. A regional head’s 60-person team holds exactly one of them. Another leader’s 302-person team holds none. So the thing that decides which person you mean is not how the name looks. It is whether the person reports into you.
That is why putting names in a search index fails, and the first reason is not privacy. Seven near-identical strings sit almost on top of each other, so ranking among them is noise, and the fact that separates them, that one reports to you, is not in the text at all.
The second reason is order of events. The index is searched before any permission check, so any of the seven could surface in any leader’s question, and the assistant would then report another region’s sales plausibly and invisibly. The third is age: a snapshot of 5,561 dealers and staff goes stale the day somebody leaves.
And the dealer records carry tax numbers, phone numbers, email addresses and outstanding balances next to the names. It was never just names.
What we changed
We proposed and built a live lookup instead. The assistant asks the database who matches, inside the asker’s own team, before any candidates exist, matching exactly first and approximately after. It returns the employee code, designation and depot, so when several people match, the assistant can ask which one you mean. This was built and tested on the test system.
What it did not fix
The 17 per cent is not a software fault. Managers have not been recorded against 218 people, and until they are, anything limited to your team will say “not in your team” about real colleagues. That is a data fix for the client, and we reported it as one.
How to check your own records
- Count the people with no manager recorded. Any question limited to “my team” cannot find them.
- Pick a common first name and see how many people with that name each leader can actually reach.
- Ask whether names are searched before or after the permission check. Before is the wrong order.
- Ask whether the assistant looks people up live, inside the asker’s own team, or from a copy of names exported into an index that goes stale.
Where this ends up
An assistant is only as good as the records that say who reports to whom, which is the work of our AI data assistants service.