The advice now going to technology chiefs is that bias in company AI cannot be fixed with a one-time audit or by buying a better model. The tool needs an owner for as long as it runs. That is right, but a manager can also do something alone: understand why the bias is there, and correct for it in their own questions.
The assistant reads the record, not the company
When you ask an assistant about your business, it reads what exists in writing: chat threads, tickets, emails, meeting notes. Writing is a sample of what happened, and the people who like to write choose the sample. A confident colleague posts every day. A phone call leaves nothing behind. A decision from the top gets retold in a chat group, a little different each time. The assistant cannot tell which things are missing, so it answers with the same confidence either way.
For example, a shop owner with three branches asks the assistant to read the shared customer inbox and say which branch needs help. One manager hands every customer a card with the branch's email address. Another deals with complaints at the counter and never writes anything down. The assistant points at the first branch. It has measured which branch invites emails, not which one is doing badly.
Fixing the words you can see does not work
Amazon's résumé tool, which it built in 2014 and later dropped, shows why you cannot patch this. Amazon found that the tool marked down one word and made it ignore that word. But the tool had also favoured language that men used more often, and a list of banned words cannot catch a pattern that is spread through everything the tool read. The slant was in the record, so it came out in many small ways.
The shop owner has the same problem. Telling the assistant to ignore how many emails each branch sent will not help, because the whole inbox is shaped by who writes. What works is changing the record. If each manager spends five minutes a week typing the complaints they heard at the counter into the same log, the next summary has both branches in it. That is under half an hour a week across three branches, and it makes every later question about customers more honest.
A confident answer is the part people believe
The second problem is on our side of the screen. A tidy, confident paragraph is easy to accept when you are busy, and the assistant writes tidy paragraphs whether it has seen the whole picture or a corner of it. At the network security company Infoblox, every AI-assisted decision that matters has a named business owner, and the checking grows with what the AI is allowed to do. A tool that summarizes documents gets a light look, and an agent that moves money needs a person's approval.
A small business can copy that without any committee. The more an answer will move money or touch someone's job, the more it needs one named person who checks it against something other than the assistant.
A useful way to think about the tool is as a new hire who has read every document the company ever produced and has met nobody. Such a person knows a great deal, and knows nothing that was said in a corridor. Before you rely on that person, you test them. Ask the assistant three questions whose answers you already know, such as which branch is really struggling or why you changed supplier last year. Where it gets them wrong, look at which way the mistakes lean. That direction shows you whose voice is too loud in your record, and what to add to it.