How the man who audits AI for a living decided to trust one

Half an hour on two sentences

Chris's email is relationships: introductions, invitations, the follow-up after a good conversation. The volume is not the problem. The care is.

“I've been known as the type of person who will sit down and stress over two sentences for half an hour.”

And the ordinary fix does not fit. “You can't really automate labels or tags. If you don't know the address it's coming from, you can work on keywords a little bit, but that's not what any traditional email client filtering system is based on.”

An inbox made of relationships, not volume

His mail is introductions, invitations, and the follow-up after a good conversation. It is not high-volume and that is exactly why the usual tools miss it.

“You can’t really automate labels or tags. If you don’t know the address it’s coming from, you can work on keywords a little bit, but that’s not what any traditional email client filtering system is based on.”

A filter needs a rule, and a rule needs something stable to match on. A first approach from someone he has never met has no address he knows and no keyword that separates it from a newsletter. The thing that makes it matter is what it is asking for, which is the one thing a filter cannot see.

Why a man in his job checks the reasoning first

Cognify Solutions exists to ask whether AI systems can be accounted for. His own description of the work: making sure systems are built intentionally, and that you can sit down and say why one agent answered this customer the way it did, and that customer differently.

He has seen what happens when nobody asks. A lender brought AI in to clear a backlog and started refusing people. The complaints came before the diagnosis did, and the diagnosis was that the training data reached back decades into an area’s financial history. The model had reinvented redlining, faithfully, from the record it was given.

So his interest in our reasoning panel is not curiosity. It is his profession, applied to a tool he was being asked to trust with his own correspondence:

“How far can we really get down to what is this thinking, why this decision was made, can I be comfortable with this?”

Show the reasoning beside every draft

Chris runs an AI governance practice. His job is deciding whether a system’s decisions can be explained, and he has watched what happens when nobody checks. One lender’s model learned from decades of an area’s financial history and quietly reinvented redlining.

So on his first look at Clarity he went straight past the output to the reasoning shown beside it.

“If I'm not comfortable with how it got to the result, I cannot trust that it would do that predictably.”

“A black box is an anti-pattern. It's very significantly an anti-pattern in any type of user experience engineering. You get a gold star from me just for including that in there.”

Learn his manner from the threads he already wrote

The proof he looked for was not accuracy. It was recognition: had the system picked up the same details about a contact that he would have picked up.

“If you're picking up on details in messages that I would pick up, then I know you should be more well aligned to start to speak for me.”

He gave it a real thread to learn from, a client raising a problem with the reply softened and rebuilt, because that is where a person’s actual manner shows. “Capturing the actualness of that is brilliant.”