You can't demo your way to trust.

In the early days, my product partner and I had these multi-hour Zoom sessions — no agenda beyond "let's think out loud until something breaks loose." Most of them ended with more questions than answers. This one gave us our north star.

I was working through how trust actually gets built, out loud, badly, the way you do. Two things kept surfacing. Trust isn't granted, it accrues — you earn it in small increments by being right about small things, repeatedly, in the open. And just as importantly, you earn it by knowing when you don't know and saying so. That second one became a real guardrail: Charlotte had to be able to recognize the edge of its own knowledge and admit it. An assistant that answers everything is an assistant nobody believes.

We both lit up. That was the one. I said the sentence; my product partner immediately started mapping it against how security teams actually level their people. That's the move that turned a decent analogy into a roadmap, and it happened inside the same hour.

What the metaphor bought us wasn't a nicer way to describe the product. It was a sequencing argument we could take to anyone. Charlotte starts as an intern. It earns analyst. Then investigator. Then, eventually, teammate. Every rung has to be paid for with a track record from the rung below it.

And the ladder climbs toward people, not over them. Each rung is defined by what work leaves someone's plate, not by who gets replaced. The most autonomous thing at the top is still a peer working a hard problem next to you.

A metaphor that only lives in a deck is decoration. Ours needed something operational underneath it, so my product partner and I wrote three principles in one of those same working sessions, took them to leadership, and got real buy-in before we did anything else with them.

They read as modest. They were not. They were a shared test we could run out loud in a room without it becoming personal — which is the part I care about most. Disagreeing with someone's feature is a fight. Checking a feature against three rules everyone already agreed to is just Tuesday.

They didn't kill ideas. They made ideas clearer. If a response came back with more than the prompt asked for, we cut it down. If a feature couldn't show its work, it got rewritten until it could. Scope creep and slippery-slope arguments both lose their footing fast when there's a shared standard to point at.

I walked the team through them with examples, and then we started using them. They began as design principles and became product principles. Later, they were one of the first things we shared with any team who wanted to build AI into their own part of the platform — before the patterns, before the components. Here's how we think. Start here.

Leadership said this a lot. And the ideas were genuinely, really cool. That was the hard part — saying no to a bad idea is easy.

But the answer was never no. The answer was not yet. You don't kill a great idea. You hold onto it until you can execute it well, because you respect it. That distinction mattered enormously to how those conversations went. Nobody stops bringing you ideas when the response is "let's earn our way to that one."

The reasoning was almost always the same: we could do that once, on a stage, and it would be magnificent. We could not do it reliably, across the range of real questions real people ask, on a Tuesday afternoon in someone's actual environment.

If the assistant is confidently wrong once in front of a security professional, you don't get a second look — you've just confirmed everything they already suspected.

So we spent longer than anyone wanted proving Charlotte could answer straightforward questions excellently before we let it answer impressive ones at all. Boring, on purpose, for a while. That was the bet, and it's the one I'd defend hardest.

The thing I didn't anticipate is how far a good metaphor travels once other people find it useful.