In some industries, an AI that's brilliant nine times out of ten and confidently fabricates on the tenth is worse than no tool at all. Wine compliance, financial decisions, customer relationships — trust isn't a feature of the product. Trust is the product.
Not every industry needs AI to be trustworthy in the same way. There's a whole category of applications where being confidently wrong some of the time is fine — the stakes are low, the output is easy to verify, and a bad answer costs you a few seconds and a shrug. Move fast and be occasionally wrong; nobody gets hurt.
Then there are the industries I've chosen to build in, where that posture is disqualifying. Wine compliance, where a wrong answer breaks a regulation. Financial decisions, where a wrong answer costs real money. Customer relationships, where a wrong answer erodes something you spent years earning. In these places, a tool that's brilliant nine times out of ten and confidently fabricates on the tenth is worse than no tool at all, because it teaches people to trust it right up until the moment it fails them. In industries like these, trust isn't a feature of the product. Trust is the product.
Building for that standard changes nearly every design decision. You stop optimizing purely for the impressive output and start optimizing for the believable one. Provenance stops being a nice-to-have and becomes structural: every claim should be able to point at where it came from. Showing your work stops being a transparency gesture and becomes the core value proposition, because in a high-trust industry the reasoning is the deliverable. And the system has to know its own limits — it has to be willing to say "I'm not sure" instead of smoothing over the gap with a plausible guess.
I learned the sharpest version of this lesson building a feature for wine label recognition. The tempting design was full automation: snap a photo, auto-fill every field, feel magical. But in a product whose entire premise is helping people trust their own record of what they tasted, a single confident auto-fill error does more damage than a hundred manual entries. One wrong vintage silently written into someone's tasting history, and they stop trusting the whole system. So the right call was to favor trust over convenience — to surface what the system saw, flag its uncertainty, and let the human confirm, rather than to optimize for the slickest possible demo. Convenience that quietly corrodes trust is a bad trade in any industry where trust is the franchise.
This is the thread that connects everything I work on, across wine and consumer insights and consulting. They look like different businesses, but they keep teaching me the same thing: in the domains worth building for, the winning AI is not the one that's most impressive in a demo. It's the one that earns its place — that shows its work, knows its limits, and never spends the user's trust to look smarter than it is. The model will keep getting better on its own. The trust is the part you actually have to build.