What shipping AI taught me.
Not hot takes. Working principles, formed by shipping AI into places where a wrong answer costs money, licenses or trust.
From data gatherer to judge
Sit with a financial crime analyst for a day and count the minutes spent actually deciding anything. Most of the day goes to gathering: pulling transactions, screening names, copying context between tabs, assembling the picture. The judgment part, the reason the job exists, gets squeezed into whatever attention is left.
That is backwards, and agentic AI is how it gets fixed. Not a chatbot that answers questions faster. An agent that does the work: analyzes the transactions, screens the counterparties, drafts the narrative, and shows up with a case file instead of a search box.
The goal was never to replace the human. It is to promote them. When the machine gathers, the human judges. The institution does not just get faster; the decisions get better, because the person making them finally has the time and the context to make them well.
Glass box or no box
In consumer software, an unexplainable model is a curiosity. In a bank, it is a liability with a deadline. Regulators do not accept “the model said so”, and they should not.
That changes what you build. Explainability is not a feature you add before a demo; it is the license to operate, designed in from the first line. Every finding traceable to its evidence. Every recommendation auditable after the fact. Every automated step separable from the human decision it feeds.
The discipline pays beyond compliance. Teams trust what they can inspect. Customers adopt what they can defend internally. And products built as glass boxes survive contact with the real world far better than black boxes with better benchmarks.
The simplest useful version
Every product I have shipped that mattered started embarrassingly small. The overengineered ones died before they could embarrass anyone.
When the risk of overbuilding is high, and with AI products it always is, I start with the simplest version that is genuinely useful, put it in real hands, and let contact with reality set the roadmap. Configuration over hardcoding. Prompts over pipelines, until the pipeline earns its complexity. Version 1.0 out the door, and honest hunger for version 2.
This is not about lowering the bar. It is about refusing to spend months polishing assumptions. The fastest way to a great product is a real one.