Deciding when to trust the model.
Evidence has to be inspected before a recommendation can become authoritative. The system retrieves and proposes; people review, override and publish.
About
I’ve designed products across ServiceNow, Compass Healthcare Digital, and Charles Schwab, spanning enterprise platforms, healthcare operations, and financial services.
I tend to get involved while the product is still unsettled: competing assumptions, unclear ownership, workflows that don’t match reality, or a new capability without an agreed product shape.
My role is usually to find the structure underneath that ambiguity and turn it into something a team can actually build.
Case studies
Two different questions. One is what an AI product should let a person do before its output counts as a decision. The other is what has to be true of a system before an organization can build on it at all.
Judgment
Deciding what an AI system may do on its own, and what it has to hand back to a person, is the call I make on every product I work on. It is rarely a model question. It is a question about consequence.
Where can the system act?
Low-cost, recoverable actions can move without a person.
Where must it ask?
Consequence and permissions can make review part of the product.
What happens when it’s wrong?
Recovery, reversibility, and accountability have to be designed in.
How much of the mistake can be taken back.
One click undoes itTakes work to undoPermanent
What it costs when the system is confidently wrong.
A minor annoyanceCostly to fixReal harm
How far the consequence travels past the person who triggered it.
Only the person actingTheir whole teamEveryone downstream
Recommended operating mode
Easy to take back and costly to get wrong, while the impact stays with the person acting. The system can prepare the action, but a person makes the commitment.
From the dials
My call: let it prepare the move, but the yes stays with a person. Ask me why, or move a dial and I may change my mind.
Experiments
Where I test these decisions in code, interfaces, and policy. These are unfinished on purpose: each one exists to answer a question I could not answer by reading about it.
Working notes
Ideas I keep returning to while building AI products.
Compare notes
System acts
Where errors are bounded and recoverable.
Pauses for approval
Where a person sees the evidence before the system commits.
Hands control back
With a clear record of what changed, what triggered it, and why.
Send a note. I reply within two business days.