Notes from the field.
Where product design is heading, how I lead through it, and what thirteen years in lending, crypto and identity says about what comes next.
Latest
All posts
Give your managers something to approve
A third of managers are considering a move back to IC work, and the jump is biggest among managers of managers. The craft isn’t what’s pulling them.
Designers disagree and the judge decides
Twenty trained designers rated the same generated screens and agreed at a kappa of 0.25. The judge tuned on their labels inherits that argument.
A tying ban needs a decline path
The Fed’s new stablecoin proposals ban tying for every issuer. That ban has always been hard to enforce because nobody specifies the screen.
All posts
Design leadership
Give your managers something to approve
A third of managers are considering a move back to IC work, and the jump is biggest among managers of managers. The craft isn’t what’s pulling them.
The interview was never the expensive part
Greenhouse found 51% of candidates who finished an AI interview never got an outcome. The new capacity went where hiring was already cheap.
A new role isn’t a new process
Leaders picked design engineer as the next core role while half of teams changed nothing about how design gets approved.
Getting hands-on costs you the crit
Managers already spend a median 40% of the week on individual contributor work. The advice to get hands-on with AI tools comes out of the crit.
Promote on decisions you wrote down
67% of companies tie promotion to AI skills. Designers say they can’t fully account for their own output, so the evidence has to be recorded early.
The bottleneck moved and the budget didn’t
Developers and engineering leaders both name reviewing AI output as their top constraint. Teams are spending the gain on deferred hires instead.
Count the drift, not the adoption rate
Design system buy-in fell from 42% to 32% in a year. Teams are told to build a business case from a return only 5% of them measure.
Train juniors on what they reject
AI closed the output gap between junior and senior designers and hid the judgment gap. Review what juniors throw away, not just what they ship.
Your team stopped taking your advice
Peer learning among designers went from 24% to 70% in a year while learning from leadership halved. That’s structural, and most leaders answer it wrong.
Where product design is going
Designers disagree and the judge decides
Twenty trained designers rated the same generated screens and agreed at a kappa of 0.25. The judge tuned on their labels inherits that argument.
The agent edits the standard it follows
NN/g watched seven power users of agent tools. They let the agent rewrite the files that define good, and rarely read what it changed.
The summary dropped the user who disagreed
Researchers trust AI most for summarizing. Across 18 models, bias against minority views ran 34.4 times larger than random. The objection goes first.
Generated screens still need a record
A Google banking prototype scored 84.38 on usability against 53.96 for the fixed app. The protocol that ships those screens keeps no copy of them.
A design review can’t see accessibility
WebAIM found WCAG failures on 95.9% of a million home pages and 56.1 errors on the average one. Those defects don’t show up in a screenshot.
Disclose the review, not the tool
58% of creatives have used AI on client work without saying so. The label they’re all arguing about carries a penalty and no information.
Prototypes that cannot fail are demos
43% of companies expect a working prototype as the deliverable. A deliverable gets approved. A prototype is supposed to be able to come back wrong.
Your design system checks spelling, not judgment
Eighteen of 21 major design systems now ship an MCP server for agents. What they enforce is the vocabulary. The usage guidance is the part getting cut.
Stop hiring for AI fluency
Half of design leaders now weigh AI fluency in hiring, while the average toolstack went from three tools to seven in a year. That screens for recency.
Fintech, web3 and trust
A tying ban needs a decline path
The Fed’s new stablecoin proposals ban tying for every issuer. That ban has always been hard to enforce because nobody specifies the screen.
The kiosk warning contradicts the refund right
Arizona makes crypto kiosks warn that losses aren’t recoverable. The same statute gives defrauded new customers a full refund.
Explainability ends on the decline screen
The guidance saying a checkbox isn’t a reason was withdrawn in May 2025. Regulation B still lets a denial notice say ‘length of residence.’
Revoking an agent doesn’t cancel what’s pending
Six banks and a bill in Congress agree an agent’s access must be revocable at any time. Neither says what the screen tells you when you use it.
The dispute form has no agent option
Agent purchases get disputed under codes written for stolen cards, and Regulation E’s own definition may defeat the box a person picks.
A name match is not a verdict
Europe just updated its payee name check and the UK wants to keep its own for good. A match confirms the name. Banks are printing verified.
Stablecoin trust is decided at redemption
Congress spent the year fighting over stablecoin rewards. The OCC’s proposed rule already defines the bad day: a 10% run turns two days into seven.
Metered pulls make frequency a consent question
Banks now charge aggregators for each data pull. How often an app checks your account is about to be decided somewhere. It belongs on the consent screen.
Agent payments are a consent problem
Mastercard and Visa are building payment rails for AI agents. Liability follows authorization, so delegation is an interface problem first.
Nothing matches that search.
