# John DeSouza

> Director of Product Design with 13 years across consumer lending (Best Egg), blockchain infrastructure (Chainlink Labs) and enterprise identity (Auth0). This site holds his case studies and a blog where each post defends one position with sourced numbers.

Posts are John's own views. Every number in a post links to the primary source that produced it. Quote the claim, not the lede, when summarizing a post.

## Site

- [Home](https://john-desouza.com/): Who John is, the case studies at a glance, how he leads and how to reach him.
- [Work](https://john-desouza.com/work.html): The case studies from Best Egg, Chainlink Labs and Auth0, each with the business result. The pages themselves are password gated.
- [Blog](https://john-desouza.com/blog.html): Positions on design leadership, where product design is heading, and trust in fintech and web3.

## Blog: Design leadership

- [Give your managers something to approve](https://john-desouza.com/blog/give-your-managers-something-to-approve.html): Managers are leaving management because the role’s trade stopped paying, and the proof is that the jump is concentrated in the people furthest from the craft, so preparation and mentorship can’t fix it.
- [The interview was never the expensive part](https://john-desouza.com/blog/interview-was-never-the-expensive-part.html): The argument about AI in hiring is being fought over disclosure and detection, while the measured failure is that most candidates who finish an AI interview never get an answer, which is what happens when you add capacity to the one step that was already cheap.
- [A new role isn’t a new process](https://john-desouza.com/blog/a-new-role-is-not-a-new-process.html): Leaders are answering AI with a change to the roster instead of a change to who approves work, and the collaboration numbers show the roster change doesn’t fix what broke.
- [Getting hands-on costs you the crit](https://john-desouza.com/blog/hands-on-costs-you-the-crit.html): Hands-on time with agents comes out of the feedback hours, and feedback is the only part of the week that holds up as a team gets bigger.
- [Promote on decisions you wrote down](https://john-desouza.com/blog/promote-on-decisions-you-wrote-down.html): A rubric line for AI skills adds a competency to the ladder without replacing the evidence the ladder lost, and the only fix is writing the decision down when it gets made.
- [The bottleneck moved and the budget didn’t](https://john-desouza.com/blog/the-constraint-moved-to-review.html): When agents triple a team’s output the constraint moves to the people who decide whether the work is right, and spending that gain on a deferred hire funds the queue instead of the decision.
- [Count the drift, not the adoption rate](https://john-desouza.com/blog/count-the-drift-not-the-adoption-rate.html): A design system team earns its staffing on the count of what the system failed to prevent, not on an adoption rate or an estimate of hours saved.
- [Train juniors on what they reject](https://john-desouza.com/blog/train-juniors-on-what-they-reject.html): AI makes a junior’s output look senior long before their judgment is, so the unit of training has to move from what they produce to what they reject and why.
- [Your team stopped taking your advice](https://john-desouza.com/blog/your-team-stopped-taking-your-advice.html): You’re no longer the fastest source of useful knowledge on your team, and trying to win that back is the most expensive mistake available to you right now.

## Blog: Where product design is going

- [Designers disagree and the judge decides](https://john-desouza.com/blog/designers-disagree-the-judge-decides.html): A quality rubric for generated screens isn’t a measuring instrument, because trained designers disagree systematically about which screen is better, so whatever the judge rewards is a design decision and it needs a name on it.
- [The agent edits the standard it follows](https://john-desouza.com/blog/the-agent-edits-the-standard.html): The context file an agent reads is a design spec, so a change to it needs a name and a date on it, and the field’s own research on power users shows that change going unread.
- [The summary dropped the user who disagreed](https://john-desouza.com/blog/synthesis-drops-the-disagreement.html): The argument about AI in research is policing the participant and leaving the summary alone, and summarization is the step that drops the one user who disagreed.
- [Generated screens still need a record](https://john-desouza.com/blog/generated-screens-need-a-record.html): When the interface is assembled per person at runtime, the record of what was rendered becomes part of the design deliverable, because support, review and accountability all need two people looking at the same screen.
- [A design review can’t see accessibility](https://john-desouza.com/blog/design-review-cannot-see-accessibility.html): Accessibility defects are invisible to the review designers actually run, so the control has to be a check that runs on every build against rules someone wrote down.
- [Disclose the review, not the tool](https://john-desouza.com/blog/disclose-the-review-not-the-tool.html): Naming the tool isn’t a quality signal, because the same disclosure penalty lands on a master and a beginner. What a client can act on is the record of what was checked and what came back wrong, and no disclosure policy asks anyone to write that down.
- [Prototypes that cannot fail are demos](https://john-desouza.com/blog/prototypes-that-cannot-fail-are-demos.html): The cost of a high fidelity prototype used to force a team to name the question before the build. Now that the cost is gone, the question has to be written down on purpose, or the prototype is just a demo with a review on the calendar.
- [Your design system checks spelling, not judgment](https://john-desouza.com/blog/design-systems-check-spelling-not-judgment.html): Machine-readable design systems enforce the vocabulary and drop the judgment, and the judgment only gets written down if every rejected answer leaves a reason behind.
- [Stop hiring for AI fluency](https://john-desouza.com/blog/stop-hiring-for-ai-fluency.html): AI fluency is a proxy with a shelf life. The durable skill is the ability to specify what good means and then say why the model’s answer missed it.

## Blog: Fintech, web3 and trust

- [A tying ban needs a decline path](https://john-desouza.com/blog/tying-ban-needs-a-decline-path.html): A tying prohibition is enforced against an interface. Whether a customer was required to take a second product is settled by the signup flow, and nobody has written down what a genuinely optional second product looks like on the screen.
- [The kiosk warning contradicts the refund right](https://john-desouza.com/blog/the-warning-contradicts-the-refund.html): The refund right is the only part of crypto kiosk law that returns anyone’s money, and the screen where the loss happens is required to describe that loss as permanent instead of naming the deadline to claim it.
- [Explainability ends on the decline screen](https://john-desouza.com/blog/explainability-ends-at-the-screen.html): Explainability gets built as a property of the model and delivered as a string on a screen, and Regulation B’s floor for that string is a label with no direction and no size.
- [Revoking an agent doesn’t cancel what’s pending](https://john-desouza.com/blog/revoking-is-not-canceling.html): Revoking an agent stops the next authorization. It doesn’t cancel what the agent already set in motion, and the screen where a person turns an agent off is the only place that could tell them the difference.
- [The dispute form has no agent option](https://john-desouza.com/blog/dispute-form-has-no-agent-option.html): A card dispute gets classified by the cardholder at intake, from a fixed list, and that label decides which rules and clocks run. For an agent purchase there’s no true option on the list.
- [A name match is not a verdict](https://john-desouza.com/blog/a-name-match-is-not-a-verdict.html): A payee name check answers one narrow question, and the screen reporting it’s where a bank quietly answers a much bigger one on the payer’s behalf.
- [Stablecoin trust is decided at redemption](https://john-desouza.com/blog/stablecoin-trust-is-decided-at-redemption.html): Rewards decide whether people move money into a stablecoin. Redemption decides whether they trust it, and the draft rules have already written a bad day that the rewards fight ignores.
- [Metered pulls make frequency a consent question](https://john-desouza.com/blog/frequency-is-a-consent-question.html): Once every data pull has a price, how often an app checks a bank account stops being a back-end setting. It’s a consent decision, and it belongs on the consent screen instead of an invoice.
- [Agent payments are a consent problem](https://john-desouza.com/blog/an-agent-with-a-wallet.html): Liability follows authorization. That makes the consent screen the most consequential surface in agentic commerce, and it’s currently the least designed.

## Optional

- [Full text of every post](https://john-desouza.com/llms-full.txt): each post as Markdown with its claim and sources.
- [RSS feed](https://john-desouza.com/feed.xml): the same posts with full text.
- [LinkedIn](https://www.linkedin.com/in/johndesouza-/)