Start Safe
Ship deterministic AI before generative AI.
Every AI product I have shipped at scale started with a deterministic layer underneath the model. Math, eligibility, and any number a user is going to trust should be computed by code, not generated by a model. The model's job is to summarize, explain, and route. The deterministic layer's job is to be right. This split is not glamorous, but it is the difference between an AI product users trust and one they screenshot for the wrong reasons.
At Credit Karma, the Debt Agent shipped this exact split. The math behind every payoff plan was deterministic. The model wrapped that math in plain language, prioritized which insight to show, and adapted the tone to the user's situation. We measured a 2.4x lift in engagement on the deterministic plus AI version compared with the math-only version. The lift was not from the AI being clever. It was from the AI explaining the math in a way users actually wanted to read.
Hallucinated math is a CFO problem, not a UX problem. Design for that reality on day one.
Show The Work
Make the AI's reasoning visible and inspectable.
Trust in AI does not come from confidence. It comes from inspectability. Show the user where the answer came from. Cite the source, link to the underlying data, and make it easy to see the inputs that drove the recommendation. Users do not need to read it every time. They need to know they could.
This is also a leadership defense. The first time a regulator, partner, or executive asks you to explain why the model said what it said, having that inspection layer already in the product is what keeps you in business.
A practical pattern that has worked across multiple launches: design every AI response as two layers. The headline answer the user reads first, and a clearly labeled "why" panel one tap away with the source data, the assumptions, and the calculation. The why panel is invisible to ninety percent of users. The ten percent who open it are the ones who decide whether your product can be trusted at scale.
Design For The Failure
Plan for the worst response, not the average one.
Most AI launches die from a long tail of bad responses, not from the average response. Designing for the worst case is what separates a demo from a product.
Concrete example. On one launch, the model returned an answer in the 99th percentile of helpfulness in nine out of ten test cases. The tenth case was wrong in a way that would have triggered a regulatory issue. We delayed the launch by three weeks to add a refusal path and a human review queue for that specific case. The delay felt expensive at the time. The cost of shipping without it would have been an order of magnitude higher.
- Define what the model is allowed to do, and what it must refuse.
- Design the refusal experience as carefully as the success experience. Most teams skip this.
- Add a hard handoff path. A human, a different surface, or a fallback flow when the model is unsure.
- Run evals on the unhappy paths before launch. Average quality is a vanity metric.
Earn The Agent
Do not give the AI agency until you have earned the user's permission.
Agentic products are interesting. They are also the fastest way to lose user trust if you ship them before you have earned the right. Move up the ladder: summarize, then recommend, then suggest the action, then take the action with confirmation, and only then take the action autonomously. Skipping levels almost always backfires. Users tolerate AI that helps them decide. They do not yet tolerate AI that decides for them without consent.
Think of the rungs of the ladder as separate products. Each one earns the right to the next by demonstrating accuracy, transparency, and reversibility. Even when the technical capability is there to take action autonomously, the product call is usually to ship the confirmation step first, watch the data for a quarter, and remove the confirmation only when the user behavior tells you they are bored of it. That patience is the most undervalued asset on an AI product team.
Measure The Right Thing
Trust is the leading indicator. Engagement is the lagging one.
Most AI products are launched with engagement metrics: time in product, messages per session, retention. Those are fine for the first read, but they hide failure modes. A product can have high engagement and rapidly falling trust if users are sending angry messages or screenshotting bad responses. Instrument the trust signals from day one: thumbs down rate, manual override rate, escalation to human rate, and qualitative feedback tagged by failure type.
The teams that ship the best AI products review trust signals weekly, not quarterly. They invest in fast feedback loops between users, designers, and model owners. When something breaks, the fix lands in days, not in the next quarter's roadmap. That speed is the only real moat in AI product design right now.
Privacy And Data
Treat user data like a regulator is watching, because eventually one will be.
The AI products that will survive the next regulatory wave are the ones that designed for compliance as a feature, not as a tax. The teams that treat privacy as a checkbox at the end of the project are the ones that ship slowly, recall publicly, and lose user trust in ways that take years to rebuild.
- Never send personally identifiable information to a model unless your contract and data flow explicitly allow it.
- Default to anonymization, redaction, and on-device processing where the use case allows.
- Get product, legal, and security in the same room before launch, not after.
- Write the data handling story down in a doc you would be comfortable handing to a regulator on day one.
Working With The Org
AI design is a team sport, not a discipline.
The AI products I have shipped that landed well were never built by a design team in isolation. They were built by a triangle: design, ML engineering, and the business owner of the surface. The right artifact at every step is not a Figma file or a model card. It is a shared doc that names what the model is supposed to do, what it must not do, how it will be measured, and how it will be improved. The team that writes that doc together is the team that ships AI users actually trust.