This year a long-running industry report put a name to the shift. John Maeda's Design in Tech Report 2026 framed it as the move from UX to AX, user experience to agentic experience. The argument is simple and consequential. For most of software's life the design question was how to help a person do a thing. Now that agents can take the action themselves, the question becomes how to help a person know whether the thing was done well. The same year's design surveys show the field adopting AI almost universally while trusting its judgment far less than a human's. Designers reached for the tools fast and stayed skeptical of the output, which is exactly the posture every ordinary user is about to take toward every AI product a business ships.
Capability has stopped being scarce. The top models are close, rented from the same short list, and able to do the visible task in a demo without trouble. So a working demo no longer earns anything, because everyone has one. What remains scarce is the willingness of a real person to let the system act without watching it. That willingness is trust, and it does not come from the model being more capable. A more capable model that gives a person no way to see, doubt, or reverse what it did is not more trusted. It is more alarming, because it does more, faster, with less visible accounting.
Trust in an AI product is not a mood. It is built from specific, drawable decisions. The product shows its work, so a person can see what was used and why. It signals confidence honestly, so a person knows which outputs to check. It marks the boundary it will not cross without asking, so nothing irreversible happens by surprise. It makes correction cheap, so a mistake is an undo rather than a phone call. Each of these is a design decision made before a model is ever chosen, and together they are the surface where a person decides whether to rely on the system. The studio has called the agent a contract. This is the contract made visible, the place the user reads the terms.
Capability is rented and roughly the same for everyone. The surface where a person decides to trust the work is designed, and it is the only part a competitor cannot buy.
The common failure is to spend all the design effort on the action and none on the account of it. A team builds an agent that drafts the reply, books the appointment, reconciles the invoice, and ships it the moment the action works. The action working is the demo. What they skipped is the harder half: the interface that lets a person confirm the draft is right, the readout that flags the one reconciliation in fifty that looks off, the clear path to undo the appointment booked into the wrong slot. Without that half, the first visible mistake ends the relationship, because the person had no way to catch it coming and no fast way to take it back. The capability was never the problem. The absence of a trust surface was.
The shift from doing to verifying changes what a product is for. When a person did the work themselves, they trusted it because they did it. When an agent does the work, that built-in trust is gone, and the product has to manufacture it deliberately, at the exact volume the agent operates. An assistant that handles ten tasks an hour needs a way for a person to certify ten outcomes an hour without reading every one in full. That is a design problem with real craft in it: summarizing what was done, surfacing only what deserves a second look, making approval and reversal fast enough to keep up. Get it right and a person supervises a stream of actions with confidence. Get it wrong and they either rubber-stamp blindly or abandon the tool. The design of that verification loop is now the product.
So the instruction for a business this year is to stop buying AI on the strength of what it can do and start judging it on how well it lets a person trust what it did. The model is rented and roughly the same as a competitor's. The trust surface is yours, composed from your understanding of the work, your judgment about what must defer to a person, your decisions about what good looks like. It is the part a competitor cannot rent off the same price list, because it is not a capability. It is a drawing. Treat the agent as a contract and the verification surface as where that contract is read, and trust stops being a thing you hope to earn after launch. It becomes the thing you designed in before you shipped.
Field note prompted by John Maeda's Design in Tech Report 2026 and its framing of the year's shift from user experience to agentic experience, from UX to AX, alongside the year's design surveys showing near-universal AI adoption among designers and persistent distrust of AI judgment. The principle outlasts the report: capability is rented, the trust surface is designed.