AI Products Differ Through Checkable Judgment and Delivery
As connecting models and generating interfaces get cheaper, lasting differentiation in AI products comes from turning judgment into executable rules, delivering reliably, and sustaining revenue after subsidies fade.
The presentation layer of AI applications has become very cheap. Being the first to plug in a new model or generate a passable interface is increasingly unlikely to create a lasting difference.
Evaluating an AI product now mainly comes down to three things: whether the team has written its judgment into vocabulary and a do-not list that the model can execute; whether it has turned single-shot generation into stable delivery; and whether revenue remains after resources and subsidies are pulled from the product. All three can be concretely checked. Fast generation does not mean a product holds up.
Access, generation, and delivery are three clocks
An AI product runs on three clocks at once. The first is access, looking at how quickly and cheaply a new model reaches users; the second is generation, measuring how close the default output is to being ready to use as-is; the third is delivery, testing whether failures, adulteration, queuing, and retries are kept under control.
A company can lead on the first clock while losing points on the third. Whenever model capabilities jump, access speed is the most visible, but whether users stay usually depends on delivery.
Vocabulary is part of the product spec
A vague "make it look better" cannot be signed off on. Yet whitespace, rhythm, a sense of quiet, error message phrasing, and whether an agent can smoothly pick up the next step can all be written into explicit rules. Only after these rules are placed into the product's outer harness does output quality start to become repeatable.
AI has raised the floor for design. Above that floor, what is scarce is the vocabulary to state requirements clearly, the restraint to know what not to do, and judgment that someone is willing to be accountable for.
Keep the right to judge on the human side
A long plan from a model can help a team move faster, or it may make the decisions for you.
When the model makes the decisions, the team slowly loses its own record of trade-offs. The next time a similar problem comes up, without that record, the team can only keep accepting the defaults. Over time, the product grows into the model's default shape.
Look at scale and new products separately
Relying on model distribution makes scaling revenue easy; integrate fast, raise money fast, and the growth story holds together. In this process, delivery and gross margin are often left to be verified later.
Evaluating an AI application company requires separating two things: how much retention an older product keeps after resources are pulled away, and how much of the gross margin comes from subsidies and resale spreads.
Exploration teams and delivery teams will clash
Treating an agent as a creator that needs room to perform, versus treating a generator as a production line that must ship on time, corresponds to two different organizations. When these two teams argue inside the same company, the conflict itself is information, showing which clock the company is stuck on.
Where I'm still unsure
Some judgments are not final. When model capabilities jump, access speed itself might decide who captures user mindshare first, while the value of a delivery system may only appear after models slow down and pricing becomes transparent. Specialized vocabulary that pulls output away from the average today may be folded into the defaults of the next generation of models tomorrow, becoming the new baseline. Agent-facing interfaces are becoming more important, yet visual interfaces still matter because humans must understand the process, grant authorization, and take final responsibility.
Questions I want to figure out next
I want to figure out several specific questions next: What is the minimum set of observable items a design do-not list must contain for results to reliably break away from the default generated style? When an agent-facing product description conflicts with a human-facing marketing page, which layer should govern? What externally visible data on pricing, latency, and failure rates can test "model wholesale-to-retail" business models?
References
- *What Happens to Design After AI?*: an a16z podcast episode in which Anish Acharya talks with Microsoft VP of Design John Maeda and Impeccable founder Paul Bakaus, June 24, 2026. The points that a raised design floor leaves vocabulary, restraint, and judgment as the differentiators, and that vocabulary belongs in the harness, come from this conversation.
- *Liblib，你的梦想是什么？* (*What Is Your Dream, Liblib?*): an in-depth report on Liblib / Evoken by Wu Ruirui for elsewhere. The discussion of revenue built on model distribution, the test of delivery and gross margin, and the clash between exploration and delivery teams draws on this report. It includes claims from anonymous sources and investors; figures on funding and margins still need cross-checking against company materials.