The chat-box tax
A blank conversation is flexible, but it asks the user to understand the system, assemble context and describe a process from scratch. The result can be impressive for experts and frustrating for everyone else. Repeating the same background information is a sign that the product has not modeled the job.
Good vertical products already know the relevant objects: a claim, customer, ticket, design or pull request. They collect the right context and present choices in the language of the work. Natural language remains useful, but it sits inside a structured workflow rather than replacing one.
From answer to outcome
An assistant creates more value when it can deliver a reviewable artifact or complete a bounded action. That requires tools, identity, permissions and a definition of success. It also requires conventional product details such as progress, undo, error states and history.
Agents do not remove interface design. They make it more important. Users need to see what the system plans to do, what evidence it used and where approval is required. A magical box that hides those details may demo well and operate poorly.
A product test for founders
Describe the recurring job without using the word AI. Identify the inputs, decision, output and person accountable for the result. Then decide which steps need generation, retrieval, deterministic code or human judgment. This prevents the model from becoming the architecture.
The strongest AI companies will not necessarily expose the most capable model. They will make a difficult workflow simpler, safer and easier to measure. Chat can remain the front door, but the product needs rooms behind it.
Sources & further reading
Social-media activity is treated as a signal of attention, not proof. Product claims are attributed to the linked publisher or announcement.