Assistants and in-product features

Scope is narrowed deliberately and then stated in the interface rather than hidden behind a general-purpose prompt. Something that answers questions about an order and says plainly that this is what it does will outperform something that appears to answer anything, because the second sets an expectation it then spends every interaction failing.

An escalation path to a person exists in the first version, not a later one. Handing a conversation to a human being, carrying the context so the customer does not repeat themselves, is simultaneously the feature most likely to be deferred and the one most certain to be needed.

Buttons beat a text box wherever the set of useful answers is small. A constrained interface for a constrained task is faster for the user, cheaper to run, and testable in a way free text is not — and a great deal of what gets built as conversation would be better as three options and a link.

What the system can say is separated from what it can do. Reading is one permission surface; writing — issuing a refund, cancelling an order, changing a record — is a different one, and actions in the second category are confirmed explicitly rather than triggered by a sentence that sounded like a request.

Latency is treated as a design constraint, not a metric. People wait differently for a conversational response than for a page load, streaming output changes what a delay feels like, and a response that is thorough and slow frequently loses to one that is adequate and immediate.

Conversation history is a retention decision made deliberately. What is stored, for how long, who can read it, and whether it reaches a vendor are questions with answers that belong in the privacy policy, and transcripts routinely contain more personal detail than the forms the same business is careful about.

In-product features often beat a chat interface outright: a summary appearing where the decision is being made, a draft already filled in, a classification applied to a queue before anyone opens it. The strongest version of this work is usually invisible, and it is a harder sell for exactly that reason.

The measure is task completion, not engagement. Time a customer spends talking to an assistant is a cost they are paying, and a system that holds attention while resolving less is failing at the only thing it was installed to do.