Perception · 2026-09-16 · 1:38

The hidden cost of an unreliable assistant: people stop asking

Name: The hidden cost of an unreliable assistant: people stop asking
Uploaded: 2026-09-16
Duration: 1 min 38 s
Description: Whether you build assistants or buy them, the question is the same: does it save time, or just move the time into checking it? Verification overhead is the real bill.

There is a cost to a slow, unreliable assistant that shows up in no benchmark: once checking its work takes longer than doing the task yourself, people quietly stop asking — and stop trusting the brand behind it.

In favour

A useful design metric: judge an assistant by how much you have to verify it, not only by its capability scores.
Trust is earned by reliability over time, not by announced features.

Worth watching

Capability on paper does not equal adoption in practice — a model can top benchmarks and still be routed around.
The disengagement is silent; there is no error message that says 'the user gave up'.

Our takeWhether you build assistants or buy them, the question is the same: does it save time, or just move the time into checking it? Verification overhead is the real bill.

Source: «The hidden costs of a bad AI assistant» (commentary)

← All Focus posts

AI-narrated · AI-generated content

Transcript

Here is a question worth asking about any assistant. What does a slow, unreliable one really cost? There is a cost here that shows up in no benchmark. You will not find it in a capability score. You find it in what people stop doing. The tipping point is simple. Once checking its work takes longer than doing the task yourself, people quietly stop asking. And they stop trusting the brand behind it. That makes for a useful design metric. Judge an assistant by how much you have to verify it, not only by its capability scores. Count the checking, not just the capability. It also says something about trust. Trust is earned by reliability over time, not by announced features. It builds slowly, one reliable answer at a time. Two things worth watching. Capability on paper does not equal adoption in practice. A model can top benchmarks and still be routed around by the people meant to use it. And the disengagement is silent. There is no error message that says the user gave up. Nothing breaks, nothing alerts. People simply stop asking. So whether you build assistants or buy them, the question is the same. Does it save time, or just move the time into checking it? Verification overhead is the real bill.