The verification tax: what checking costs, and when it exceeds the gain

Naming the handover point, as the last piece in this series covered, tells you where a human takes over. It does not tell you what taking over actually costs. Every handover has a checking step attached to it, and that checking step has a price in time and attention that almost never appears in the productivity claim that justified automating the task in the first place.

This piece is about that price, the verification tax, and the honest question that follows from it: once you have counted what checking actually costs, is this task still worth automating?

What the verification tax actually is

Checking a tool's output is not free reading. It is reading closely enough to catch an error you cannot predict the shape of, which usually means re-deriving part of the answer, comparing it against a source, or holding the original request in mind while judging whether the output actually satisfies it. That is real cognitive work, distinct from the work of producing the output, and it does not shrink just because the output looks polished.

Why the tax is invisible in most productivity claims

The most-cited study on this trade-off found consultants using an AI tool completed 12.2% more tasks, 25% faster, and rated higher on quality, but on a task deliberately placed outside the tool's competence, the same consultants were 19 percentage points less likely to reach the correct answer than a control group working without it. The speed and quality gains are the number every adoption pitch quotes. The 19-point figure is the number that tells you when the checking step stops being optional and starts being the whole game, and it rarely makes the slide.

What decides whether the tax is worth paying

• The error is cheap to catch and expensive to miss. A wrong number in a footnote is easy to check and cheap if missed; a wrong number in a board report's headline figure is exactly as easy to check and considerably more expensive if missed, which is why the same-looking task can carry a completely different verification price depending on where the error would land.
• The checker has a genuinely faster path to verification than to redoing the work. Checking a citation against a source is quicker than researching the citation from scratch. Checking a subtle judgement about a grant applicant's eligibility can take as long as making the judgement yourself, because there is no shortcut that skips holding the whole case in mind.
• The task sits outside the tool's known competence, which is exactly where the productivity data breaks down. The 19-point gap above was measured precisely on tasks chosen to sit outside the tool's comfort zone, and any task your organisation has not explicitly tested for competence should be assumed to carry that same risk until proven otherwise.
• Someone is actually doing the checking, not skimming it under time pressure. A verification step that exists in the workflow diagram but not in practice is not a control, it is a false sense of one, and the tax is only paid for real when the checker has the time and the standing to say no.

A saving that isn't verified isn't a saving. It's an unpriced risk.

The trade-off in taking verification seriously

Pricing the verification tax honestly is uncomfortable, because it often reveals that a task's net time saving is much smaller than the raw output-speed number suggested, or in the worst case, negative once a competent person's checking time is added back in. The honest answer is that skipping this accounting does not make the cost disappear, it just moves it downstream to whoever discovers the error after it has already caused damage, and that version of the cost is always higher than the version paid upfront by a careful checker.

The practical takeaway

Before crediting a task with an AI productivity gain, price the verification separately: if checking costs nearly as much as the work it saves, the honest net gain is close to zero, and if the task sits outside the tool's demonstrated competence, that checking cost is not optional.

Sources

• Dell'Acqua et al., "Navigating the Jagged Technological Frontier" (Harvard Business School Working Paper 24-013): consultants using the tool were 25% faster and completed 12.2% more tasks, but on a task placed outside the tool's competence were 19 percentage points less likely to reach the correct answer than a control group (n=758).