Checking It Is Quick. Explaining It Is Not.

A client replies to your report with one line. Can you walk me through how you got to this number? The number is right. You know it is right. And you still lose most of a morning putting the reasoning back together, because whatever produced it may have assembled an answer without preserving enough of the reasoning behind it. There is no trail to follow. A question that took eight seconds to ask costs you a morning to answer.

This is the part that got skipped when the tools came in. The test many teams applied was whether the output was correct, and in some workflows, correctness turns out to be the cheaper part. Checking is fast when you can hold the answer next to a source and see that it matches. Explaining is slower, because explaining means reconstructing the path: what went in, what was assumed, what was ruled out, and why this figure rather than the other one. Surveyed finance professionals have reported that time saved by AI is partly canceled out by the time spent verifying and explaining it.

Not all of your work carries that risk. Plenty of it never leaves the room. An internal summary, a first draft, a shortlist someone else will sift through, those are far less likely to be interrogated later. But some work is essentially an explanation with a number attached. A fee calculation, a valuation, a claim, a scope, anything a client or an insurer or an auditor can come back to in six months. Those are two different jobs, and a lot of businesses have pointed the same setup at both without noticing.

The instinct, once this bites, is more review. Add a second pair of eyes before anything goes out. That can help with accuracy, but it may do very little for the situation above, because review confirms the answer is right today without leaving behind anything you can hand to someone next spring. This is a design decision, and it sits earlier than most people put it. It belongs in the same conversation as what technology should carry and what stays with a person. If an output will have to be defended, it should be built to produce its working as it goes. Capturing the reasoning at the point it exists can be far cheaper than reconstructing it months later, which is the same problem as work that was never written down in the first place.

There is a useful sorting exercise in here too. Some of what you want carried is a rule, and rules can often be documented clearly enough that someone else can follow the logic later, which is worth separating out from the judgment calls you were actually paying someone to make. Pick three things your business sent out last month. For each one, ask who would have to answer if a client challenged it in six months, and how long it would take them. If the answer is you, and the answer is a morning, you know which work needs to start showing its reasoning.