Before and after: one grant report with AI, honestly measured

This series has spent six pieces building a vocabulary: task, not job. Handover points. The verification tax. Prompt patterns for real documents. Context engineering. The four checks before anything leaves the building. Vocabulary is only useful if it survives contact with one ordinary task, so this piece walks through a single, real one: drafting a mid-sized foundation grant report, before and after redesign, honestly measured.

Honestly measured means net of the checking time, not just the headline drafting-time saved. A number that only counts the fast part is not a measurement, it is a sales pitch.

The task before redesign

An experienced fundraising officer drafting this report by hand spends roughly six hours: pulling figures from three separate systems, writing the narrative by hand, checking her own numbers as she goes, and reformatting everything into the funder's specific template. The whole job sits with one person from first figure to final sign-off.

The task after redesign

Applying the task decomposition this series covered earlier, the report splits into distinguishable sub-tasks with different risk profiles: gathering and structuring the figures, drafting the narrative, framing the one paragraph that explains a missed target, and final sign-off. The first two moved to AI assistance; the last two stayed entirely human, which is the handover point in practice, not just in theory. Getting there took real context engineering: feeding the model the funder's actual template and the previous year's report structure, not a generic instruction, and naming what to ignore, this funder's boilerplate legal text, up front.

What the four checks actually cost, on this report

• Checking the source took fifteen minutes. Every figure pulled into the draft was traced back to the finance system it came from, catching nothing this time, which is itself useful information.
• Checking the boundary case took twenty minutes and caught a real error. One grant line had changed structure mid-year, and the drafted figure had quietly applied the old category.
• Checking the judgement call took twenty-five minutes and needed a genuine rewrite. The paragraph framing a missed programme target was drafted competently but too blandly; the programme lead rewrote it herself, because that framing was never delegable in the first place.
• Checking what was missing took ten minutes and caught a missing appendix. The funder's own report checklist named a data table the draft had silently omitted.

The honest number was two hours saved, not six, and two hours saved is still worth having.

The trade-off in reporting the honest number

A two-hour saving looks unimpressive next to a vendor claim of a report drafted "in minutes," and reporting the smaller number risks making the redesign look like it barely worked. The honest answer is that the smaller number is the only one that survives someone else trying to reproduce it: the vendor number never counted the checking time, and the first time a fundraising lead discovers that gap for herself, every future claim from that source loses credibility along with it.

The practical takeaway

When you measure a redesigned task, net the checking time against the drafting time saved and report the smaller, real number rather than the larger, unverified one; two genuine hours are worth more to your case for adoption than six hours nobody can actually reproduce.

No external statistic cited; this article presents a worked internal example rather than third-party evidence.