Every Budget, Built Once: Human-in-the-Loop AI for Grant Budgeting

01The problem

Grant managers built every budget twice.

A UW grant manager carries about 50 active awards, and 75 or more at peak. Each budget is built once in Excel and again by hand in SAGE, across 8 or more disconnected systems: about 2 to 3 hours of copying per award before the judgment work begins.

I ran 6 contextual inquiries with grant managers in 6 departments and watched the real work: which systems they opened, how many tabs stayed up, where they copied and pasted. All 6 used Excel as a scratch space, and 5 looked up rates by hand across scattered sources.

02What I decided

SAGE became the place the budget is built.

UW treated SAGE as a system of record, where finished budgets are stored and checked. We made it a system of work: the worksheet moved inside SAGE, in the Excel layout grant managers already knew, and the AI took only the work in between. Then I set where the AI leads, assists and refuses.

1. 1

   AI leads the routine work

   It reads the award letter and the rate sites, fills budget lines into SAGE and flags mismatches, each linked to its source.

2. 2

   AI assists, and the grant manager decides

   Every AI-filled value is reviewed, then accepted or edited, before it counts.

3. 3

   AI refuses anything that carries accountability

   The final submission, conflict-of-interest sign-off and effort certification stay with a person, always.

03How it works

Built, checked and submitted in one workspace.

The AI fills each line with its source and its confidence. Rates and the dollar and percent math run on fixed rules, never on the AI. Mismatches with the award are resolved in a separate reconciliation step, and every fix is previewed before it saves.

Microsoft HAX, Google PAIR and the University of Idaho's AI4RA framework became concrete parts of the screen: a confidence chip and a source on every value, math that opens on a click, and a preview before anything saves.

04Looking back

Define where AI belongs before designing where it goes.

I took the sponsor's mandate, automate award setup, as the starting point and narrowed it afterwards, through the six inquiries, to the finding that grant managers build every budget twice. The narrowing was right, and the order was not. Next time I would first map where AI fits the existing workflow and why it adds value there, then decide what to build.