Trustworthy AI for Research Budgets
Integrating AI inside the university's research administration system. Deciding what the AI is for, how AI automates, and where humans make the final call.
TL;DR
Problem
A grant manager carries ~50 awards (up to 75+ at peak) and builds each budget twice: once in Excel, once by hand in SAGE, across 8+ disconnected systems, ~2–3 hrs of copying per award. This is not a tooling problem but a coordination gap.
Strategy
Embedded the worksheet inside SAGE and put AI only on the in-between work, gated by a clear AI-vs-human boundary.
Impact
- 2+ hrs saved per award
- 25–33% workload cut (projected 75–150 hrs/GM/yr)
- Trust design validated in moderated usability sessions with GMs
Problem solved
01. The work between the systems had become the work.
A UW grant manager carries ~50 active awards, up to 75+ at peak, and rebuilds each budget twice: ~2–3 hours of hand-copying across 8+ disconnected systems before the work only they can do (rates, exceptions, compliance, sign-off) even begins.
- ~50 active awards per GM on average, up to 75+ at peak
- 8+ disconnected surfaces per award: the NOA PDF, Excel, 5–6 rate sites, Workday, SAGE
- ~2–3 hours of hand-copying per award, before any judgment work begins
How I validated
02. One pattern recurred in all six interviews.
Ran 6 contextual inquiries with grant managers across 6 departments (67+ combined years, 200+ active awards), watching real work (which systems they open, how many tabs stay up, where they copy and paste) over asking them to describe it.
“GMs build every budget twice.”
Built once in Excel for the proposal, again in the portal after the award.
6/6 used Excel as a playground system · 5/6 did manual rate lookups across scattered sources.
The decision that mattered
03. Reframed what the internal admin platform is for.
University of Washington treated SAGE, the admin system, as a system of record, somewhere to store and verify finished budgets. We reframed it as a system of work: the surface where the budget actually gets built. That shift is what makes “bring the worksheet inside SAGE” a structural fix for double-entry, not just a new feature.
✗ rejected · standalone tool
Instead of introducing a standalone tool, I integrated into the existing user experience, to minimize friction while improving efficiency.
✓ selected · worksheet inside SAGE
- Embed the worksheet in SAGE; put AI only on the in-between work.
- Respected GMs' existing Excel-based workflow, using their mental models and minimizing re-training.
✓ selected · AI-powered budgeting with manual adjustment
AI fills the routine lines and does the math; the GM reviews and adjusts before anything is saved. Rates and dollar/percent math run on fixed rules, not AI guesses, and every AI-filled value links to its source so it can be checked.
Where the AI leads, assists, and refuses
- Reads the document and websites
- Auto-fills budget data from the worksheet into the admin system
- Flags mismatches to review, each linked back to its source
- Reviewing and adjusting every AI-filled value before it counts
- The final budget submission
- FCOI sign-off
- Effort certification
How I built · trust by design
04. Made the AI's work easy to check.
Trust fails two ways: distrust the AI and re-key everything by hand, or accept a wrong number without accountability. The AI clears the copying; the GM keeps every judgment call that counts.
AI assists, the human decides.
The GM keeps every final call: review, sign-off, and submit. AI never commits anything on its own.
Turned trust guidelines into real UI.
Applied Microsoft HAX, Google PAIR, and the University of Idaho's AI4RA framework as concrete parts: confidence chips, a source label on every value, and a preview before anything saves.
Kept drafting and checking apart.
You build in the worksheet; you resolve mismatches in reconciliation. Every AI fix is shown and confirmed.
SourceWorkday
“Annual salary 201,912 USD, appointment 0.10 FTE, 9-month basis.”
Matches the source. Safe to accept at a glance.SourceOPB rate table
“FY25 faculty blended fringe rate, effective 07/01/2024.”
⚠ Confirm if the rate's effective date is older than 6 months.
One thing to check first. Confirm the effective date, then accept or edit.Estimated fromsimilar past budgets
“Averaged across 3 prior ARVO trips (2021–2024).”
⚠ Verify before accepting; not from a system of record.
Not from a system of record. Verify the figure; edit is the likely path.▲ hover, tap, or tab to a green-underlined value
Before & after
05. One workspace. Zero round-trips.
Scattered tools get absorbed into SAGE: Excel, the NOA PDF, and the rate websites now live inside one surface. Plays on its own when it scrolls into view; hit ↻ Replay to watch again.
The same work, before and after
Impact delivered
06. What it moved.
Validated in moderated usability sessions with grant managers.
Reclaimed per grant manager
projected · 2+ hrs/award × award volume
Workload cut
estimated across ~50 concurrent awards (P1)
Round-trips per budget
time-to-submission: several hrs → ≤30 min
Returned org-wide
across UW's 1,000+ active grants, at ≤30 min/project
Video
The concept film frames the scale: UW runs 1,000+ active research grants, with grant managers tracking up to 75+ projects each at 2 to 3 hours per budget. Trustworthy AI for Research Budgets cuts that to under 30 minutes per project, returning 3,000+ hours a year across UW's research portfolio.
What I'd do differently
Define where AI belongs before designing where it goes.
I took the sponsor's "automate award setup" mandate as the starting point and narrowed it afterwards, through six contextual inquiries, into the diagnosis that grant managers build every budget twice. The narrowing was right; the order was not. I would map first where AI can genuinely integrate into the existing workflow and why it creates value there, and only then decide what to build. Starting from a mandate means arguing your way back to the problem. Starting from the workflow means the problem is already the brief.
Key takeaway
The hardest part of integrating AI isn't the model performance. It's narrowing what the AI is for.
Responsible AI is calibrated trust: people catch the AI's mistakes without second-guessing what it gets right.
Keep exploring
You just viewed: Trustworthy AI for Research Budgets