When AI isn’t enough
300+ backlogged billing packages, a federal mandate, and an AI extraction layer nobody could trust. I led the design of the platform that turned auditors’ work back over to them.
AT&T’s platform and data are proprietary. All screens recreated from the originals — branding removed, data synthetic.
Role
Lead UX Designer, stepped into product ownership.
Scope
GenAI-assisted billing package platform, built for tax and finance auditors.
Outcome
~16 hours of manual work saved per billing package. ~4,800 hours, ~2.5 person-years. Multi-million dollar tax savings.
THE PROBLEM
Search was keyed to the wrong thing
I was brought in for a visual refresh. What I found was a data model built backwards, an AI output nobody could correct, and no approval workflow at all. Getting engineering to agree MVP 0 wasn’t the product was the first problem — the design work came after.
THE APPROACH
Requirements, built one working session at a time
There was no requirements doc to inherit — I built one across a series of working sessions with the supply chain leads: the data model, role-based access, scope, and flow. Engineering pushed back at every stage; I iterated where it helped and held ground where it mattered.
THE PIVOT
Correctable beats impressive
We tried…
The AI scraped source documents straight into a dense cost table — populated, polished, done.
Why it failed…
There was no way for auditors to review, correct, or question what the AI had pulled — an output nobody could put their name behind.
Instead…
Designed for editability with traceability — soft deletes that preserved the original record, and a direct link to the source invoice on every row.
THE SOLUTION
A single page carrying the whole product
The approval chain couldn’t be automated — the next reviewer depends on the package. Admins assigned each package manually through a sequential sign-off, reviewer by reviewer, with a full per-package history logging who changed what, when.
The AI was a capable but fallible extraction layer. The design problem wasn’t making it impressive — it was making its output correctable and auditable, so a human stayed accountable for every number.
THE PROOF
Outcomes
Piloted with 20–30 tax and finance auditors before company-wide rollout.
~16 hours
Of manual work saved per billing package
~4,800 hours
~2.5 person-years saved clearing the 300+ package backlog
15%
Reduction in process overhead at MVP1 pilot
Multi-million
Dollar tax savings for AT&T at full deployment
