Replacing Manual
Tax Work With In-Tool Analysis
Tax users could update transaction data directly, replacing a multi-step download-edit-reupload cycle that had taken 1-2 days. What had been a manual, error-prone reconciliation became a single in-tool edit.
Tax obligation pattern analysis that once required 40-80 hours of spreadsheet compilation was surfaced automatically.
Introduction
When I took over the design, the Use Tax Calculation Tool had already completed its MVP phase and was in the enhancement stage. To identify opportunities, the product team held bi-weekly discovery sessions with stakeholders. During those sessions, we identified two major pain points:
updating transactions data for tax calculations
analyzing tax patterns for long-term strategic planning
To understand each area in depth, I interviewed pilot users with the product team present.
User Insight
Users compiled spreadsheets listing thousands of transaction details, and uploaded them to the tool for tax calculation. Once uploaded, the files were tedious to work with and demanded close concentration to update individual line items. The process typically meant downloading the latest spreadsheet, locating the correct entry, verifying the change, and re-uploading. It was error-prone, and the sheer volume of cell-level data made mistakes hard to spot.
Detecting those errors was itself a manual part of the tax team’s analysis. By examining trends and patterns in tax obligations, the teams could assess accuracy, gaps, exposure, taxability classification, among other financial indicators. But that analysis was constrained by the small team’s bandwidth, since compiling and interpreting the spreadsheet data consumed so much of their time.
The Work
Replacing the Spreadsheet Workflow
I designed an editable table that gave users an alternative to the spreadsheet workflow, letting them manage updates directly within the tool. The mockups gave stakeholders a shared framework for negotiating what could be built immediately versus phased in over time.
Designing Insights Under Constraint
Analyzing tax patterns consumed the largest share of the team’s time, so we scoped generative AI into the workflow to reduce it. My initial direction was conversational: users would query trends and anomalies directly and get answers back in natural language. That direction hit a technical wall. The insight-generation pipeline took 12-24 hours to process, which made live querying unworkable. Rather than force a conversational pattern that couldn’t be supported, I pivoted to a surfaced-insights model: the tool proactively presented trends and anomalies the team would otherwise have compiled by hand.
The tradeoff was that users lost the ability to ask open-ended follow-up questions, which was the part of the original vision they were most excited about. However, the stakeholders agreed the surfaced-insights version was worth shipping anyway, because it removed the highest-cost work, even without the conversational layer. I recommended phasing the conversational capability in later, once engineering had investigated further. Meanwhile, I validated the direction by iterating on the tax team’s feedback across five rounds, until it was the pattern they were satisfied with.
The editable table let users update line items in place, replacing the download-edit-reupload cycle.
Iterations of the insights view, refined across five rounds of feedback.