Loading page…
PISTA’s small user studies examined active oversight. USC now reports Microsoft funding, while an Excel add-in remains a future goal.
Listen to this story
PISTA, a USC research prototype, is designed to make spreadsheet agents auditable mid-task: it breaks work into reviewable actions and lets users revise an earlier choice on a separate branch, regenerating downstream steps while preserving the original. USC says Microsoft provided $100,000 after early discussions with Sumit Gulwani; an Excel add-in remains a future goal, not a released product. Evidence so far comes from an eight-person initial study and a 16-person evaluation against a baseline, not a workplace deployment.
The authors report that participants found errors they say a review after task completion would have missed, and described a sense of co-ownership over the output.
In USC’s example, adding a missed student-loan-interest deduction creates a new branch and rebuilds later steps without deleting the original.
The paper was submitted to arXiv on April 22, 2026, and reports an initial study with eight participants followed by a 16-person evaluation.
USC researchers have developed a spreadsheet AI agent that lets people interrupt, revise and compare its work before accepting the result. In an October 15 account, the university reported $100,000 in Microsoft funding for PISTA, a research prototype whose proposed Excel add-in remains a future goal rather than a released product.
The researchers’ paper argues that waiting for a finished spreadsheet makes oversight harder. Users cannot inspect assumptions, catch errors before they spread or redirect the agent while it works.
USC professor Rini Chattopadhyay and PhD student Sadra Sabouri developed PISTA. USC says development took three months, beginning in January. The resulting paper, Auditing and Controlling AI Agent Actions in Spreadsheets, was submitted to arXiv on April 22, 2026.
“In spreadsheets, the AI isn’t just generating text—it is making decisions directly in cells that users own and are ultimately responsible for,” USC professor Rini Chattopadhyay said in the university’s October 15 account.
PISTA splits spreadsheet work into smaller actions, such as creating headers or adding formulas. It pauses for approval and supplies a plain-language explanation of what it is doing and why. Users can ask questions during execution instead of waiting until the spreadsheet is finished.
Its branching feature lets a user return to an earlier decision, change it and continue along a different path without overwriting the original. Subsequent steps are regenerated around the correction, leaving both versions available for comparison.
USC illustrates this with a tax spreadsheet: someone who forgot a student loan interest deduction can branch from that point and have later steps rebuilt. PISTA also suggests follow-up questions, including why a particular number was used in a formula.
The April paper describes an initial study with eight participants and a subsequent evaluation with 16. In the latter, participants used both PISTA and a baseline agent. These are small research samples, not evidence from a broad workplace deployment.
The authors report that active participation affected task outcomes, understanding of the task and perceptions of the agent. Participants detected errors that the authors say a review after completion would have missed. They also reported a sense of co-ownership over the output, with their own intentions reflected in the agent’s actions.
USC says Microsoft’s funding followed early discussions with executive and scientist Sumit Gulwani.
Loading discussion...
Join the conversation
Explain where you would want the agent to pause.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.