Arushi Tandon
AI-first product designer
A1
Arushi Tandon
AI-first product designer
Senior Product Designer, Microsoft. I lead design for Excel, making complex work feel simple, intuitive, and increasingly agent-powered. Sixteen years: retail display designer, author, founder, enterprise product designer, design lead, and builder — HAWK AI, a funded Microsoft product, was designed and coded by her. Design lead for Excel on mobile and iPad. At Microsoft since November 2021, promoted twice. Gurgaon, India.
Spreadsheets on mobile were built for reading, not doing: people could check numbers and make light edits, but real work waited for the desktop. On iPhone, the jobs are triage in mobile moments, prepping on the go, fire and return without losing context, and fixing issues without a desktop. On iPad, the jobs are desktop parity, complex analysis, confident work on complex sheets, and Agent for desktop-heavy workflows. The moments that matter in a spreadsheet happen away from the laptop. We shipped the Excel Agent on iPad and iPhone so people can carry it, and new users can start on the device in their hand. Every artefact is agent state made legible at a moment of decision: clarifying questions before it acts, chain of thought while it works, change highlights and selective undo after, fire-and-return for the people who switch apps. Trust scales with risk. Mobile becomes a command centre, a true companion to the desk. The honest limit: the agent takes longer than a typical mobile session, and we say so. Design lead, with three designers to my direction.
I'm moving the team from designing in design tools to designing, auditing and fixing directly in code. HAWK AI, the Experience Analyser, replaces a 1.5-month, 8-vendor manual testing cycle: 4–5 days saved per audit, 7,000+ videos analysed, 25+ critical issues caught in its first run, six specialised AI agents. The Microanalyzer measures flickers frame by frame. PRISM carries design into code and files a validated handoff with the work item.
Finding UX defects meant people watching recordings for four to five days and still missing the one-frame flicker. I built HAWK AI, an internal Microsoft tool that watches recordings and files what it finds with visual evidence and a severity. It became a funded product with engineers scaling it, and grew into a factory: the analysers, a micro-analyser that catches flickers a human eye skips, a bug fixer where designers fix defects in code, and PRISM, where agentic design and engineering work together. An eval system weights bug-finding accuracy 40, triagability 30, hallucination control 30. On the previewer it is right about half the time, and that number went in the leadership deck. Word and PowerPoint are next.
Excel did not support interactive charts, and Copilot could draw a chart nobody could question. Ten interviews across three regions: every participant called interactivity a basic need, and four of our hypotheses were overturned, including mine. Consumption dictates creation. So the chart became the instrument: Copilot delivering charts with tooltips, series toggling, zoom and drill-down for makers and takers on one surface.
An analysis takes multiple steps, tools and people, and cell after cell of numbers does not tell a story. At a Microsoft hackathon I reimagined Excel as a Copilot-first data hub with Python underneath: one prompt becomes a plan, Copilot gathers internal and external data, runs each step with a preview, writes the report, and the reader can question it in place. A patent application was filed in May 2025.
Every app an organisation runs has to be deployed to the right people, shown to them in the right state, and governed by default. I owned that system end to end: central deployment across host products, bundles, usage, the Deployed / Available / Blocked model, the two admin centres synced, and default-state controls so admins could set security and trust once. When Copilot arrived the model held: a review desk where admins see an agent's capabilities, knowledge and actions before they publish or reject it, designed in 2023 before autonomous agents reached public preview. A usability study with admins at six global enterprises found they navigated by search while state lived in tabs; two of three recommendations shipped.