PwC Built an AI Agent for Spreadsheets. Here's Why That Matters.
When Matt Wood joined PwC in 2024 after a stint as vice president of AI at Amazon Web Services, he noticed something unusual: every screen in the office was buried under spreadsheets. Not the kind youd use for a soccer team budget, he told Business Insider. These were "financial engines" — massive, interlinked workbooks with millions of cells, dozens of tabs, embedded charts, receipts, and formulas that span across files.
The irony was striking. While the AI world obsessed over flashy demonstrations — generating images, writing code, holding conversations — one of the most critical pieces of enterprise infrastructure remained stubbornly unreachable.
"They just kind of shrug and give up," Wood said of conventional AI systems when faced with enterprise spreadsheets.
So PwCs engineers built something different.
The Unsexy Problem Worth Solving
The breakthrough, announced this week, is a "frontier AI agent" capable of reasoning across massive, multi-sheet workbooks — the kind that power deals, supply chain modeling, healthcare analytics, and audit walkthroughs at the worlds largest organizations.
The numbers tell the story: the agent delivers 3x higher accuracy on real-world enterprise spreadsheets while using 50% fewer tokens. It can analyze up to 30 workbooks containing nearly four million cells — the scale actually found in corporate finance, not laboratory conditions.
"Were talking about spreadsheets that are more like financial engines than spreadsheets," Wood said.
How It Works
Unlike traditional AI approaches that dump an entire spreadsheet into context, PwCs agent mimics how experienced practitioners work: scanning, searching, jumping across tabs, integrating charts and receipts, and reasoning through complex dependencies.
The system combines multimodal pattern recognition with a retrieval-augmented architecture — essentially teaching the AI to navigate spreadsheets the way a human accountant would, rather than treating them as flat text.
For audit teams that once spent weeks manually gathering and validating evidence across countless complex files, the difference is dramatic. Upload the files, and the agent maps the structure, extracts relevant data, and performs validation — tasks that would otherwise require combing through millions of rows by hand.
The Bigger Picture
This matters beyond spreadsheets. PwCs push into engineering talent represents a broader shift among the Big Four. The firm launched a dedicated tech engineering career track in January — previously, it offered only consulting and accounting paths.
The numbers are staggering across the industry:
- Accenture: Added nearly 40,000 AI and data professionals in the last two years (roughly 10% of global headcount)
- EY: Added 61,000 technologists since 2023
Wood described the work as having two modes: "transforming today" (improving current workflows) and "building for tomorrow" (reimagining professional services from scratch in an AI agent world).
The spreadsheet agent is firmly in the first category. But it demonstrates something important: the real AI divide isnt between companies with flashy demos and those without. Its between organizations willing to tackle the unglamorous, foundational problems — and those waiting for someone else to solve them first.
The spreadsheet was never sexy. But it might be the most important AI problem nobody was paying attention to.