PwC Built an AI Agent for Enterprise Spreadsheets — Why This Matters More Than You Think

On February 20, 2026, PwC quietly announced something that most people scrolled past: an AI agent that can reason over enterprise-grade spreadsheets. Not summarize them. Not extract a few numbers. Actually reason across millions of cells, dozens of linked workbooks, embedded charts, and buried formulas.

If that doesn't sound exciting, you've never watched a finance team spend three weeks manually cross-referencing Excel files for an audit. This is one of the most practically important AI breakthroughs of the year — and it signals exactly where business automation is heading.

What PwC Actually Built

PwC calls it a "frontier AI agent" — internally referred to as From Rows to Reasoning (FRTR). It's not a chatbot bolted onto Excel. It's an agent that mimics how experienced financial practitioners actually work with spreadsheets: scanning tabs, jumping between sheets, tracing formula dependencies, integrating charts and receipts, and building a coherent understanding of the whole workbook.

The numbers from PwC's internal research are striking:

• 3x higher accuracy than previously published methods on real-world enterprise spreadsheets
• 50% fewer tokens — meaning it's faster and cheaper to run
• Can process up to 30 workbooks containing nearly 4 million cells
• Handles complex cross-sheet formulas, the kind found in deals, supply chain modeling, and healthcare analytics

Matt Wood, PwC's Global and US Commercial Technology & Innovation Officer (previously VP of AI at Amazon Web Services), put it bluntly: traditional AI systems "just kind of shrug and give up" when they encounter enterprise spreadsheets. These aren't your school budget templates — they're financial engines with millions of interconnected cells.

Why Traditional AI Fails at Spreadsheets

Here's the thing most AI demos don't show you. The models behind ChatGPT, Claude, and Gemini are brilliant at text, code, and images. But hand them a 50-tab workbook with 200,000 rows, nested VLOOKUP chains, and conditional formatting logic? They choke.

The reasons are structural:

• Context window limits — A single enterprise workbook can exceed the token capacity of any current model
• Spatial reasoning — Spreadsheets aren't linear text. Cell A47 references Sheet3!B12 which pulls from a named range on Sheet7. Models struggle with this kind of multi-dimensional navigation
• Multimodal complexity — Real workbooks contain charts, images, comments, conditional formatting, and embedded objects alongside raw data
• Precision requirements — In audit and tax work, "close enough" isn't acceptable. You need exact traceability

PwC's solution uses a retrieval-augmented architecture combined with multimodal pattern recognition. Instead of trying to stuff an entire workbook into a context window, the agent dynamically scans, searches, and retrieves relevant sections — the same way a human analyst would flip between tabs and zoom into specific ranges.

Real Use Cases Already in Production

This isn't a research paper or a demo. PwC is already deploying FRTR across its business:

Audit Walkthroughs

Audit teams previously spent weeks manually gathering and validating evidence across complex spreadsheets. Now they upload the files, and the agent automatically maps the structure, extracts relevant data, and runs validation and consistency checks. What took days now takes hours.

Engagement Documentation

PwC teams work with highly structured workbooks documenting business processes, controls, and procedures. Despite being "standardized," these files vary wildly in practice — column names shift, fields appear in different orders, structures change between engagements. The agent handles this variability automatically.

Tax and Advisory Analysis

Complex tax scenarios involving multi-entity structures, cross-border regulations, and interlinked financial models are exactly the kind of spreadsheet work that burns hundreds of billable hours. The agent can trace logic across sheets, explain outcomes, and flag inconsistencies.

Why This Matters Beyond PwC

PwC built this for their own consultants. But the implications are enormous for every business that runs on spreadsheets — which is, frankly, every business.

The Spreadsheet Problem Is Universal

According to a 2023 study by the International Data Corporation (IDC), knowledge workers spend an average of 2.5 hours per day working with spreadsheets. That's roughly 30% of a workday spent navigating cells, formulas, and pivot tables.

For finance teams, it's even worse. Month-end close processes, budget consolidation, variance analysis — these workflows are 80% spreadsheet manipulation and 20% actual decision-making. An AI agent that can handle the manipulation frees humans for the decisions.

AI Agents Are Moving From Chat to Work

This is the real signal. The first wave of AI was chatbots — ask a question, get an answer. The second wave is agents that do actual work. Not just answering "what's the formula for compound interest?" but actually navigating your financial model, tracing the logic, and telling you where the numbers don't add up.

PwC's agent is a Big Four firm saying: the era of AI agents doing real professional work has started. Not in a lab. In production. With client data.

The Consulting Model Is Changing

PwC launched a dedicated tech engineering career track in January 2026 to attract software engineers — not consultants, not accountants, but builders. They're reshaping what a consulting firm even is.

When a Big Four firm invests this heavily in engineering AI agents, smaller firms need to pay attention. The competitive advantage is shifting from "we have more analysts" to "our AI agents can process your data faster and more accurately."

What This Means for Your Business

You don't need to be PwC to benefit from AI agents that understand your business data. The same principles apply at every scale:

• Identify your spreadsheet bottlenecks — Where does your team spend hours on repetitive spreadsheet tasks? Month-end reporting? Inventory reconciliation? Client onboarding data?
• Think agents, not chatbots — The value isn't in asking AI questions about your data. It's in having AI work with your data autonomously — reading, validating, cross-referencing, and flagging issues
• Start with structured workflows — AI agents perform best on tasks with clear inputs, defined steps, and measurable outputs. Audit processes, financial reconciliation, and compliance checks are perfect starting points
• Don't wait for perfection — PwC's agent isn't replacing analysts. It's making them 3-5x more productive. The goal isn't full automation — it's augmentation

The Bigger Picture: AI Is Eating the Boring Work First

There's a pattern emerging in 2026. The most impactful AI deployments aren't the flashy ones — they're the ones tackling the tedious, time-consuming work that nobody wants to do but everybody needs done.

Spreadsheet reasoning. Document processing. Email triage. Calendar management. Invoice matching. Compliance checking. These are the workflows where AI agents deliver immediate, measurable ROI — not because they're impressive demos, but because they save real hours from real people doing real work.

As Matt Wood told Business Insider: the real way to judge a company's AI expertise isn't in the flashy headlines. It's in the "unsexy" work rolling out behind the scenes.

Bottom Line

PwC's spreadsheet AI agent isn't just a Big Four flex. It's proof that AI agents are ready for real enterprise work — not just answering questions, but actually processing, reasoning over, and validating business-critical data.

The businesses that move first on AI agents for their core workflows — the boring, repetitive, spreadsheet-heavy ones — will have a structural advantage that compounds over time. Every hour saved on data wrangling is an hour spent on decisions that actually move the needle.

The question isn't whether AI agents will handle your spreadsheets. It's whether you'll be the one deploying them — or competing against someone who already has.