Why Your Finance Team Quietly Stopped Using Copilot (If They Ever Even Started)

If you worked in an office between 1997 and 2007, you remember Clippy. Microsoft’s animated paperclip would appear, uninvited, to offer help you hadn’t asked for. Microsoft killed it in 2007. Nearly two decades later, they’re spending $37.5 billion a quarter on something that feels remarkably familiar. This time the paperclip costs $30 a seat, and your team can’t turn it off.

Last week, a senior analyst needed to reconcile 14,000 transactions across three subsidiary ledgers. She typed a prompt into Copilot. It responded with a paragraph explaining how to use VLOOKUP. She closed the pane, pasted the data into a web-based model, and had her match exceptions flagged in four minutes. That scene is playing out across thousands of finance teams. Only 3.3% of Microsoft’s 450 million commercial M365 users actually pay for Copilot.

The structural problems run deep. Copilot in Excel caps at 2 million cells. It only works on files saved to OneDrive or SharePoint, formatted as proper Excel Tables. It can’t read PivotTables or see underlying formulas. If you ask a follow-up question, it loses context and starts over. Your GL extract sits on a shared network drive. Your consolidation workbook has merged cells from a template that predates your tenure. Copilot sees none of it.

Even when the data fits, the reasoning falls short. In independent benchmarks, Copilot scored 67 on reasoning tests while frontier models scored above 117. Analysts who try to build linked forecast models or walk through a five-year DCF describe the same pattern: Copilot times out, produces numbers that don’t tie, or explains how to build the model yourself. Your company is spending $30 per user per month for an AI assistant that tells your team how to do their own job, then watches while they do it.

Wharton professor Ethan Mollick calls the employees who’ve quietly found better tools “secret cyborgs.” They’ve automated their work with capable AI but won’t tell their employers. In finance, the data flowing through those unsanctioned tools includes GL balances, forecast models, board materials, and customer payment records. A recent survey found 59% of employees use unapproved AI at work. Among executives, the number is 93%. Shadow-AI-related breaches cost $670,000 more per incident than those involving sanctioned tools.

Three things to do this week: pull actual Copilot utilization data from your M365 admin center, map the gap between what Copilot handles and what your team actually needs, and surface the cyborgs. If your finance team doesn’t have a governed path to use capable AI tools for complex analytical work, they are already using ungoverned ones. The fix starts with asking the question and making it safe to answer honestly.