Copilot is only as good as the data underneath it, so before judging the AI, I'd check the files it reads.

Why it matters. FinanceSignal.ai is about turning a business's numbers into clear decision signals. Copilot can speed that up, or add noise that looks like signal. Copilot Chat comes free with Microsoft 365 business plans and the full licence is a paid add-on, but either way your data decides the answer.

Raw output versus Confirmed figure — read Copilot as a signal, confirm the figure.

What it reads

Microsoft's own documentation says Copilot grounds its answers in your organisational data through Microsoft Graph: documents, emails, calendar, chats and meetings. For finance, that's the Excel model, the ledger export on SharePoint, the supplier email thread and the Teams chat where the forecast changed.

The same documentation says Copilot only surfaces data a user already has at least view permission for. That's sound security, but it cuts both ways. If last year's draft budget sits in an over-shared folder, or three versions of the debtors report live side by side, any of them can come back in an answer. Microsoft also notes AI responses aren't guaranteed to be 100% factual.

Readiness is patchy. Gartner surveyed 1,203 data management leaders in July 2024 and found 63% of organisations either don't have, or aren't sure they have, the right data management practices for AI.

Read the answer as a signal

So I treat every Copilot answer as a signal to check against a confirmed figure: the reconciled balance, the closed month, the signed-off forecast, never the conclusion.

A peer-reviewed Harvard and BCG experiment with 758 consultants using GPT-4 (not Copilot) found that on a harder task outside AI's strengths, people with AI were 19 percentage points less likely to reach the right answer than those without it.

Measure it like any other number

Four measures show whether it's working: time per task, error or correction rate, weekly active use, and the share of outputs that tie back to source.

In a randomised field experiment across 66 large firms, run by Microsoft Research with a Harvard Business School co-author, Copilot users spent about two fewer hours a week on email, with no significant change in meeting or document time. In that study, over 90% of people given Copilot tried it, but weekly use settled at just under 40%. Measure, don't ask: in Microsoft's own 2023 studies (vendor research, not peer-reviewed), people guessed Copilot saved them 36 minutes when the measured saving was 12.

What to try

  1. Clean up first: archive old versions, fix over-shared folders and mark the confirmed file for each key number.
  2. Pick two or three tasks built on solid data: month-end commentary from a closed ledger, a meeting recap, supplier email triage.
  3. Time each task for two weeks, then again with Copilot.
  4. Tie every figure in an output back to a confirmed source, and log the share that ties.
  5. Log corrections: how often someone had to fix a number or fact.
  6. Track weekly active use, and keep Copilot only where time drops and the tie-out rate holds.

Soft close

Copilot doesn't create signal. It amplifies what's already in your data. Clean the inputs, read the output as a signal, and confirm the figure before it goes near a decision. For simple measures and tools, start with the Signal Lab at financesignal.ai.

Sources: Microsoft Learn, Data, Privacy, and Security for Microsoft 365 Copilot (now titled for Microsoft Copilot), checked 30 Sep 2026 — learn.microsoft.com/…/microsoft-365-copilot-privacy · Gartner, Lack of AI-Ready Data Puts AI Projects at Risk (press release, survey of 1,203 data management leaders, July 2024), 26 February 2025 — gartner.com/…/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk · Fabrizio Dell'Acqua et al. (Harvard Business School, Wharton, MIT Sloan, Warwick, BCG), Navigating the Jagged Technological Frontier, Organization Science, 2026 — doi.org/10.1287/orsc.2025.21838 (open-access PDF: hbs.edu/…/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_….pdf) · Eleanor Wiske Dillon, Sonia Jaffe, Nicole Immorlica and Christopher T. Stanton (Microsoft Research; Harvard Business School), Shifting Work Patterns with Generative AI, arXiv:2504.11436, v4, November 2025 — arxiv.org/abs/2504.11436 · Alexia Cambon et al., Microsoft Research, Early LLM-based Tools for Enterprise Information Workers Likely Provide Meaningful Boosts to Productivity, MSR-TR-2023-43, December 2023 — microsoft.com/…/early-llm-based-tools-for-enterprise-information-workers-likely-provide-meaningful-boosts-to-productivity