Insights

Copilot in Excel: better finance workflows still need review

24 September 2026

Explore Microsoft's Copilot in Excel finance updates and the checks teams need for budgets, forecasts, formulas and variance analysis.

Microsoft has announced finance-focused updates to Copilot in Excel, including reusable workflow instructions, additional financial data connectors and features intended to make analysis easier to trace. For Australian business owners, finance managers and analysts, the announcement is relevant to budgeting, forecasting and monthly variance reporting.

The opportunity is to reduce repetitive preparation. The responsibility remains with the finance team: validate formulas, reconcile source data and review the conclusions before using an AI-assisted workbook to make decisions.

What Microsoft announced

In its 25 June 2026 announcement, Microsoft describes several developments for finance users:

  • Reusable skills: instructions for recurring processes, such as refreshing reporting models or preparing variance analysis. Microsoft says teams can define custom skills through a `SKILL.md` file saved in OneDrive.
  • Personalisation and workbook rules: preferences and workbook-level conventions intended to guide structure, naming and formulas.
  • More financial data connectors: additional ways to bring external financial information into Excel, building on previously announced connectors.
  • Greater emphasis on traceability: capabilities intended to help users inspect how financial outputs were produced.

Microsoft also describes internal finance use and evaluation work with the Financial Modeling Institute. That is context for the product’s development, not evidence that a particular business workbook will be correct.

Before planning a rollout, confirm which features are available in your tenant, the relevant licensing and any connector requirements. The announcement includes future distribution options for skills; do not assume every described capability is immediately deployable.

Budgeting: approve the assumptions before generating the model

A budget can look consistent while relying on the wrong drivers. Asking Copilot to build a departmental budget will not resolve whether revenue growth, salary changes or supplier increases have been authorised.

Start with a controlled assumptions table. Record each driver, its source, effective period, owner and approval status. Keep these assumptions separate from calculations so reviewers can see what changes the result.

For example, a staffing budget should distinguish existing employees, approved vacancies and proposed hires. A model that treats all three as committed expenditure could materially distort the budget.

Before accepting an AI-assisted budget, check:

  • The financial year and monthly periods match your reporting calendar.
  • GST treatment is consistent with the report’s purpose.
  • Salary start dates, employer superannuation assumptions and other on-costs are explicit.
  • Annual increases apply from the intended month rather than across the whole year.
  • Department totals reconcile to the consolidated budget without duplication.

Ask for an explanation of the model’s assumptions and inspect the actual cells. A clear narrative is not a substitute for reviewing the formulas underneath it.

Forecasting: separate actuals from estimates

A rolling forecast combines recorded results with assumptions about future trading. Its usefulness depends on keeping those two categories distinct.

Establish a reporting cut-off and identify which periods are closed actuals, provisional actuals and forecasts. Preserve the approved baseline before asking Copilot to assist with updates, so the team can compare the new version with the previous one.

Useful checks include:

  • Timing: does revenue follow the delivery or recognition schedule, rather than the invoice date by default?
  • Cash conversion: are debtor and creditor payment assumptions separate from profit assumptions?
  • One-off items: have unusual costs been excluded from recurring run-rate calculations where appropriate?
  • Scenario control: do base, upside and downside cases use documented changes rather than unexplained adjustments?
  • Period boundaries: are actuals included once, with forecast formulas beginning in the correct month?

Test a change deliberately. If customer receipts move out by one month, does the cash forecast respond in the expected period? This kind of sensitivity check can reveal broken links that a headline total will not expose.

Variance analysis: verify the calculation and the explanation

Copilot may help prepare comparisons and draft commentary, but an observed difference does not establish its cause.

Lower revenue could reflect fewer sales, changed pricing, delayed delivery, currency movements or missing transactions. Without supporting evidence, commentary should identify those as possibilities rather than conclusions.

Define your variance conventions before generating the report. Revenue and expenses may require different favourable or unfavourable labels. Percentage variances also need an agreed treatment when the budget is zero or negative.

A practical review sequence is:

  • Reconcile actuals to the approved source report.
  • Confirm the budget or forecast version used for comparison.
  • Recalculate selected dollar and percentage variances independently.
  • Inspect significant differences by account, department or project.
  • Obtain evidence for the proposed explanation from the responsible owner.
  • Approve the final commentary before distribution.

For example, a lower contractor expense is not necessarily a saving. It may reflect an invoice that has not arrived and requires an accrual. Finance judgement determines the interpretation.

Reconcile data before trusting connected analysis

Additional connectors can reduce manual data handling, but a connection does not establish that information is complete, current or suitable for the task. External market data also does not replace reconciliation of internal accounting records.

For each input, document the source system, extraction time, reporting period, currency and filters. Record whether values are transaction-level details, summarised balances or estimates.

Use control totals appropriate to the report. A profit and loss analysis should reconcile to the approved ledger report. A receivables forecast should reconcile to the debtor balance, with reconciling items explained. Check for duplicate rows, unmapped accounts, missing entities and inconsistent date formats.

Where external information informs a forecast, identify the provider and observation date. Confirm that the data’s geography, currency and definition match the assumption being modelled.

Validate formulas and retain an approval trail

Treat AI-generated workbook changes like changes made by a new analyst: review them before relying on them.

Check formula ranges, absolute and relative references, lookup behaviour, hidden rows, error handling and hard-coded values. Test representative cells, subtotals and boundary cases, including blank inputs and zero denominators. Compare important outputs with an independently calculated result.

Reusable skills and workbook rules can document your preferred process, but they are not substitutes for enforced access controls or approval. Give recurring instructions an owner and version, and test changes on a copy of a known workbook.

Retain the source extracts, assumptions, reviewed workbook and approver’s record together. Restrict access to payroll, customer and commercially sensitive information through appropriate Microsoft 365 permissions. Confirm the applicable data handling and residency arrangements rather than assuming Copilot is a private Australian-hosted service.

Start with one controlled monthly report

Choose a recurring report with stable inputs and a clear reviewer. Establish its control totals, formula tests and sign-off criteria before trialling Copilot. Evaluate whether preparation becomes easier without creating additional reconciliation work or weakening review.

For workflows beyond Excel, Tech Engine’s Financial and Management Reporting Agent can ingest authorised financial and operational information to prepare draft reports and commentary. Tech Engine’s managed AI agents run in a private Australian-hosted environment, integrate with existing systems and retain human approval. Our guide to invoice processing and human approval explains a related finance control approach.

Tech Engine can help configure, secure and manage Microsoft 365, including assessing Copilot readiness and finance workflow controls. Contact sales@techengine.au or call 1300 088 324 to discuss a controlled starting point.

Want this applied to your business?

Request an AI Blueprint and we will map the processes worth automating first.

Request an AI Blueprint