AI Spreadsheet Workflows Need a Recalculation Handover

The workbook looks finished. That does not mean the numbers updated when the next person opened it.

A generated workbook is not yet a working handover

The dangerous moment in an AI spreadsheet workflow is not always when a formula is written. It can be the moment the file leaves the assistant and somebody else opens it. The workbook may look complete, display plausible totals and contain formulas, yet still fail to update correctly in the application where the business actually uses it. That is an operational handover problem, not simply a question of whether the assistant understands arithmetic.

Consider an illustrative cash forecast for a UK wholesaler. An assistant produces a workbook and a short explanation of the expected cash balance. The finance manager opens the file, changes a customer payment assumption and forwards the summary to an owner. If dependent figures have not recalculated, the explanation and the visible workbook can describe different versions of the business. Neither the professional formatting nor the confident commentary proves that the change reached the final result.

The current Vals AI Excel Modeling Benchmark makes this distinction concrete. Its tasks cover seven financial-model categories and are expected to take human experts at least five hours each to build from scratch. The methodology explicitly expects workbooks to open cleanly and core input changes to propagate through the model. Those are useful acceptance properties, not a promise that any particular assistant is safe for your accounts.

For a business owner, the question is therefore more specific than whether AI can build spreadsheets. Can this workflow deliver a file that the next person can open, recalculate, understand and use without the original assistant present? Define that boundary before automating the email, dashboard export or management-report attachment. Otherwise the automation can make an unfinished calculation look like an approved result.

A plausible total is not proof that the delivered workbook recalculates.

Give calculation and data refresh separate jobs

Two different activities are often bundled into the phrase update the spreadsheet. Recalculation evaluates formulas using the inputs currently available. Data refresh brings new information into the workbook from an external source. A file can recalculate perfectly while still using last week's export. Equally, a new export can arrive while the displayed dependent totals remain stale. Your workflow needs to establish which activity happened, in which order, and against which data.

Microsoft's Excel calculation guidance explains automatic and manual calculation, and warns that changing calculation options in desktop Excel affects all open workbooks. It also describes circumstances where a linked source worksheet has not been completely recalculated. Do not assume a setting visible on one developer's machine will establish the state of a colleague's session.

In practice, assign the import step a clear input contract: source system, reporting period, extraction time and expected totals. Give the calculation step a different contract: supported application, required functions, external dependencies and a completed calculation before export. Then let the publishing step consume only the checked output. An AI assistant can help perform or explain these steps, but its narrative should not replace evidence from the workbook and source systems.

For the wholesaler, this could mean confirming that the debtor export is the intended Monday snapshot, completing the supported refresh procedure, recalculating the workbook and only then producing the owner's PDF. Record the actual sequence with the delivered file. If refresh fails, mark the output as incomplete rather than quietly substituting the last successful result. A useful fallback is an explicitly dated previous report, not an apparently current report built from unknown inputs.

Refreshing the data and recalculating the formulas are different controls.

Test what changes when an input changes

A workbook that passes one static comparison can still be an unreliable operating tool. The most revealing handover check is often a small, deliberate change to a known input. Choose a driver whose consequences the finance owner can predict, then follow the effect through the relevant schedules to the final decision figure. This is not a request for the model to explain its own work. It is an observation of the delivered workbook in use.

Use a disposable copy with synthetic or appropriately protected figures. For example, increase an illustrative debtor-days assumption while leaving revenue unchanged. Ask the finance owner to specify which cash-flow lines should move and which profit measures should not change purely because collection is slower. The exact expected effect depends on the model's definitions, tax treatment and period conventions. Establish those details before interpreting the result, rather than allowing the assistant to invent the acceptance answer afterwards.

Include a switch between the workbook's approved scenarios and a check at a period boundary. A monthly model can behave correctly in November and break when the next financial year begins. Follow formulas across the whole forecast horizon, not just the first populated column. Where an output is intentionally fixed, label it as a snapshot so the recipient does not mistake it for a live calculation.

This approach fits the emphasis in ICAEW's September and October 2026 financial-modelling series: consistent columns, formula consistency, avoidance of hardcoding and explicit checks. Its overview describes how the income statement, balance sheet and cash flow interact. The practical lesson for an AI workflow is straightforward: deliver a working chain of financial relationships, not merely a sheet that matches a screenshot once.

Change a known driver and observe the downstream effect in the delivered file.

Make the recipient environment part of the workflow

A spreadsheet workflow ends in an application, not in a filename. State whether the approved destination is desktop Excel, Excel for the web or another spreadsheet application, and test the important functions there. Do not infer compatibility simply because the file opens. Formula support, external connections, macros and refresh behaviour can differ. A workbook that is suitable for one controlled destination may not be suitable for every person who receives an attachment.

Start with tools already available in the organisation. Excel's calculation controls, formula inspection and existing finance templates may be sufficient. You do not need to build a custom recalculation service merely because an assistant can write automation code. Where a maintained automation route is appropriate, verify its documented support for your workbook features and operating environment. A library that writes a formula into a file should not be assumed to calculate that formula as well.

ICAEW's 8 September 2026 AI and Excel webinar overview describes four areas of assistance: Clean, Create, Code and Customise. It names Microsoft Copilot and external assistants, while stressing validation, auditability and professional scepticism for business-critical work. That is a useful division of labour: AI can accelerate preparation without becoming the authority that certifies the delivered numbers.

In practice, arrange one handover rehearsal using the recipient's normal application and access rights. Have that person open the file, locate the source date, change an approved assumption, complete the calculation procedure and reproduce the published output. Record unsupported features as restrictions. If the intended recipient only needs a fixed report, provide a clearly dated PDF alongside the controlled workbook rather than pretending every recipient needs an editable live model.

The destination application is part of the specification, not an afterthought.

Carry the calculation evidence into the report

Even a correctly calculated workbook can become misleading when the final commentary refers to an earlier state. An assistant may draft its explanation before a refresh finishes, or an export step may read an older copy of the file. The result is a tidy report with inconsistent evidence. Design the reporting handover so that the narrative, charts and figures all derive from the same accepted calculation state.

Give each run a recognisable identity and retain the source snapshot date, workbook version, completion time and reviewer. These need not become a complex platform. A controlled folder, an existing document-management system and a short run record can be enough for a small business. The important property is that somebody can reconstruct which file supported the decision without searching through a conversation history or guessing which attachment was latest.

Only generate final commentary after the calculation and reconciliation checks have passed. Ask the assistant to cite the relevant output cells or named measures where the tool supports that, then compare its headline numbers with the exported report. If a check fails, stop the report handover and preserve the diagnostic evidence. Do not let a downstream email step treat file creation as equivalent to finance approval.

Keep the distinction from a general spreadsheet change review clear. A change review asks whether edits were authorised and logically sound. This handover asks whether the accepted workbook actually calculated and whether its delivered report reflects that state. Both can matter, but they catch different failures. For broader model inspection, ICAEW's earlier June guidance also recommends checking hidden sheets, external links and hard-coded figures. Recalculation does not excuse those checks or prove that the business assumptions are reasonable.

The commentary must describe the same calculation state as the attached report.

Keep the controls proportionate to the decision

The strongest objection is practical: if somebody has to check all of this, where is the time saving? It is a fair question. Requiring a full model audit for every exploratory calculation would be wasteful. The answer is to distinguish a disposable analysis from a recurring workflow that other people rely on, then concentrate the checks on the points where an unnoticed stale result would change a commitment.

For an internal exploration, clear labelling and a quick independent calculation may be enough. For a recurring cash forecast, investment submission or lender report, a controlled calculation and handover procedure is more appropriate. A complex tax, financing or valuation model may require specialist review as well. These are recommended operational controls, not a claim that UK law mandates one particular spreadsheet checklist for every business.

The initial effort should produce something reusable: an approved template, a short set of input-change checks and a named finance owner. Repeat those checks when material formulas, connections, application versions or automation steps change. On ordinary runs, use a smaller set of reconciliations and completion evidence agreed by that owner. Do not ask staff to perform every possible check indefinitely simply because an assistant was involved.

Begin with one report that already has clear sources and a stable definition of success. Measure total preparation, checking and correction time before and after introducing AI assistance. Include the recipient's time spent resolving inconsistencies. If the workflow only moves work from the analyst to the finance manager, the apparent saving is not recovered capacity. If it produces a reproducible, current report with less total handling, expand carefully. The goal is not to eliminate human judgement. It is to make the handover reliable enough that human judgement is applied to the business decision rather than to discovering which numbers actually updated.

Automate preparation, but make completed calculation a condition of handover.