Write Your Modelling Conventions Somewhere the Model Can Read Them

Abstract illustration of a rulebook sheet travelling attached to a stack of documents, with edits marked in two distinct colours

Write Your Modelling Conventions Somewhere the Model Can Read Them

Two of the Excel Copilot features that went generally available in June are control artefacts, not productivity features. They have been sitting there for three months.

On 25 June 2026 Microsoft published a piece about Copilot in Excel aimed squarely at finance teams, and with it made a cluster of capabilities generally available across Excel for Web, Windows and Mac: personalisation, workbook rules, pre-built skills, federated data connectors, Plan with Copilot, and Copilot attribution in the Show Changes pane.

Most finance teams skimmed it and moved on, and I understand why. It arrived as a feature list. Feature lists are answers to questions nobody asked, and the connector half of that list — CB Insights and other market-data providers, on top of the LSEG and Moody's connectors released in May — is aimed at investment analysts, not at anyone running a $12 million NDIS provider. If you read the announcement and concluded it was not for you, you were two-thirds right.

The other third matters, because two items on that list are not productivity features at all. They are control artefacts, and they answer a question this blog has been circling for months: how do you tell, six months later, which numbers in a shared model a person produced and which an AI produced?

The two that are controls

Workbook rules. A sheet inside the workbook that captures your structure, naming and formula conventions, and travels with the file. Copilot reads it and works to it.

Copilot attribution in Show Changes. When a collaborator makes a change using Copilot, the entry in the Show Changes pane is now marked as Copilot's, alongside the human edits.

Read those two sentences again as an auditor would rather than as a user would. The first is a written statement of how a model is meant to be built, stored with the model. The second is an edit log that distinguishes machine edits from human ones. In any other part of a finance function we would call those a documented standard and a segregated audit trail, and we would not describe either as a convenience.

3 months
Generally available since late June 2026. This is not news — which is the point. It has been available for a full quarter and is still being treated as a feature rather than a control.
1 sheet
A rules sheet lives in the workbook and travels with it. Unlike a modelling standard in a policy folder, it cannot be separated from the file it governs.

Why "it travels with the file" is the whole idea

Most finance functions have modelling conventions. Inputs are blue. No hard-coded numbers in a calculation row. One row, one formula, all the way across. Assumptions live on one sheet and nowhere else. Prior year never gets overwritten.

And in most of them, the conventions live in one of two places: a document nobody opened after their first week, or somebody's head. When a model moved — to a new analyst, to an auditor, to a funder, to a consultant — the conventions did not move with it. That is why a model that was clean in March is a thicket by November, and why the person who inherits it rebuilds rather than extends it.

A rules sheet inside the workbook changes the failure mode. Not because the technology is clever, but because the artefact is in the right place. It is the difference between a policy and a label on the jar.

And there is a second-order benefit that has nothing to do with AI at all. Writing your conventions down in a form precise enough for a machine to follow forces you to discover that your team does not agree on them. In many teams the conventions are assumed rather than agreed, and that only becomes visible once somebody tries to write them down. The disagreement is worth having once, on purpose, rather than continuously and by accident through inconsistent workbooks.

What belongs in the sheet

Keep it short enough that somebody will actually maintain it. Five sorts of rule cover most of the value:

Structure. Which sheets exist, what each is for, and which direction time runs. Where assumptions are allowed to live, and the statement that they are not allowed anywhere else.

Naming. How periods are labelled, how a scenario is named, the distinction between actual, forecast and budget columns, and how that distinction is shown rather than remembered.

Formula conventions. No constants inside calculation rows. Consistency across a row. Which lookup approach you use and which you have banned. How a division guards against a zero denominator.

Presentation that carries meaning. The input/calculation/output colour convention, units, and sign convention — the one that causes more wasted hours than any other, because half of finance treats costs as negative and half does not.

What must not be touched. Locked historical periods, audited comparatives, anything reconciled to a statutory figure. This is the most important section and the one most likely to be left out.

The other one: knowing who changed what

Plan with Copilot is the companion piece. Before it acts, it can set out which ranges, worksheets, formulas and assumptions it intends to update, with clarifying questions. That is useful, but it is a pre-flight check, and pre-flight checks only help the person watching at the time.

Attribution in Show Changes helps the person who was not there. Six months after the fact, when a figure turns out to be wrong, the question is never "what did the tool propose?" It is "who changed this, and did anyone look at it?" Until now, a Copilot edit in a shared workbook was indistinguishable from a human one after the event. That was a genuine gap, and it is the gap I said in early September an auditor would eventually ask you to close.

On the data: a rules sheet is harmless. The workbook it governs may not be — payroll detail, participant or client data, unpublished results. Before pointing any AI tool at a model containing those, confirm whether the vendor trains its models on customer inputs, and confirm it from your licence terms and tenant settings rather than from a product page. The safest posture is a tool where your data is not retained for training.

What this does not give you

Three honest limits, because a control you have overestimated is worse than one you know is partial.

A rules sheet is guidance to a model, not a lock. It makes compliant behaviour much more likely. It does not make non-compliant behaviour impossible, and it is not a substitute for protected ranges on the cells that must not move. Microsoft's own documentation notes that rules may behave differently across models and over time, and are only fully supported in English.

Attribution tells you an edit came from Copilot. It does not tell you the edit was right, and it does not tell you anybody reviewed it. Who approved the change is still a human record you have to keep.

And none of this is retrospective. It applies from the day you set it up. Every model you already have remains exactly as documented as it was last week.

Three things to do this week

Check whether you actually have it. These capabilities require a Microsoft 365 Copilot licence, and availability, regions and licensing vary. Confirm what your tenant has before planning around it — this takes one conversation and avoids designing a control on top of a feature you do not own.

Write the rules once, for your most important model. Not all of them. The one the board sees, or the one the funder relies on. Half a page. If your team cannot agree on the sign convention in twenty minutes, you have already learned something worth the time.

Decide what your record of AI edits is. Attribution in Show Changes is now part of the answer, and for many small finance functions it can be most of it. But decide deliberately, and write it down, rather than discovering the answer during an audit.

None of this is exciting. Controls rarely are. The reason to do it in a quiet week is that the alternative is doing it in a loud one.

Could you hand your key model to a stranger with its rules attached?

Most organisations have modelling conventions and no artefact that carries them. PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations — including making the way a model is built explicit enough that it survives the person who built it.

Talk to PFL →
Timothy, CPA is Managing Director of Professional Financelink (PFL), providing senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations. 20+ years in finance leadership across NFP, NDIS and SME.

Comments

Popular posts from this blog

Google Gemma 4 Just Launched — And It Might Solve Finance's Biggest AI Privacy Problem

Claude vs Gemini for Australian Finance: An Honest Comparison After 12 Months of Using Both

Why NFP Boards Are Finally Talking About AI — And What the Finance Team Should Do Before They Ask