Before AI Edits the Forecast, Give the Workbook a Rulebook

The variance pack is due Monday morning, and it's 4:50 on Friday. The workbook has forty-one tabs, three of them named some version of "old_v2," and an AI assistant that now lives inside the file itself instead of a side chat window. Ask it to refresh the variance analysis against actuals, reformat the summary tab to match the board template, and note what changed, and it hands back something that looks finished in under two minutes.

Nobody wrote down the rules it just followed, though. Whether "variance" here means budget-to-actual or forecast-to-actual. Whether the board template wants basis points or percentage points. Whether the connected source is this month's general ledger pull or last month's export somebody forgot to refresh. Those rules exist. They live in a Slack thread from March, in an analyst's memory, in a footnote three tabs deep in last year's file. The AI didn't skip them. It couldn't find them, so it filled the gap with a plausible guess, and the guess was fluent enough that nobody caught it before the numbers reached the deck.

That's the gap Microsoft just tried to close directly inside Excel, and closing it is a bigger story than another AI feature drop.

What actually changed

On June 25, Microsoft's Brian Jones announced a set of Copilot in Excel features built specifically for finance teams, in a post titled "Copilot in Excel: Built for the era of Frontier Finance." His framing was blunt about why finance is a different surface than the rest of Copilot: the software has to show its work, and every number needs a way back to where it came from. He put it more plainly on X: "Finance teams hold software to a high bar. Every number has to trace back, every change has to be explainable, and the data has to come from a source you trust."

Four features carry that argument.

Skills teach Copilot to run a repeatable process the same way every time: building a DCF, closing the books, refreshing a monthly reporting model, preparing a variance analysis. Microsoft ships a sample finance skills library, and teams can also write their own, using an open-standard markdown file, SKILL.md, saved to a folder in OneDrive. Per Microsoft's own documentation, "a skill extends the capabilities of an AI model. Skills provide Copilot in Excel with information to complete repeatable tasks." You invoke a skill with an @mention, and switching one off only stops it from firing automatically. It's still callable by name.

Financial connectors, built with partners including LSEG, Ramp, Rogo, Samaya AI, Velixo, and Vena, keep Copilot pulling from sources finance teams already trust instead of whatever data happens to be open in another tab. And traceability, surfaced through Show Changes-style attribution, answers the question a reviewer always asks first: what did the AI touch, and why.

Then there's workbook rules, which capture the structure, naming, and formula conventions of one specific file as a worksheet inside that file. According to Microsoft's support documentation, rules live on a worksheet titled .Rules, they have to stay visible to take effect, and they travel with the workbook. Share the file, and the rules ship with it. That last part is the one worth sitting with.

Microsoft says its own finance organization pressure-tested all of this across FP&A, accounting, tax, compliance, and treasury before shipping it, which is itself a small confession about how much can go sideways when nobody tests it that hard.

The four features are answering one question

Each of those four pieces solves the same underlying problem from a different angle. Copilot in Excel isn't a chat window that suggests a formula anymore. It can read a workbook, follow a repeatable process inside it, and write the result back. That makes it a workflow runner, not merely an assistant. And a workflow runner needs exactly what any other automation needs before it gets write access: operating instructions that are visible, reviewable, and impossible to misplace, instead of instructions scattered across the tools people used to compensate for the AI's absence.

That's what "rules travel with the workbook" actually means. It isn't a formatting nicety. It's an admission that instructions kept anywhere else (a chat history, a prompt someone typed once, an analyst's memory) don't survive the moment the file gets forwarded, duplicated, or opened by someone new. Microsoft's own guidance even warns that "Copilot in Excel and available models are evolving rapidly. Rules may behave differently across models and over time as Copilot in Excel features expand." If a workbook's AI behavior can shift under it without anyone touching a formula, the rules and the review trail stop being optional polish. They're the only thing holding the process still.

It's the same lesson BaristaLabs has been tracking since the first wave of spreadsheet AI showed up in Excel: a model that can edit a live file is only as trustworthy as what it's allowed to see and how visibly it shows its work. It's also the finance-specific version of a pattern we've written about before: agents that need a map of the buried context behind a decision before they're allowed to touch it. The difference here is that the map now has a literal address, a worksheet in the file, not a policy doc somewhere else.

The workbook AI control packet

Before handing Copilot write access to a workbook that feeds a board deck, a lender, or an audit, it's worth building one small, unglamorous artifact for that specific file: not a company-wide AI policy, just the operating rules for this workbook. Call it a workbook AI control packet. It fits on one page and answers nine questions.

Scroll sideways to see all 2 columns.

Field	What goes in it
Owner	The person accountable for this workbook's numbers — not whoever uses Copilot most
Workflow	The exact process being automated: "refresh Q2 variance vs. budget," not "help with finance"
Allowed data	The specific connected sources Copilot may pull from, named — not "the usual sources"
Disallowed assumptions	The defaults Copilot must never apply on its own: a rounding convention, a fiscal calendar, a currency
.Rules sheet	Visible, current, and written the way Microsoft's own guidance recommends — one rule per cell, grouped under headings like ## NUMBERS or ## CHARTS, with a short example
Skill definition	The SKILL.md file, if one exists, and where it lives in OneDrive
Review owner	Who signs off on AI-touched cells before the file leaves the team
Change trace	Where Show Changes output gets checked, and by whom, before the workbook is shared
Stop condition	The specific case where Copilot has to stop and ask — a missing connector, a rule conflict, a number that moves past a set threshold

None of these fields require a platform decision or a governance committee. They require one person to sit with one workbook and write down what has, until now, lived in that person's head. This packet is a smaller, workbook-scoped cousin of the AI workflow controls worth using anywhere an AI assistant gets write access, and the change-trace field does the same job as the state ledger any AI agent should carry before it acts on live data.

Start with the workbook that's already a mess

Pick one workbook, the messiest one that already matters, not a test file nobody depends on. Write the .Rules sheet first, in plain sight, before turning Copilot loose on anything in that file. If a rule can't survive being written in one cell with a short example, it wasn't a rule. It was a habit.

Build one skill for the single most repeated task in that workbook, not five skills for five workflows that aren't automated yet. Connect only the data sources named in the control packet; anything else Copilot might reach for stays disconnected until someone approves it. Require a named review owner to check Show Changes output before the file moves anywhere the analyst doesn't control. Then re-test the whole packet after any model or feature update. Microsoft's own warning about rules behaving differently across models isn't a footnote. It's a maintenance schedule.

This is the kind of narrow, workbook-specific mapping that process automation work should start with: one file, one owner, one visible set of rules, before the second and third workbook get the same treatment.

Do this for one workbook before scaling to ten. The packet takes an afternoon to fill out. What it prevents (a fluent, wrong number reaching a board deck because nobody wrote down what "trusted data" meant for that file) takes a lot longer to walk back.

Map one workflow before Copilot gets write access

Copilot in Excel and tools like it are going to keep moving from suggestion to execution inside the files finance already runs on. That's genuinely useful. It's also exactly why the boring paperwork matters more than the model version behind it. Map one spreadsheet workflow — the rules, the trusted sources, the review owner — before giving AI write access to it. Start with one workbook.