How to build a live, drillable BvA in Excel with AI

Bottom line: yes — with an MCP connection between your FP&A data and Claude for Excel, an agent can roll a budget-versus-actuals report forward a month and draft the variance commentary against real drivers, inside the spreadsheet you already use. The roll-forward is the easy part. What takes iteration is the commentary register and the account mapping, and both are fixable once.

We ran this end to end on a live webinar in August 2026: roll a departmental BvA from March to April, write the commentary, then package the whole thing into a reusable skill. You can watch the full recording here, including the parts that went wrong.

Can AI roll a BvA forward in Excel?

It can, and the mechanics are less impressive than the result. With the spreadsheet connected to a governed data layer, the instruction is close to plain English: roll this report forward to April, then write variance commentary. The agent updates the period, pulls the new actuals, and produces a first pass.

Two things reliably cost time on the first run. The agent spends effort working out which underlying table it should be reading, and it does not always verify that the values actually refreshed before moving on. Both are exactly the kind of friction worth capturing in a skill so the second run does not repeat it.

Why the first commentary draft is always wrong

The first draft is too long. Nobody senior reads eight lines explaining a single variance, and a CFO reads less than that. Ask for concision and the second draft usually overshoots the other way and becomes too thin to act on.

The instruction that lands is specific rather than tonal. Naming what the commentary must contain works better than asking for a length: flag the main vendors driving the variance, state whether each one is favourable or unfavourable, and separate one-off items from run-rate changes. A line like "tech spend came in at twenty thousand against six thousand budgeted, with two unbudgeted one-offs accounting for the entire variance" is the register you want.

Once that register is right, it is worth preserving. That is the whole argument for turning it into a skill rather than re-prompting monthly — see our guide to Claude skills for finance.

What makes a BvA drillable rather than just automated

Automation without a drill path produces commentary you cannot defend in a board meeting. The difference shows up the moment someone asks why.

The drill matters more than the drafting. An explanation you can trace to transactions is a finding; one you cannot is a guess with good grammar. More on that distinction in AI variance detection for FP&A.

The account mapping problem nobody mentions

Here is the part that quietly decides whether this works. Your ledger might carry a hundred accounts. The P&L your CFO actually reads has ten or twelve lines. Something has to hold that mapping.

Left to an agent alone, the mapping ends up in a markdown file that somebody maintains by hand. That works until the chart of accounts changes, and then it silently stops being true — the commentary keeps reading fine while the numbers drift. A governed layer codifies the mapping instead, and flags you when new accounts appear so the mapping gets updated deliberately rather than discovered in a board pack.

This is the same reason a semantic layer matters generally: an agent reading raw contract data has to guess whether you meant start date, close date or renewal date, and it will pick one without telling you. See getting live financial data into Claude and ChatGPT.

What the roll-forward actually looks like

The sequence is worth spelling out, because the useful detail is in the order rather than the prompt wording.

1. The agent identifies the source. It has to work out which underlying table holds the actuals for the report you are pointing at. On a first run this is where a surprising amount of the time goes, and it is the first thing worth hard-coding into a skill.
2. It advances the period. March becomes April across the report's column structure.
3. It pulls the new actuals for the department in scope and lands them against the existing budget lines.
4. It should verify the refresh — and by default it often does not. Ask explicitly for a check that the values changed, or you will occasionally get last month's numbers under this month's heading.
5. It drafts the commentary against whatever variances it now sees.

Steps one and four are the two that justify building a skill rather than re-prompting. They are pure overhead, they recur identically every month, and they are invisible in the output when they go wrong.

What to put in the skill

A skill that makes month two genuinely fast holds five things:

• The source. Name the table or report the actuals come from, so the agent stops rediscovering it.
• The commentary register. Length, whether to name vendors, whether to label variances favourable or unfavourable, and whether to separate one-offs from run-rate.
• A materiality threshold. Below what value or percentage a variance is not worth a sentence. Without this the commentary explains rounding.
• The checks. Confirm the values refreshed; confirm the period is closed.
• What not to do. The corrections you found yourself making twice are the most valuable lines in the file.

That last point is the practical trigger for skillifying anything: the moment you notice you are correcting the same thing you corrected last month, that correction belongs in a skill. Anthropic's announcement of Skills covers the underlying model.

Where this breaks

Four failure modes account for most disappointing results.

Running against an open period. The commentary reads perfectly and the numbers are wrong. Build a close check into the skill rather than relying on the calendar.

Unmapped new accounts. A new GL account appears, nothing in the mapping catches it, and it silently lands outside your reporting lines. Your P&L still foots; it is just missing something.

Commentary drift. Without a stored register, each month's tone and depth depends on how the prompt was phrased that day, which makes month-over-month packs feel inconsistent to whoever reads them.

Explaining the wrong thing. An agent with no materiality threshold will spend three sentences on a two-hundred-dollar variance and one on a material one. See AI variance detection for how to set that boundary.

Which tools support a live drillable BvA

The differentiator is not whether the tool can draft commentary — most now can — but whether the number behind the sentence drills to a transaction.

Capabilities reflect public vendor documentation as of August 2026; confirm current state directly with any vendor.

What you need in place

• A connection, not an export. An MCP link between the spreadsheet and your actuals, so the refresh is a data pull rather than a copy-paste.
• Codified account mapping from GL detail to your reporting lines, with alerts when the chart of accounts changes.
• A drill path from any reported number back to transactions.
• A skill holding the commentary register, so month two does not restart the iteration.

For the Excel-side setup specifically, see using Claude with Excel for FP&A. Anthropic's own agent skills documentation covers how skills are structured.

Build it in Aleph

Aleph keeps the BvA in Excel or Google Sheets over live ERP and CRM data, so rolling a period forward is a refresh rather than a rebuild, every figure drills to source, and the same model filters per department owner. The commentary draft then runs against real drivers instead of a pasted snapshot.

See financial reporting and variance analysis, or start from the six ready-made skills in the Claude skills for finance guide.