Part of our series on Claude for finance teams. The surface map this playbook builds on is in Claude Has Four Surfaces.
FP&A has a strange job description: be the most forward-looking function in the company while spending most of your time looking backward. The typical analyst's week is data assembly — exporting actuals, fixing broken links in the planning workbook, chasing department heads for inputs — with analysis squeezed into whatever hours remain. The industry's open secret is that "financial planning and analysis" is mostly financial plumbing and reconciliation.
That's exactly the shape of work Claude is good at — if it can reach your numbers. Here's the playbook.
The setup: Claude needs a live ledger, not a CSV ritual
Everything below assumes one thing: Claude is connected to a real book of record through MCP — either Artifi as your ledger, or Artifi bridged to QuickBooks, Xero, or NetSuite with your data synced into one queryable book. Without that connection, "Claude for FP&A" degenerates into pasting exports into a chat window, which is a demo, not an operating model.
With it, the four surfaces divide the work cleanly:
- Chat — interrogation. "Why is gross margin down 2.1 points in May?" and every follow-up.
- Cowork — assembly. "Build the May management pack" as one delegated, multi-step job.
- Skills and Plugins — your conventions. Report formats, allocation rules, budget structures encoded once, in readable text, so every output follows house style.
- Managed Agents — the calendar. The weekly FP&A snapshot that lands every Friday at 16:00 whether or not anyone remembered.
Scenario 1: Variance analysis as a conversation
The monthly ritual: actuals close, you produce budget-vs-actual, and then the real work starts — explaining the variances. Today that means pivoting, filtering, opening invoices, asking AP what the charge was.
In conversation it looks like:
You: Budget vs actual for May, all departments, flag anything over 10% or €5k. Claude: Three lines flagged. Marketing software +€8.2k over budget — driven by two invoices from a new vendor posted to 6420. Want the documents? You: Yes. And is this recurring or one-off? Claude: Both invoices reference "annual subscription." Treated as one-off in May; if it's annual, accrual treatment would spread €683/month. Want me to propose the reclass entry?
Note the last line: Claude proposes the entry; it posts only when a human signs it. That boundary holds everywhere in this series, and it's load-bearing — variance analysis frequently ends in correcting entries, and corrections in a finance system need a signature.
Drill-down is what makes this work: every number traces to journal lines, every line to its source document. That's a property of the ledger underneath, not of the model.
Scenario 2: Scenario modeling in real Excel
The "what if" requests — what if we hire six engineers in Q3, what if churn doubles, what if we delay the Berlin launch — usually die in a workbook called model_v7_FINAL_jk_edits.xlsx.
With the Claude Excel round-trip, the pattern changes: ask for a scenario, get a real .xlsx artifact built from live actuals — pivot-ready, formula-intact, editable. Adjust assumptions in Excel, push back, ask for a revision. The base data is never stale because it's queried, not pasted.
What Claude is genuinely good at here: building and flexing the model, stress-testing assumptions, narrating the deltas between scenarios. What it doesn't do: know your business's drivers better than you. Garbage assumptions in, confident garbage out — the analyst's judgment is still the product.
Scenario 3: Budget season without the template war
Budget season's worst part isn't analysis — it's orchestration. Templates out, versions back, consolidation, resubmission.
The plugin approach (Artifi ships a budget-planning Plugin): the budget structure, granularity, and rules live as readable text. Building from prior-year actuals plus growth assumptions is one instruction. Mid-year, a forecast that blends YTD actuals with projections for remaining months is another. Variance reviews against either version come with the drill-down from Scenario 1.
/create-budget 2027 from 2026 actuals, +15% revenue, headcount per the hiring plan, monthly granularity
Scenario 4: The Friday snapshot that sends itself
The recurring FP&A artifact — weekly KPI snapshot, monthly flash, cash-adjacent dashboards — is Managed Agent territory: scheduled, running against the live ledger, output in Excel or in your inbox, with a full log of what it did. The difference from "a scheduled report" in a BI tool is that you can reply to it: the snapshot lands, you ask "why did DSO tick up?", and you're in Scenario 1.
What's real today — and what isn't
Real, shipping now: live budget-vs-actual with drill-down to documents; Excel artifacts round-tripped from live actuals; a budget-planning Plugin (create, copy-forward, forecast, variance review); cost-allocation runs by headcount or driver; P&L by dimension (department, project, cost center); multi-entity, multi-currency consolidation underneath all of it; scheduled FP&A snapshots as agents.
Not there yet — and we'd rather tell you: a dedicated driver-based planning cube (think Anaplan-class dependency graphs) — scenarios live in Excel artifacts, which is enough for mid-market and honest about it; native HRIS integrations for headcount planning (headcount comes from your assumptions or your payroll data, not from a BambooHR sync); autonomous forecasting — Claude drafts forecasts and explains them, but no one should let a model commit a number to the board deck without an analyst signing it. We built the sign-off in for a reason.
How FP&A teams adopt this
Almost never by replacing anything on day one. The common path: bridge Artifi to the existing ledger, point Claude at the unified book, and start with the interrogation work — variances, drill-downs, ad-hoc analysis — because it requires zero process change. Excel scenarios come next, then the scheduled snapshot, then budget season runs on the Plugin. Each step is one workflow, reversible, and earns the next.
See the walkthroughs for real runs, or start with why we built it this way.