AI for FP&A: a Claude setup for the planning cycle, and where it guesses

Budget season. Forty cost-centre owners, forty spreadsheets, one template that half of them have quietly edited. The first week goes on finding out what everyone actually submitted. That week is where AI pays for itself fastest, and it involves no forecasting at all.

The evidence says to keep a person on the numbers. On SpreadsheetBench V2, 321 end-to-end business workflows including financial models, the best published score is 34.89% (checked October 2026). On Vals AI's Finance Agent test, the leader scores 64.37% (Vals AI, v1.1, June 2026). And in a Datarails survey of 270 US finance leaders, teams reported spending 26% of their working week checking AI output (Datarails, October 2026; a vendor survey).

None of that argues for avoiding AI in FP&A. It argues for using it where the output is easy to check: consolidations that must sum, back-tests with a score, scenario diffs you can read. We'd do it in Claude, with the setup below.

The planning cycle runs on three things: actuals, plan versions and models. Our setup gives Claude read-only access to the first two and puts it inside the third.

1. A project per planning cycle, say "FY27 budget". Upload the plan template, last year's actuals by cost centre and your definitions, and paste the instructions below (what are projects). Use a Team plan for company data (plans).
2. Claude for Excel for the models. Install "Claude for Microsoft 365" from Microsoft AppSource ("Get it now"), or have an admin deploy it from the Microsoft 365 admin centre. Put formats and sign conventions in the add-in's Instructions field, which is separate from the project (Claude for Excel docs).
3. Actuals, read-only. If they live in NetSuite, add Oracle's free AI Connector Service as a custom connector on a view-only, non-administrator role (NetSuite FAQ; custom connectors). If they live in a warehouse, Snowflake's managed MCP server is SELECT-only by default (docs) and BigQuery's execute_sql can be blocked with IAM (docs).
4. Planning platform, if you have one. Pigment offers an MCP server for Claude, and Datarails sells a FinanceOS AI Connector (vendor pages, checked 22 October 2026). Connecting the plan beats exporting it, because the version is then never in doubt.
5. Before a Research run, switch off write tools: Anthropic says Research "can invoke tools from your connectors automatically without further approval".

FP&A rules for [Company]:
- "The plan" = [FY27 budget v3, dated ...] unless I name another.
- Forecast accuracy = [weighted MAPE at product level, 3-month
  horizon]; also report bias.
- ARR = [definition]. Fiscal year starts [1 Feb]; [4-4-5] periods.
- Variances: favourable/adverse by line type, not +/-.
- Do arithmetic in code. Never fill in or change a submitted
  number. Label every input ASSUMPTION or give its source.

If your company runs ChatGPT instead, put the same rules in a ChatGPT project; Oracle says the NetSuite connector supports ChatGPT, and OpenAI lists ChatGPT for Excel on Plus and Pro. The prompts below work unchanged.

Ask an assistant for "forecast accuracy" and it will usually compute 1 minus the mean absolute percentage error (MAPE) and give you a confident percentage. Whether that's the measure your board has been seeing for three years is another matter.

MAPE has known problems. Percentage errors are "infinite or undefined" whenever an actual value is zero, so one product line with a zero month breaks the measure. They also "put a heavier penalty on negative errors than on positive errors" (Hyndman and Athanasopoulos, Forecasting: Principles and Practice, 3rd edition). And because MAPE can exceed 100%, "1 minus MAPE" can go negative, which no one presenting "accuracy" expects. The same book recommends scaled errors such as MASE, which compare your forecast with a naive one.

The other traps we see most:

Put the right-hand column into your Claude project instructions once, as in the block above, and every chat starts from your definitions instead of the internet's.

The useful job is the dull one: reading every submission and finding what doesn't fit.

Attached: [n] cost-centre budget submissions for FY27. Read
FY26 actuals by cost centre and account from NetSuite.

Using code:
1. Consolidate the submissions into one table by cost centre and
   account. Report any file that doesn't follow the template and
   what you did with it.
2. Flag lines where FY27 differs from the FY26 run-rate by more
   than [15%], headcount without matching salary costs, and
   accounts present last year but missing this year.
3. Do not change or fill in any submitted number. Questions go in
   a separate list, addressed to the cost-centre owner.

The check. The consolidated total must equal the sum of the submission totals, file by file. Then open three flagged lines and confirm the flag is right. Read the "files that didn't follow the template" list first: that's where a whole cost centre goes missing.

A forecast from an AI looks exactly as authoritative as one from your best analyst. The difference shows up only against history, so make history the first thing it faces.

From the warehouse: monthly actuals for the last 36 months.
Attached: our current forecast (version dated [date]).

1. Propose a driver tree for [revenue] and calculate each driver's
   history from the actuals.
2. Back-test with code: fit on months 1-30, forecast months 31-36,
   and report MAE and MASE against a seasonal naive forecast
   (same month last year).
3. Produce a rolling forecast for the next 12 months.
4. Write a forecast walk from our current forecast to yours: each
   change, by driver, in currency, adding up exactly to the total
   difference.

The check. Read step 2 first. A MASE above 1 means the model did worse than "same as last year" on your own history; stop there. If it passes, check the walk adds up to the difference between the two forecasts, to the unit. A walk that doesn't sum is hiding a change. The assumptions register in prompts for financial analysis pairs well with this.

The arithmetic of a variance is easy to get right with code. The story is where AI goes wrong, because a plausible story costs it nothing. So split them: first the candidate explanations and the evidence each would need, then a test of each against data you can connect.

For the [Q3 gross margin] variance against budget:
1. Calculate it with code, and split it into price, volume and
   mix using the attached product-level data.
2. For the two largest components, list possible causes. For each
   cause, name the data that would confirm or rule it out.
3. Check each cause against the data available to you (the
   attachments and any connected sources). Mark it CONFIRMED with
   the evidence, RULED OUT, or UNTESTED.
4. Draft commentary using CONFIRMED causes only. List the rest
   for the budget owner.

The check. Open the evidence behind every CONFIRMED cause. Then look at UNTESTED: if it's empty, ask how the model tested causes for which you gave it no data.

Scenario work fails quietly. You ask for a downside with churn 2 points higher, and the model also trims hiring "for consistency". Now your downside has two changes and you only know about one.

On a copy of the model, build three scenarios. Change only these
drivers:
- Downside: [new customer growth -30%, churn +2 pts]
- Upside: [new customer growth +20%]
Everything else stays as in the base case.

Show a table of each scenario's revenue, EBITDA, closing cash and
minimum cash month. Then list every input cell you changed in
each scenario. If any output can't be produced without changing
another input, stop and ask.

The check. Read the list of changed inputs: it should contain exactly the drivers you named. Then confirm the balance sheet still balances and closing cash ties to the cash flow in every scenario. Run this one in Claude for Excel, on a copy: Anthropic says it adjusts assumptions "while keeping formula relationships intact", and its cell citations make the list of changed inputs quick to verify. Anthropic still advises reviewing changes before you finalise (see Claude for finance).

Only 4% of the finance leaders Datarails surveyed trusted AI with the month-end close, and 5% with board-ready reports without review. We think that's about right for now. Use AI to get to a checkable draft faster; don't use it to skip the check.

A two-minute check on any FP&A output

1. Totals first. Does the consolidation sum, the walk add up, the balance sheet balance?
2. Definition. Ask: "Which definition of [metric] did you use, and which alternatives would change the answer?"
3. Source of three numbers. Trace them to the actuals, the plan version, or the cell.
4. Assumptions. Ask: "Which inputs came from you rather than the data?" Every one should already be labelled.
5. History. For any forecast, has it beaten a naive forecast on your own past data?

The hardest of these to do by eye is step 4, because an assumption looks exactly like a fact once it's in a table. Count makes it visible: connect it to Claude or ChatGPT and each claim is labelled as a citation, an assumption, model knowledge or a hallucination, with the reasoning chain showing the query or calculation behind it. For the team-wide setup, see AI for finance teams.

How is AI used in FP&A?

Will AI replace FP&A analysts?

What is the best AI tool for FP&A?

Can ChatGPT or Claude do forecasting?