Questions to ask an agentic AI FP&A vendor, and my answers

Vantage is the orchestrated agentic AI solution for finance, and these are the eight questions I would put to any vendor in this category, Vantage included. A separate checklist for any supplier is What to ask an FP&A supplier before you trust its numbers.

Where do the figures come from, and can I trace one to a source cell?

From your own exports. In a diligence package, each figure in the written memo carries a tab and cell reference into the workbook, and the structural identity checks are live formulas, so you can click into a check and trace it. Bring a figure you already know the answer to and I will walk the trace with you on a call.

How is the output checked, and by whom?

Two layers. A deterministic gate runs on the build, and every check reports PASS, FAIL or SKIP with a stated reason. A warning prints alongside a check's own result and does not by itself stop the package. A check that reports nothing at all is itself a failure that stops the run. The second layer is an independent review by a reviewer that did not build the package and cannot edit it, and a person dispatches that review. The gate tests structure, arithmetic, internal consistency and stated basis.

What happens when data is missing or does not agree?

Where neither test basis is available for a period in a diligence package, the red flags section reports that period as untested, so an empty section does not read as a clean bill. A forecast line whose data does not meet its method's precondition falls back on the sheet. In a diligence package, the structural identities are recomputed independently each period, and a break fails the package at the gate.

Can I see the workings behind a variance explanation?

Each income statement line is compared on seven scopes, and on an eighth against the prior forecast once a forecast vintage exists for that month. In the monthly package, a line that clears a dollar or percent threshold is flagged and sized, and each flagged line above EBITDA becomes a step in the Reported EBITDA bridge. Where your general ledger journal detail is on hand, the largest flagged variance is explained down to the journal entries behind it. The detail is proved to foot to the package's own source data first, and if it does not, no explanation is written. Most builds carry no journal detail, and then the package states what moved and not why.

How are add-backs decided?

Every add-back we identify is shown with the rule that decided it, and by default only what clears a stated test is included in Adjusted EBITDA. An analyst can include an item by hand, and the row shows that it was.

How does it work with my accounting system?

Every build starts from books the company has already closed and exported. Send whatever your system exports, ideally .xlsx or .csv, unlocked, one column per month. Where the converter cannot open something, it says so and names exactly what to ask for instead.

Which budget does the variance use?

The budget is an input the company supplies, and variance and the rest-of-year reforecast are measured against your own budget.

Does the package name the diligence workstreams it did not perform?

In a diligence package, yes. The package names the diligence workstreams it did not perform, why, and what would close each one, on the tab the buyer opens. The list is built per draft, so a workstream whose input did not arrive on this engagement appears on it, and the gate asserts it so it cannot quietly go missing.