Before You Put AI Numbers in a Board Deck: A Verification Checklist

A board director once stopped a meeting on slide four. The deck claimed a $14 billion addressable market. She had sat on a competitor’s board the previous year and the number she remembered was $9 billion. The presenter froze. Nobody in the room could say where the figure came from, because it had come from a chat prompt three days earlier, and the prompt was gone.

That is the new failure mode. Decks used to be wrong because someone fat-fingered a cell. Now they’re wrong because a language model produced a confident, plausible, completely unsourced number and a junior analyst pasted it into a slide at 11pm. The number looks right. It reads like research. And under questioning, it evaporates.

This is not a reason to ban AI from deck prep. It’s a reason to verify before the deck leaves your hands. Below is the checklist we’d hand any finance team that uses AI to draft board or investor materials. It assumes the skeptical reader in the room is your sharpest director, not your friendliest.

The three errors that actually show up

Most AI deck mistakes fall into a small number of repeat offenders. Knowing the pattern is half the catch.

Stale or invented market sizing. TAM, SAM, and SOM slides are where models hallucinate most freely, because the “right” answer is genuinely fuzzy and the training data is full of optimistic press-release figures. Tools that auto-populate a market-size slide tend to pull numbers that are either years out of date or fabricated outright, with no citation attached. Investors flag inflated market numbers pulled from thin air as a top credibility killer, and the damage is immediate: if you can’t defend the figure in the room, the rest of the deck inherits the doubt.

Annual-versus-quarterly confusion. This one is quieter and more dangerous. Ask a model for “revenue” across a few entities or periods and it will happily mix a quarterly figure into an annual roll-up, or annualize a one-time number, without telling you. The slide footnote says “FY2025.” The bar chart underneath blends Q4 actuals with full-year projections. No error is thrown. The total is just wrong by 3x in one segment.

Uncited figures that can’t be traced. A number with no provenance is a liability even when it’s correct, because you can’t prove it. If a director asks “what’s behind the 41% gross margin,” and the honest answer is “the AI said so,” you have already lost. Every figure on a board slide should be answerable with a system, a period, and a calculation, not a vibe.

The pre-send checklist

Run this on every figure before the deck circulates. It takes minutes per slide and it has saved meetings.

1. Source every number to a system, not a chat

For each figure, you should be able to name where it lives: the GL account, the ledger, the data warehouse table, or the specific statement it was pulled from. “ChatGPT generated it” is not a source. If the figure traces back only to a model output, treat it as unverified until you tie it to a system of record.

2. Confirm the period and the unit

Check that every figure on a slide shares the same period basis. Annual with annual, quarter with quarter, trailing-twelve with trailing-twelve. Confirm currency and scale (thousands versus millions is a classic). Mixed periods are the single most common silent error in AI-assisted decks, and they survive a casual read.

3. Re-derive the math yourself, once

Growth rates, margins, ratios, and CAGRs are where models reason badly. A model may invert a ratio, confuse absolute change with percentage change, or get the sign wrong on a cash figure, and present the result with total confidence. Pick the load-bearing calculations on each slide and reproduce them in a spreadsheet or calculator. If you can’t reproduce it, it doesn’t ship.

4. Reconcile totals against the audited or closed numbers

The summary figures on your deck should tie out to your last close or your audited statements. If the deck’s “total revenue” doesn’t match the number your controller signed off on, find out why before, not during, the meeting. A figure that doesn’t reconcile to the ledger is not a board figure.

5. Pressure-test the outliers

Any number that would make a director sit up (a market size, a growth rate, a margin that jumped) gets a second pass. Is it plausible against last quarter? Against the competitor set? Against physics? Outliers are where hallucinations hide, because they’re also where genuine good news hides, and the two look identical on a slide.

6. Keep the provenance attached

When the deck goes out, the backup should exist: for each headline figure, which source, which period, which calculation. You may never need it. The one time a director pushes, having it is the difference between composure and a scramble.

If you read that list and thought “this is most of my Thursday,” you’re right. The checklist works, and it does not scale. Doing it by hand on every deck, every quarter, is exactly the kind of manual reconciliation that AI was supposed to remove. So the better question is structural: why is verification a separate, manual step at all?

Make the figures traceable by construction

The reason the checklist is painful is that the AI and the source of truth are two different things. The model produces a number; you go hunting for where it came from. Flip that. If the AI can only answer with figures that already trace to a reconciled source, most of the checklist disappears, because verification was done before the number existed.

That’s the model we build toward at Rexfin. Connect your accounting and financial-data systems (QuickBooks, Xero, NetSuite, Sage, SAP, Oracle, your warehouse) or upload statements, and the platform builds one reconciled financial model that ties out to the ledger. When AI then answers a question for your deck, it retrieves figures from that model rather than inventing them, and any calculation (a margin, a growth rate, a CAGR) runs through a deterministic engine, not the language model’s probabilistic guess.

The payoff is exactly the checklist, automated. Every figure carries a source, a period, and a calculation path. Annual and quarterly stay distinct because the model knows which is which. Totals reconcile to the close because that’s what they were built from. And when a director asks “what’s behind this,” the answer is a trace, not a shrug.

This matters more as the inputs get messier. If your figures come from scanned PDFs or document images, extraction accuracy drops sharply, and a deck built on misread source documents fails the checklist before you even reach the math. Reconciliation upstream is what keeps the downstream slide honest. For the broader picture of where these errors come from and how to test for them, the financial-hallucinations pillar walks through the full set.

None of this means trusting the AI blindly. A director should still ask hard questions, and a good system should make the answers easy. The shift is from “verify the output after the fact” to “the output can only be a verified figure.” One of those scales to every deck you’ll ever ship. The other is your Thursday.

The next time a deck claims a number that would make a board member pause, you want the answer ready before the question lands. Numbers that trace to a reconciled source give you that. Book a demo and bring your messiest deck, we’ll show you how each figure on it would tie back to the ledger.