You're handed a financial report with an AI-assisted projection in it — a revenue forecast, a cost estimate, a churn prediction — and you have limited time before a decision needs to be made. What do you actually check, quickly, before treating that number as something to act on? This is a practical, fast checklist for exactly that moment, built for someone reading the report, not someone who built the model behind it.

Question one: is this a measured fact or a modeled estimate?

The first and most important question, and often the hardest to answer because reports don't always make the distinction obvious: is this specific number something that was actually observed (a transaction that happened, a cost that was actually incurred) or something a model produced based on patterns in historical data? These deserve very different levels of trust, and a report that doesn't clearly distinguish between the two is itself a red flag worth raising before you go any further.

A practical habit: scan the report specifically for language that signals which category each number falls into — "actual," "recorded," "confirmed" versus "projected," "estimated," "forecast." If a report uses these terms inconsistently or not at all, ask directly, before using any number in a decision, which category it belongs to.

Question two: what historical window and assumptions produced this?

An estimate is only as good as the assumptions behind it, and those assumptions are rarely visible in the headline number itself. Ask specifically: what time period's data was this based on, and does that period still resemble current conditions? A forecast built on the last two years of data will be a poor guide if something material has changed in the last month — a new competitor, a pricing change, a shift in customer behavior — that the historical window doesn't yet reflect.

Tip

A fast, practical version of this question: "what would have to be true about the future for this number to be wrong?" If you can immediately think of something plausible that would invalidate the number — a known upcoming change the model has no way to know about — that's specific, actionable information the raw number itself doesn't convey, and it belongs in the conversation before the number gets used to justify a decision.

Question three: what's the actual range, not just the point estimate?

A single number ("$1.2M in projected revenue") conveys false precision, because every forecast has some real range of plausible outcomes even when only one number gets presented. Ask whether a range or confidence interval exists behind the headline number, and if it does, look at how wide it actually is — a narrow range suggests a more stable, well-understood pattern; a wide range suggests genuine uncertainty that a single confident-sounding number obscures. If no range was calculated or reported at all, that's worth flagging directly, because decisions made against a false sense of precision are a common, avoidable source of bad outcomes.

Question four: has this forecasting approach been checked against reality before?

Ask whether anyone has actually compared this model's past forecasts against what really happened. A forecasting process with a real track record of being checked against outcomes deserves more trust than one that's never been evaluated this way, regardless of how sophisticated the underlying method sounds. If nobody can answer this question — if the model's actual accuracy has simply never been measured — that absence of evaluation is itself important information, not a neutral unknown.

Question five: does this pass a basic sanity check against what you already know?

The final, fastest check: does this number make sense given context you have that the model likely doesn't — a contract that just closed, a customer relationship that's visibly deteriorating, a market shift that's common knowledge internally but wouldn't show up in historical transaction data yet? This is the same discipline discussed elsewhere as the core defense against over-trusting AI-generated forecasts: a human with current, real context is the final check the model structurally cannot perform on its own, because the model only knows what was in its training or input data, and recent real-world developments often aren't in either yet.

What to actually do if a number fails this checklist

Failing one of these checks doesn't mean throwing the number out — it means asking for more information before treating it as decision-ready. "What assumptions is this based on" and "has this approach been checked against past results" are reasonable, specific questions to bring back to whoever produced the forecast, and a forecasting process that can answer them clearly and specifically deserves more trust than one that responds with vague reassurance instead of a concrete answer.

Why this is a literacy skill worth building deliberately

As AI-assisted numbers show up in more everyday business reporting, the ability to run this kind of quick, specific skepticism check becomes a genuinely valuable professional skill — not because AI forecasting is untrustworthy by default, but because the confident, fluent presentation of any AI-generated number makes it easy to skip exactly the questions that determine whether it deserves that confidence.

The AI-Powered Finance and Decision Engine builds this checklist out into a full practical guide, including real examples of forecasts that passed and failed each of these checks, and what the actual outcome turned out to be in each case — useful calibration for how much these specific questions actually matter in practice, not just in theory.