Every finance team has heard some version of “just have AI do your forecasting” in the last two years. Most of that advice skips the part that actually matters: what AI changes in FP&A, and, just as important, what it doesn’t.
What FP&A Actually Is, Before AI Enters the Picture
Financial planning and analysis is a three-part loop: build a forecast, track actuals against it every period, and explain the variance between the two in terms an executive team can act on. None of that loop is new, and none of it is replaced by AI tooling. It’s accelerated by it, in specific places.
Where AI Genuinely Helps
Scenario modeling. Historically, testing “what happens to runway if CAC rises 20% next quarter” meant manually reworking a spreadsheet, often taking the better part of a day for a non-trivial model. AI-assisted modeling tools can iterate that scenario in minutes, which changes the practical number of scenarios a team can actually explore before a decision gets made.
Variance analysis. Explaining why actuals diverged from forecast used to mean manually cross-referencing line items against assumptions. AI tooling can draft a first-pass variance explanation directly from structured accounting data, which a human then verifies and refines, turning a half-day task into a review task.
First-pass drafting. Board decks, investor updates, and monthly reporting narratives can get a competent first draft generated from the underlying model output, cutting the time from “numbers are final” to “the deck is ready.”
Where AI Does Not Help, and Can Actively Hurt
The failure mode worth naming directly: an AI tool will generate a confident, fluent-sounding forecast or variance explanation regardless of whether the underlying data is clean. If your transaction categorization is inconsistent, if your revenue recognition logic isn’t encoded correctly in the model, or if your assumption set is stale, AI tooling will not catch that. It will simply produce a fast, wrong answer instead of a slow, wrong answer. Fixing the data structure and assumption set has to happen before AI acceleration adds real value; skipping that step is the single most common way FP&A “AI initiatives” fail to deliver anything durable.
The judgment calls (which assumptions are defensible, what a variance actually means for the business, what a board needs to hear versus what’s merely accurate) still require a human who understands the business, not a model trained on generic financial patterns.
A Practical Adoption Order
The same discipline applies here as in sequencing AI adoption anywhere else in the business: fix the foundation before automating on top of it, and don’t skip stages.
- Fix the data structure first. Clean, consistent transaction categorization and a model architecture that reflects how the business actually operates.
- Introduce AI-assisted scenario modeling second. Once the base model is sound, use AI tooling to expand how many scenarios get tested before a decision.
- Add AI-assisted variance drafting third. Let AI produce a first-pass explanation that a human FP&A owner reviews and finalizes. It never ships unreviewed.
- Extend to reporting narratives last. Board and investor communications benefit from AI-assisted drafting, but this is the lowest-risk, lowest-priority step. Get the model and the analysis right first.
| Order | Step | Risk if skipped |
|---|---|---|
| 1 | Fix data structure | Every later step compounds bad data |
| 2 | AI-assisted scenario modeling | Scenarios built on a shaky base model |
| 3 | AI-assisted variance drafting | Explanations drift from what actually happened |
| 4 | AI-assisted reporting narratives | Lowest risk, but only sound if 1-3 are already right |
This is the same sequencing NCR Digital applies in AI strategy work: fix the model, then accelerate the iteration on top of it, never the other way around.
Other reads
FAQ
Can AI replace an FP&A analyst?
Not the judgment calls — which assumptions are defensible, what a variance actually means for the business, what a board needs to hear. AI accelerates the mechanical parts of the loop (scenario iteration, first-pass variance drafting), not the interpretation.
What's the biggest risk of AI-assisted forecasting?
A model built on messy or inconsistently categorized data will produce a fast, confident, wrong answer just as easily as a slow one. AI doesn't catch bad underlying data — it makes the bad answer arrive faster.
Where should a team start with AI in FP&A?
With the data structure, not the AI tooling. Clean, consistent transaction categorization and a model architecture that reflects how the business actually operates has to come before any AI-assisted acceleration adds real value.
This post is part of the AI Strategy capability.
See the full AI Strategy capability