In 2023, Unilever's finance team quietly killed an AI forecasting pilot that had cost the better part of a year to build. The model was technically excellent—it beat the human baseline on backtests. It failed for a reason that has nothing to do with data science: the demand planners didn't trust a number they couldn't interrogate, and the model couldn't explain why it disagreed with them. The forecast was more accurate and less used.
This is the trap waiting for every CFO writing an AI investment case right now. The vendor deck shows accuracy lift. The board wants a headline about "AI transformation." And somewhere between those two pressures, capital flows toward use cases that demo beautifully and deliver nothing—while the genuinely valuable applications go unfunded because they're unglamorous.
Your job is not to adopt AI. Your job is to allocate scarce capital and organizational attention to the finance use cases where AI's economics actually work. That requires a framework sharper than "where can we use AI"—it requires knowing where AI's specific failure modes are survivable and where they are fatal.
Every proposed AI use case in finance should pass through four filters. Skip any one of them and you're buying a demo, not an asset.
1. Is the task prediction, or is it judgment? AI is a prediction machine. It excels where the answer is a pattern extrapolated from history and fails where the answer requires understanding a situation that has never occurred. Cash flow forecasting is prediction. Deciding whether to breach a covenant to preserve a strategic relationship is judgment. The moment a task requires weighing consequences the data has never seen, AI's advantage collapses.
2. What is the cost of being wrong, and who catches it?
3. Is the underlying relationship stable? AI assumes tomorrow resembles the training data. In finance, this holds for high-frequency, structurally stable processes (invoice matching, transaction categorization) and breaks precisely when it matters most—during regime changes, market dislocations, or one-off strategic events. The 2020 demand shock destroyed forecasting models trained on pre-pandemic patterns. Ask: how often does the ground truth here shift structurally?
4. Can you afford the explanation problem? In regulated finance, "the model said so" is not an audit trail. If a use case touches statutory reporting, tax positions, or anything an auditor or regulator will challenge, you need explainability, not just accuracy. This single constraint eliminates a large fraction of black-box applications from the parts of finance that matter most to your signature.
Run any use case through these four filters and it sorts itself into one of three tiers.
These share a profile: high-volume, pattern-rich, low cost-of-error, and a natural human checkpoint.
Anomaly detection is the strongest use case in the entire finance function, and it's underhyped precisely because it isn't sexy. Fraud detection, duplicate payments, expense-policy violations, journal-entry outliers, reconciliation breaks—these are tasks where the model doesn't need to be right, it only needs to *flag*. A human confirms. The economics are asymmetric in your favor: false positives are cheap, and every true positive caught is money or risk avoided. JPMorgan's transaction-monitoring systems don't decide anything; they narrow millions of transactions to a reviewable set. That is the correct architecture.
Transaction processing and matching—invoice-to-PO matching, cash application, intercompany reconciliation—is Tier 1 because the relationships are stable and the volume justifies the build. This is where automation and machine learning genuinely compress cost per transaction. The catch: measure the value in *cycle time and error reduction*, not headcount. The teams that framed these projects as "replace three FTEs" almost always over-promised and under-delivered on the harder tasks that remained.
Forecasting—with a critical caveat. AI improves forecasting where you have high-frequency data and stable drivers: short-horizon cash forecasting, granular demand at SKU level, collections timing. The value is real but bounded—expect accuracy improvements in the range that matters operationally, not miracles. The Unilever failure is the lesson: a forecast that isn't trusted isn't used, and trust requires the model to *surface its reasoning* and let planners override it. The winning design is human-in-the-loop, where AI produces a baseline and the finance team adjusts with documented rationale. Fully autonomous forecasting fails filter #4.
Narrative reporting is the use case that has changed most with generative AI, and where the sharpest CFO judgment is now required. LLMs can draft the MD&A commentary, board-deck narratives, variance explanations, and IR talking points from structured financial data. The productivity gain is genuine—drafting time drops dramatically. But this is exactly where filter #2 bites hardest: a generative model will *fabricate a plausible explanation* for a variance it doesn't understand. It will confidently attribute a margin decline to "FX headwinds" when the real cause was a pricing error.
The correct deployment: AI drafts, finance verifies every claim against source data, and no generated narrative reaches an external audience without human sign-off. Use it to eliminate the blank page, not to eliminate the analyst. The value is speed-to-first-draft, not autonomous authorship.
Scenario and driver-based modeling sits here too. AI can rapidly generate and stress-test scenarios, but the *choice* of which scenarios matter and the interpretation of results remain judgment. Treat AI as a way to explore the possibility space faster, not to tell you which future to plan for.
Be as disciplined about the "no" list as the "yes" list. This is where CFO credibility is won or lost.
Strategic capital allocation, M&A judgment, and one-off decisions. These fail filters #1 and #3 completely. They are judgment under genuine uncertainty, drawing on relationships, competitive dynamics, and consequences no dataset contains. An AI can assemble the analysis; it cannot make the call. CFOs who let a model's output substitute for their own judgment here are abdicating the core of the role.
Anything requiring accountability you can't delegate. Statutory sign-offs, tax positions, going-concern assessments—the cost of error is catastrophic and the explanation problem is unforgiving. AI can support the preparation; it cannot own the conclusion.
Long-horizon forecasting through regime change. The longer the horizon and the more likely a structural break, the worse AI performs relative to informed human judgment. Models are dangerously confident right up to the moment the world changes.
The pattern across Tier 3 is consistent: high cost-of-error, unstable relationships, and accountability that regulation or fiduciary duty places on a human. No amount of model accuracy overcomes those constraints.
Vérification des acquis
1. The Unilever forecasting pilot was described as 'more accurate and less used.' What is the core lesson this paradox illustrates?
2. According to the lesson's first filter, why does cash flow forecasting qualify as an AI-suitable task while deciding whether to breach a covenant does not?
3. The lesson argues a CFO's real job regarding AI is best described as which of the following?
4. Select ALL correct answers. According to the lesson, which pressures cause CFOs to misallocate capital toward the wrong AI use cases?
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers. Which statements accurately reflect the lesson's reasoning about AI errors and the 'cost of being wrong' filter?
Sélectionnez toutes les réponses correctes.
Knowing the tiers is analysis. Deploying capital against them is the job. Here's the operating discipline.
Sequence by "value density," not visibility. Rank candidate use cases by expected annual value divided by implementation cost and risk. Tier 1 anomaly detection and transaction matching almost always top this ranking because the value is captured continuously and the risk is contained. Resist the board's pull toward the flashy generative use case as your *first* project—win credibility with a boring, high-ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → anomaly-detection deployment before you touch anything that faces external audiences.
Design the human checkpoint before you design the model. For every use case, answer explicitly: who reviews the output, what does a caught error cost, and what does a *missed* error cost? If you cannot name the reviewer and the checkpoint, the use case is not ready—regardless of model accuracy. This single practice would have flagged most of the failed AI-in-finance projects of the last three years before a dollar was spent.
Measure against the right baseline. The comparison is never "AI vs. perfect." It's "AI-plus-human vs. current process." A forecasting model that's 8% more accurate but that planners override into inaccuracy delivers zero. Instrument adoption and trust, not just backtest accuracy. If usage is low, the project has failed even if the model is excellent—exactly the Unilever failure mode.
Build the data and governance foundation as the actual project. In most finance functions, the binding constraint on AI value is not the model—it's fragmented, poorly governed data. The uncomfortable truth is that the "AI project" is 70% a data-quality and governance project. CFOs who fund the model without funding the foundation buy an expensive disappointment. This is also where your control environment must extend: model governance, version control, and audit trails for AI outputs are now part of your internal controls, not an IT afterthought.
Structure the vendor conversation around your failure modes, not their features. When a vendor demos accuracy, ask: show me the false-positive and false-negative rates. Show me how a user interrogates a specific output. Show me the audit trail. Show me performance during the last market dislocation in your training window. The demo answers none of these; the diligence must.
The CFOs who create value from AI over the next three years won't be the ones who deployed the most—they'll be the ones who deployed the *right* use cases with the discipline to say no to the rest. The technology will keep improving. The framework for judging where it belongs will not change: prediction versus judgment, cost of error, stability of relationships, and the explanation you can defend.
1. Sort every AI use case through four filters before funding it: prediction vs. judgment, cost-of-error and who catches it, stability of the underlying relationship, and whether you can afford the explanation problem. Fail any one filter and the case is a liability, not an asset.
2. Lead with anomaly detection and transaction matching, not generative reporting. They have the best value density—asymmetric payoffs, contained risk, natural human checkpoints—and they build the credibility to fund harder projects later.
3. Deploy generative narrative reporting as a drafting tool with mandatory human verification. The value is speed-to-first-draft; the danger is fluent fabricationfabricationA hallucination is when an AI model generates output that is fluent and confident but factually wrong, fabricated, or unsupported by its source data.Voir la définition complète →. No AI-generated narrative reaches an external audience without a human validating every claim against source data.
4. Design the human checkpoint before the model, and measure AI-plus-human against your current process—never against perfection. A more accurate forecast that no one trusts and uses delivers zero value.
5. Fund the data and governance foundation as the real project. The binding constraint is rarely the model; it's data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète → and control. Extend your internal controls to cover model governance and AI audit trails.