The close is where most finance teams lose the most recoverable hours — but not in the way most people think. The bottleneck isn't a single slow process. It's the combination of deterministic tasks (journal entries, reconciliations, eliminations) running on manual workflows alongside probabilistic tasks (anomaly detection, narrative commentary) that require fundamentally different technology. Automating the deterministic steps with rule-based tools frees time. Applying GenAI to the narrative steps changes the nature of the work. Confusing the two is where most implementations stall.

Most FP&A teams spend 80% of their time assembling data and 20% actually analysing it. The opportunity isn't just flipping that ratio — it's recognising that the two halves require completely different technology. Data extraction and KPI calculation are deterministic: same inputs, same outputs, every time. That's automation territory. Trend identification and forecast accuracy analysis are pattern recognition problems suited to ML. And the variance commentary that lands on the CFO's desk? That's language synthesis — exactly what GenAI is built for. The teams getting real value apply all three where each belongs.

AP and AR are the highest-volume transactional workflows in most finance functions — and the ones where the technology distinction matters most practically. Invoice data capture is a pattern recognition problem: extracting fields from unstructured documents is what ML-based OCR was designed for. But the three-way match that follows? That's a deterministic check with one correct answer. Applying GenAI to transaction matching is like using a creative writer to do arithmetic. Meanwhile, the collections follow-up and dispute resolution sitting at the end of the AR cycle is genuine language work — drafting, adapting tone, referencing account history — where GenAI delivers real value.

Intercompany reconciliation is often the single biggest bottleneck in the consolidated close — and it's almost entirely deterministic. Transaction matching, balance confirmation, break identification, elimination entries: these all have one correct answer. The complexity isn't technical; it's operational — multiple entities, different systems, timing mismatches, and the human overhead of chasing counterparties to resolve breaks. Where ML adds value is in detecting patterns across unmatched items that humans miss at volume. And where GenAI fits is narrow but real: drafting the investigation narratives and counterparty communication that consume disproportionate time relative to the matching itself.

Regulatory reporting is where getting the technology classification wrong carries the highest consequences. The calculations, validations, and template populations that make up the bulk of regulatory work are deterministic — and must stay that way. There is no acceptable margin of error on a tax return or a statutory filing. GenAI has no role in the numbers. Where it does add value is in the narrative sections: disclosure drafting, management commentary for regulatory filings, and the explanatory notes that often take disproportionate time to produce. And ML-based regulatory change monitoring — scanning for rule changes across jurisdictions — is an emerging application that saves senior time on horizon scanning.

Time saved is the most commonly cited benefit of AI in finance — and the least reliable predictor of actual ROI. Research shows that productivity gains deliver no financial value unless they're explicitly converted: either downward to cost reduction (actual headcount or spend decreases) or upward to strategic capability (redeployed to business partnering that influences decisions). Without that conversion, freed capacity gets absorbed into increased demand. FP&A Trends research found only 15% of analyst time reaches business partnering activities. The question isn't how much time AI saves. It's what happens to that time.