Every finance team is being told to "use AI". Far less common is a straight answer to the only questions that matter: where does it actually pay off, and is our process even ready for it? For most finance functions the honest answer is "in a few specific places, and only once the underlying process is in order".
The 80/20 problem
AI projects in finance rarely fail on the model. They fail on data and process readiness — the unglamorous 80% — because you cannot automate a process that isn't standardised. Point a clever tool at an inconsistent, spreadsheet-driven close and it learns the inconsistency. Readiness for AI is gated by process maturity, which is why the smart sequence is: stabilise the process, then automate it, then add intelligence on top.
Where AI genuinely earns its place in finance
Across the finance cycle, the use-cases that consistently repay the effort cluster in a few areas:
- Transaction matching & reconciliation — auto-matching intercompany and bank/ledger items, surfacing only the exceptions.
- Anomaly detection — flagging unusual journals, balances or trends before they reach the pack, not after.
- Account mapping — learning how source accounts roll into the group structure to speed consolidation onboarding.
- Predictive forecasting — driver-based and statistical forecasts that update continuously rather than once a quarter.
- Narrative & disclosure — drafting commentary, board narrative and statutory disclosure from the underlying numbers.
- Conversational analysis — "ask the numbers" interfaces over governed data, so analysts stop building one-off cuts.
Notice what's not on that list: replacing judgement. The value is in removing manual gathering, matching and drafting so that scarce finance time moves to the decisions only people can make.
Ready now vs fix-first
The same use-case can be a quick win or a money pit depending on where the underlying process sits. A rule of thumb:
| Process maturity | Right AI move |
|---|---|
| Level 1–2 (manual / templated) | Fix first. Standardise the process before automating — tooling here just industrialises the chaos. |
| Level 3 (system-supported) | Efficiency AI — matching, anomaly detection, mapping. Take out the manual effort. |
| Level 4–5 (automated / optimised) | Effectiveness AI — prediction, narrative, conversational analysis. Sharpen the decisions. |
Don't buy the hype; buy the readiness
The right first step isn't a tool selection — it's an honest read of which finance processes are ready for AI, which need foundations first, and what each move is worth. That diagnosis is exactly what turns a vague "AI strategy" into a costed, sequenced plan a board will fund.
Common questions
Where does AI actually pay off in finance?
The use-cases that consistently repay the effort are transaction matching and reconciliation, anomaly detection, account mapping for consolidation, predictive/driver-based forecasting, narrative and disclosure drafting, and conversational 'ask the numbers' analysis. The value is removing manual gathering and matching, not replacing judgement.
Why do AI projects in finance fail?
They rarely fail on the model — they fail on data and process readiness, the unglamorous 80%. You can't automate a process that isn't standardised; pointing AI at an inconsistent, spreadsheet-driven process just industrialises the inconsistency. Readiness for AI is gated by process maturity.
Is my finance function ready for AI?
It depends per process. At level 1–2 maturity, fix the process first. At level 3, efficiency AI (matching, anomaly detection) pays off. At level 4–5, effectiveness AI (prediction, narrative, conversational analysis) sharpens decisions. The first step is an honest read of which processes are ready and which need foundations.