"Should we use AI?" is the wrong question — the finance function is not one job. The useful move is to sort your team's work by its shape, because AI's reliability tracks the shape of a task far more than its subject matter. We use seven patterns (they're the same seven primitives our benchmark workloads instantiate), and they conveniently sort from most to least proven.
The seven patterns
1. Extract — unstructured in, structured out. Invoices, receipts, remittance advices, contracts → fields in a system. This is the most mature pattern: outputs are mechanically checkable, volumes are high, and failures are visible. Most AP automation is this pattern. Status: production-ready with sampling controls.
2. Classify — route, tag, prioritize. Inbox triage, transaction categorization, expense-policy flagging, lead scoring for collections. Cheap per item, easy to grade, and the cost of a single miss is usually low. The right first project for most teams because the blast radius of failure is one misfiled email. Status: production-ready.
3. Reconcile — compare sources, find discrepancies. Bank-to-GL matching, intercompany eliminations, subledger tie-outs. Higher value per unit than extract/classify and a genuinely painful manual job. The catch: the pass bar must be strict — a reconciliation that finds 11 of 12 breaks is not 92% useful, because the missed break is precisely the thing you were paying to find. Evaluate with recall-based criteria, not accuracy. Status: deployable with careful pass criteria.
4. Query — questions over structured data. "What did we spend with vendor X last quarter, ex-freight?" Natural language to SQL or to a BI layer. Enormously useful for self-service analytics; the risk is silently wrong answers that look plausible. Mitigate by having the AI return the query alongside the answer, so a reviewer can check the logic rather than re-derive the number. Status: deployable for analysts, risky for unreviewed executive consumption.
5. Generate — drafts to a template. Variance commentary, customer payment reminders, policy summaries, footnote drafts. The economics are strong (drafting is slow for humans, fast for models) but grading is genuinely harder — "is this good commentary?" is a judgment call. Keep humans as editors, and check the mechanical sub-facts (every figure cited must match the source) automatically. Status: production-ready as a drafting assist; not ready unreviewed.
6. Synthesize — many inputs, one decision-ready brief. Board-pack summaries, diligence-room digests, close-status roll-ups. High value, but the failure mode — a fabricated or misattributed figure in front of your board — is expensive, and detection requires reading the sources. Deploy with fact-checklists: every number in the brief must exist in the pack, verified mechanically. Status: promising, deploy with strong verification.
7. Execute — multi-step work with tools. The agentic frontier: close-task orchestration, exception resolution end-to-end, "reconcile this account and propose the correcting entries." This is where the vendor demos are most spectacular and the production evidence is thinnest. Costs multiply too — an agent that takes twenty steps pays for twenty model calls, so a per-outcome cost measured on single steps understates agentic cost dramatically. Status: pilot-only; measure before trusting.
Reading the map
Three practical rules fall out of this sorting:
Start at the top of the list. Extract and classify have checkable outputs, high volumes, and cheap failures. Wins there fund and de-risk everything below. Teams that start with execute-shaped moonshots usually end their AI program with the first public failure.
Match the control to the pattern, not the vendor pitch. Patterns 1–4 fail visibly (a field doesn't match, a query errors); sampling controls work. Patterns 5–7 can fail invisibly (fluent, wrong); they need verification designed into the output — cited figures checked against sources, required checklists, end-state validation. The controls guide covers this in depth.
Price the pattern, not the token. A classify call costs fractions of a cent; a synthesize job reads forty pages; an execute loop calls the model twenty times. The same "per-token price" produces per-outcome costs that differ by three orders of magnitude across patterns. When comparing options, always convert to cost per successful outcome for your unit of work — the pricing guide shows the arithmetic, and our editions publish measured numbers for representative workloads.
Where to start this quarter
If you run AP: extract + classify on the invoice inbox, measured against your current cost per invoice. If you run the close: reconciliation assist on your highest-volume bank account, with 100%-recall pass criteria. If you run FP&A: generate-pattern drafting for variance commentary, human-edited. Each is a two-week evaluation (see the pilot playbook), and each produces a number — cost per successful outcome versus your baseline — that makes the next conversation with your CFO short.