Which AI use case should finance operations fund this year?
Fictional example. The company and its numbers are invented, but realistic. Every assumption is labelled so you can replace it with your own figures.
Decision: Fund supplier invoice automation. Expected payback is about 6.5 months after go‑live, on a one-time cost of €84,000.
Do not build a contract-review assistant. At current volume it would take nearly 10 years to pay back.
Stop rule: after two weeks of shadow mode, we project payback from the two numbers we observe: the share of invoices processed with no human touch, and the review time for the rest. If projected payback is over 18 months, we recommend stopping.
1. The shortlist
We interviewed 7 people in finance and IT and reviewed three months of invoices, emails, and contracts. Four candidates came out of the interviews.
| Use case | Value / year | Data readiness | Call |
|---|---|---|---|
| Supplier invoice entry | ~€182K capacity | High: 3 years of posted invoices with POs | Build now |
| Supplier email triage | ~€35K | High | Buy |
| Demand forecasting | Largest, unquantified | Low: sales history split across two ERPs | Not yet |
| Contract-review assistant | ~€22K | Medium | Don’t build |
Email triage is already offered by the client’s ERP vendor as a paid add-on; building it would duplicate a product they can switch on. Forecasting needs the two ERPs reconciled first; we will re-score it once that project lands.
2. Cost model: supplier invoice entry
| Line | Assumption | Amount |
|---|---|---|
| Manual work today | 4,000 invoices a month × 6 minutes = 400 hours, at €45 an hour fully loaded | €18,000 / month |
| Human work that remains | 30% of invoices go to a review queue at 3 minutes each; 2% of auto-posted invoices are audited at 2 minutes each | 62 hours / month |
| Capacity recovered | 338 hours a month, about 2 full-time roles | €15,216 / month |
| Running cost | Model usage €200 (about 12M tokens), document OCR €160, hosting and monitoring €900, engineering upkeep €1,000 | €2,260 / month |
| Net benefit | €12,956 / month | |
| Internal engineering | 2 engineers × 8 weeks = 80 person-days at €550 | €44,000 |
| ERP integration partner | 10 days at €1,200 | €12,000 |
| Training and process change | Finance team and audit sign-off | €4,000 |
| This Sprint | Included so the whole decision is costed | €24,000 |
| One-time cost | €84,000 | |
| Payback after go‑live | €84,000 ÷ €12,956 | 6.5 months |
For the CFO: recovered hours are capacity, not cash. They become savings only if you avoid hiring for volume growth or redeploy the team. This case assumes the two accounts-payable hires budgeted for next year are not made.
3. Sensitivity and stop rules
The two numbers that drive the case are the share of invoices processed without a human and the review time for the rest. We measure both in shadow mode before anything posts automatically.
| Scenario | No-touch share | Review time | Net / month | Payback |
|---|---|---|---|---|
| Expected | 70% | 3 min | €12,956 | 6.5 months |
| Conservative | 50% | 4 min | €9,680 | 9 months |
| Low automation, fast review | 30% | 3 min | €9,404 | 9 months |
| Weak | 40% | 4.5 min | €7,592 | 11 months |
| Stop (example) | 30% | 5.5 min | €4,154 | 20 months |
- Gate 1, after 2 weeks of shadow mode: project payback from the observed no-touch share and review time. Up to 12 months: go. 12 to 18 months: re-scope to the top suppliers only. Over 18 months: stop. Neither number decides alone: 30% no-touch with 3-minute reviews still pays back in about 9 months.
- Gate 2, before any auto-posting: at least 99.5% field accuracy on totals and supplier IDs across 500 labelled invoices.
4. Target architecture
Components
- Intake: accounts-payable mailbox and supplier portal into a document store.
- Extraction: a managed multimodal model hosted in an EU region, returning a fixed JSON schema. No training on client data.
- Validation: three-way match against purchase order and goods receipt, with business rules in code, not prompts.
- Orchestration: a LangGraph workflow that routes low-confidence invoices to a human review queue.
- Posting: the ERP’s existing API, idempotent jobs, full audit log. No direct database writes.
Token budget
- About 3,000 tokens per invoice × 4,000 invoices = 12M tokens a month.
- Hard monthly cap of €400, alert at 80%.
Policies
- AI never changes supplier bank details.
- Invoices above €25,000 always need human approval.
- No invoice data leaves the EU.
- Every posting is logged with its confidence score and source fields.
Evaluation: 500 labelled invoices scored field by field, re-run on every model or prompt change. A release that drops below the bar does not ship.
5. What we recommend against
Contract-review assistant: don’t build. The legal team reviews about 25 contracts a month. AI would save roughly 0.8 hours each, at €90 an hour: €1,800 a month. A reliable build with a clause library and evaluation costs about €70,000, and running it costs about €1,200 a month.
| Net benefit | €600 / month |
| Payback | 117 months |
If the team wants help sooner, trial a vendor contract-review tool for 60 days. Revisit a custom build only if volume passes about 70 contracts a month, the point at which payback falls under 18 months.
6. The first 90 days
- Weeks 1 to 8: two engineers build intake, extraction, validation, and the review queue. Label 500 invoices for evaluation.
- Weeks 9 and 10: shadow mode on live invoices. Nothing posts automatically.
- Week 11: Gate 1 and Gate 2. Go, re-scope, or stop.
- Week 12: auto-post high-confidence invoices only; widen as accuracy holds.