Technology is not the starting point
Seven out of ten companies have already launched one or more AI experiments. Only a quarter say they have identified real opportunities to change their processes, and 5% have actually defined their roadmap (McKinsey, State of AI 2025). The gap is revealing: what holds things back is neither access to tools nor team appetite. It is the entry point.
In the finance function, starting from the tool is all the more misleading because the options are many and all legitimate: RPA for repetitive tasks with stable rules, the AI modules that ERP, accounting, treasury and consolidation vendors build in natively, and agentic platforms for building custom agents on top of existing systems. Three complementary building blocks, and none of them tells you where to start.
The core principle runs counter to the usual instinct: the starting point is not the technology building block but the business pain point. Technology is only a means, and the choice of building block comes at the end of the reasoning, once the use case has been characterized, not before.
The useful question is not "which tool should we deploy", but "which task wastes my teams' time every week without adding anything to the quality of the analysis". In finance, the answer varies little: gathering data, reconciling, formatting and searching for information.
The same pain points, scope after scope
Nine scopes are concerned: management control, consolidation, intercompany, treasury, customer and supplier accounting in shared service centers, financial control and credit management, order management, legal, and administrative and financial management. The processes differ, but the pain points overlap enough to be grouped into six families, and the first use cases follow the same logic from one scope to the next.
Time absorbed by gathering data across several tools and hunting for explanations, before analysis even begins. AI prepares a first reading of variances, suggests management commentary and flags outliers. Further along, a multi-source agent produces the monthly review summary.
Discrepancies between entities, time-consuming follow-ups, incomplete closing documentation. An agent reconciles inter-entity transactions, flags inconsistencies and lists the missing supporting documents. The benefit is not saved data entry: it is fewer back-and-forths.
Cash positions scattered across banks and tools, weak signals drowned in volume. An agent aggregates balances, detects flow anomalies and updates the 30/60/90-day forecast. AI improves readability; it does not claim to make the call.
Large volumes, highly heterogeneous documents, uncertain allocations. On the supplier side: reading invoices, matching purchase order / delivery note / invoice, preparing the approval batch. On the customer side: sorting incoming emails and tracking payment promises.
Information scattered across orders, disputes, payments and relationship history, hard to piece together before acting. AI rebuilds the customer context, prioritizes the cases to follow up and prepares the file before each action.
Reviewing long contracts under deadline, comparing versions, extracting scattered clauses. AI summarizes, extracts sensitive clauses and prepares a structured reading note. Often the best showcase: visible value, full control.
One common denominator is worth highlighting: none of these cases asks AI to decide. All of them are about preparing.
Three maturity levels, not one big leap
Once the cases are identified, the temptation is to aim straight for end-to-end automation. That is the best way to get nothing into production. One point that is often underestimated: maturity is not declared at company level, it is observed team by team, and three levels are enough to place each one.
- Level 1 — Assist. Summarizing a document, rewording an email, preparing a draft, extracting data from an invoice. Accessible to any employee, with no technical prerequisites.
- Level 2 — Automate. Automatic reconciliation, document classification, pre-drafted follow-ups, monthly review summaries. Requires stable processes and explicit rules.
- Level 3 — Orchestrate. Multi-source agents, end-to-end closing workflows, multi-step anomaly detection, automated preparation of executive committee meetings. Requires an established IT framework: access rights, traceability, validation checkpoints.
You start at level 1, validate, and move up in maturity gradually. This progression has a valuable side effect: it builds control reflexes before the stakes become significant.
AI prepares, people decide
The dividing line must be set explicitly, from the very first use case. AI prepares, alerts, reconciles, summarizes and highlights. People validate sensitive matters, exercise business judgment, make financial, commercial and legal decisions, handle exceptions and carry final responsibility.
In finance, this boundary is not a rhetorical precaution: accounting validation, budget trade-offs, investment decisions, blocking an order or entering into a contract remain human acts. Far from slowing adoption, a clear framework makes it easier: teams accept an agent all the more readily when they know exactly what it will not do.
The framework also protects the company. Without it, usage grows anyway, but outside any control: that is shadow AI, with the regulatory risks (AI Act, GDPR) and data leakage risks that come with it. Waiting is not a neutral option; it has technical and cultural costs.
Results observed in production
Three orders of magnitude, measured on recent projects.
The first concerns an industrial group whose financial documentation amounted to more than 5,000 scattered procedure documents, with experts constantly asked the same questions. Four agents were put into production: an orchestrator, an agent dedicated to accounting policies, a systems and processes agent and a quality control agent. The teams' feedback fits in one sentence: "No more waiting for an expert to be available for a procedure question."
The second concerns a legal department whose contract monitoring was entirely manual, with no alerts on contract expirations and renewals, across a portfolio that was heterogeneous in languages, formats and clauses. Two agents were enough: an extraction agent and a reporting agent, with different views by profile (legal, finance, sales). The gain is not only time saved: it is a consolidated, real-time view of risks.
In both cases, implementation did not take long. What takes time is business calibration: formulating the real questions, testing on imperfect files, adjusting the outputs each profile expects.
Where to start?
A concrete start beats a big project. The sequence comes down to four steps: surface your teams' pain points rather than an abstract technology promise; choose two or three targeted use cases that are visible and useful for the teams concerned; experiment with users, refine, then widen; co-build the usage while keeping an explicit human validation framework.
Sorting use cases, for its part, is best done as a funnel: diagnose the value chain to find the main pockets of value, eliminate low-impact or overly complex cases, prioritize by potential ROI and complexity, check data and IT prerequisites (data existence and quality, accessibility, skills), quantify gains and costs, then re-prioritize based on real feasibility, technical as well as organizational. This double filter is what avoids the list of fifty use cases, none of which ever reaches production.
The right start is not the one that covers the most ground. It is the one that identifies where AI can make a team's daily work a little simpler, clearer and more comfortable, and that makes people want to move on to the next one.
