The list usually arrives after an offsite or a vendor day. Thirty ideas on a slide: invoice capture, a chatbot for the policy manual, AI-written board commentary, forecast models, contract review. The team wants to start with the board commentary because the CEO will see it. The audit committee chair, a week later, asks which one comes first and why. Asked to rank a list of AI use cases in finance, many finance leaders answer with a preference rather than a method.
A method is not hard to build. Score each idea on five factors, read benefit against risk, and let the risk score decide how much human review the design needs. This guide sets that out, with worked examples across payables, receivables, the close, reporting and compliance.
Why lists of AI use cases in finance get ranked wrong
They get ranked by visibility. The ideas that demo well are the ones that produce something a senior person reads, such as a narrative, a summary or an answer to a question. Those are often the hardest to get right and the most expensive when wrong.
The work that pays back is usually less visible. Gartner surveyed 183 finance leaders in May and June 2025 and, in a November 2025 release, reported the three most common AI uses in finance as knowledge management (49%), AP process automation (37%) and error and anomaly detection (34%). Most respondents (91%) described the impact as low or moderate initially. The Hackett Group's 2026 Finance Key Issues study, which does not disclose its sample, described accounts payable as "at the forefront", with 33% of organizations already scaling AI solutions there.
Neither survey tells you what to do in your own team. Both point the same way: the early ground is transactional, and impact takes longer than the first wave of expectations suggested.
Five factors to score
Score each use case from 1 (low) to 5 (high) on five factors. The first three describe benefit. The last two describe risk.
| Factor | Question to ask | Scores 1 | Scores 5 |
|---|---|---|---|
| Volume | How many times a month does this happen, and how many hours does it take? | A few times a quarter | Thousands of items, or many hours a week |
| Rule-bound | Could a careful new joiner follow written rules to get it right? | Depends on experience and context | Rules exist and cover most cases |
| Data availability | Is the input already in a system, in a consistent form? | Scattered across email and spreadsheets | In the ERP or a single source, structured |
| Error cost | If the output is wrong and nobody notices, what happens? | Minor rework | Misstated accounts, cash lost, or a filing missed |
| Control risk | Does it touch financial reporting, payments, master data or external parties? | Internal and advisory only | Posts entries, releases cash or reaches a regulator |
Two points on scoring. Volume should come from a measurement, not a guess; the time audit from last week's issue gives you that. And error cost is about the consequence of an undetected mistake, so a use case that only flags items for a person to review usually scores lower than one that acts on its own.
Worked examples across the finance function
The scores below are illustrative, for a mid-sized company with one main ERP and several entities. They are judgments, not measurements, and yours will differ. The point is the pattern.
| Area | Use case | Volume | Rules | Data | Error cost | Control risk |
|---|---|---|---|---|---|---|
| AP | Invoice data capture and GL coding suggestions | 5 | 4 | 4 | 2 | 2 |
| AP | Triage of three-way match exceptions | 4 | 5 | 4 | 3 | 3 |
| AP | Duplicate payment check before the run | 4 | 5 | 4 | 2 | 2 |
| AR | Cash application matching | 5 | 4 | 3 | 3 | 3 |
| AR | Collections prioritization and draft reminders | 4 | 3 | 4 | 2 | 2 |
| Close | Bank and balance sheet reconciliation matching | 5 | 5 | 5 | 3 | 3 |
| Close | First-draft flux commentary | 3 | 2 | 2 | 3 | 3 |
| Reporting | Questions about the numbers answered from the ledger | 3 | 3 | 3 | 3 | 2 |
| Reporting | Board pack narrative | 1 | 1 | 3 | 5 | 4 |
| Compliance | Statutory deadline tracking and reminders | 3 | 5 | 4 | 4 | 2 |
| Compliance | Tax positions and technical judgments | 2 | 2 | 3 | 5 | 5 |
Accounts payable. Invoice capture and coding scores well because the volume is high and coding usually follows the supplier and the purchase order. Three-way match triage scores higher on rules than people expect: the tolerance rules are written down, and the work is sorting exceptions into price, quantity and missing-receipt queues for the right owner. A duplicate check before the payment run is a detective control. It flags; a person decides. That keeps its error cost low.
Accounts receivable. Cash application has the volume, but the data score drops wherever customers pay without remittance details, and the match then depends on judgment about partial and combined payments. Collections drafting is useful and low-risk if a person sends every message. The rules score is middling because who to call first depends on context the aging report does not hold.
The close. Reconciliation matching is often the strongest candidate in the whole list. The rules are explicit, both sides of the match are in systems, and a person reviews the unmatched items anyway. Flux commentary is different. AI can produce a fluent first draft, but the reason a cost line moved often sits in someone's head or an email rather than in a system, so both the rules and the data scores are low.
Reporting. Answering questions from the ledger works when the figures come from the ledger and the model explains them, and less well when it is asked to produce numbers. Board narrative scores lowest on benefit: it happens once a month or quarter, there are no rules, and the cost of a confident wrong sentence in front of a board is high.
Compliance. Deadline tracking is rule-bound and benefits from systematic reminders, but the error cost of a missed filing is high, so a named owner stays accountable for every date. Tax technical positions are a poor fit for AI-led work at this stage and remain with qualified people.
How to read the scores
Add the three benefit scores and the two risk scores separately. Then place each use case in one of four groups.
| Benefit (out of 15) | Risk (out of 10) | What to do | Design for human review |
|---|---|---|---|
| 11 or more | 5 or less | Start here | AI acts within rules; a person samples and handles exceptions |
| 11 or more | 6 or more | Start, with tighter design | AI suggests; a person approves each item or batch |
| 10 or less | 5 or less | Second wave | Useful assistance; low priority for investment |
| 10 or less | 6 or more | Park | Revisit when data or rules improve |
On the illustrative scores, duplicate checks, invoice coding and collections drafting fall in the first group. Reconciliation matching, three-way match triage, cash application and deadline tracking fall in the second group, start with tighter design: worth doing early, with a person approving what the AI proposes. Ledger questions sit in the second wave. Flux commentary, board narrative and tax positions are parked. Flux commentary often moves up once the reasons for variances are captured somewhere other than people's inboxes, which raises its data and rules scores.
The risk score is not a veto. Its main job is to set the mode. KPMG's Global AI in Finance survey of 1,013 senior finance leaders, published in May 2026, found that only 42% of organizations were fully assurance-ready for AI-enabled finance processes. If yours is in the majority that is not there yet, weight control risk heavily and start in assistive mode, where the AI proposes and a person decides, with a record of both.
Running the scoring session
Bring the right people. The process owner, the controller, someone who knows how to get the data out, and whoever owns internal controls. Without the last two, data availability and control risk get scored optimistically.
Score independently, then compare. Have each person score alone first. Large disagreements usually mean the use case is described too loosely. "AI for AP" is not a use case. "Suggest GL codes for non-PO invoices from existing suppliers" is.
Check data with a sample. Before anything scores a 4 or 5 on data, pull one month of it and look. Remittance fields that are blank half the time and supplier names entered three different ways show up quickly.
Pick one, not three. Choose the highest-ranked use case in the first or second group that has a willing process owner. Record its current volume, hours, error rate and cycle time before anything changes, so the result can be measured later.
What a scoring framework will not do
It will not tell you whether a given tool works on your data. Two products addressing the same use case can perform very differently on your suppliers, your chart of accounts and your document quality. Only a scoped test on real data answers that.
It also gives false precision. A use case scoring 12 is not meaningfully better than one scoring 11. Use the scores to separate the obvious starts from the obvious parks, and to force a conversation about risk, then make the call on the borderline cases with judgment.
Frequently asked questions
What are the most common AI use cases in finance?
The most common, according to a Gartner release from November 2025 based on 183 CFOs and senior finance leaders, were knowledge management (49%), AP process automation (37%) and error and anomaly detection (34%). Reconciliation matching and cash application also tend to score well on volume and rules.
Which finance process should a CFO apply AI to first?
Start with a process that has high volume, clear written rules and structured data, where a person already reviews the exceptions. Invoice coding, reconciliation matching and pre-payment duplicate checks usually qualify.
How do you prioritize AI use cases in finance?
Score each use case from 1 to 5 on volume, rule-bound work, data availability, error cost and control risk. Rank by the benefit total, then use the risk total to decide how much human approval the design needs.
Should AI write board commentary?
Not as a first project. Board narrative is low in volume, has no fixed rules and carries a high cost when wrong, so it ranks low on benefit and high on risk. AI drafting with careful human editing can come later.
What makes a finance process a poor fit for AI?
Low volume, rules that live in people's heads, data that is scattered or inconsistent, and a high cost if an error goes unnoticed. Processes with these features need fixing or documenting before AI adds much.
Where CFOLogic fits
We work with CFOs to score and rank candidate use cases, then build proofs of concept scoped to one process, on top of the ERP or ledger the team already uses. Our designs read data rather than write it by default, and a person approves anything that leaves the team. This is part of our CFO Success Partners work. For a purchase-to-pay illustration built on fictional data, this demo shows invoices grouped by status and owner.