AI Use Cases in Corporate Finance: FP&A, Accounting, Treasury and Tax

By Jane Ha · Updated Oct 11, 2026

AI use cases in corporate finance are the specific, repeatable jobs inside the finance function where machine learning or generative AI can take over reading, matching, forecasting, drafting or flagging, while finance staff keep approval and judgment. They aren't one big project. They're a list of small ones, each with its own data, its own accuracy bar and its own level of risk.

This page is for CFOs, controllers, finance directors and finance managers who've been handed a long vendor list and need to decide what to try first. It maps the use cases by function (FP&A, accounting and close, payables and receivables, treasury, tax, audit and controls) and ends with a way to rank them. If you want the build-level detail on specific agents, see AI agents for finance. This page stays one level up, at the choice of where to start.

Key Facts

• In Deloitte's Q2 2026 North American CFO Signals survey of 200 CFOs at organizations with at least US$1 billion in revenue, 51% used AI for operational productivity tasks, 44% for financial planning and budgeting, and 41% to analyze financial data for insights (Deloitte CFO Signals, Q2 2026).
• In the 2026 AFP Treasury Benchmarking Survey (425 treasury practitioners, fielded May 2026), cash and liquidity forecasting was named the most challenging treasury activity by 49% of respondents, and treasury policies on AI scored 2.9 out of 5, the lowest of any policy area measured (AFP).
• Thomson Reuters' 2025 corporate tax report found 43% of respondents had used AI for their own work, but only 6% described their wider departments as active users (Thomson Reuters Institute).
• PCAOB staff outreach in 2024 found audit firms using generative AI mainly for administrative and research work, while some preparers used it for less complex, repetitive tasks such as preparing account reconciliations (PCAOB staff Spotlight, July 2024).

Why Finance Needs a Use-Case Map, Not a Tool List

Finance is a good fit for AI because so much of the work is structured, rule-bound and high-volume. It's also a poor place for sloppy experiments, because the output ends up in the books, in a bank instruction or in a filing. A use-case map helps because the same word, "AI," covers very different jobs:

• Reading documents. Pulling fields from invoices, contracts and bank statements. It's closely related to RPA but can handle messy layouts.
• Matching and flagging. Reconciling two data sets, or spotting the transaction that doesn't fit, which is anomaly detection.
• Forecasting. Predicting revenue, cash or collections from history, which is predictive analytics.
• Drafting and answering. Writing variance commentary or answering a policy question from your own documents, which is generative AI.

Each type fails differently. A forecast can be off by a few percent. An extraction can misread a digit. A draft can state something false with total confidence, the problem known as AI hallucination. Knowing which type a use case belongs to tells you what to test.

The Use-Case Map by Finance Function

The table below is the core of this page. "Maturity" is an editorial rating of how well-proven the use case is in finance teams today (High: widely deployed, Medium: proven but uneven, Emerging: mostly pilots). "Risk" is the cost of an undetected error.

Two patterns stand out. The "High" maturity rows are almost all extraction, matching or drafting-for-review: work where a person checks the output in the normal flow. The "High" risk rows are the ones where an error leaves the building, into a payment, a filing or an audit opinion.

Function by Function

FP&A

FP&A is where finance leaders report real use. In Deloitte's survey, 44% of CFOs used AI for planning and budgeting and 41% to analyze financial data for insights. The easiest win is variance commentary. The tool drafts "revenue missed by 3% because of delayed enterprise renewals" from the actuals and driver notes, and the analyst corrects it instead of writing from scratch.

Forecasting is harder. A model can add value on volume-driven lines (subscription renewals, headcount costs, seasonal demand), but it needs a history that hasn't been rewritten by reorganizations and account remaps. Treat it as a second opinion next to the human forecast for a few cycles before you rely on it.

Accounting and close

Close work is repetitive and deadline-driven, which makes it attractive. PCAOB staff noted that some preparers were using generative AI for less complex, repetitive processes such as preparing account reconciliations or identifying reconciling items. Matching bank lines to ledger entries, suggesting the likely cause of a difference, and flagging journal entries posted at odd times or by unusual users are the common starting points.

The control matters more than the model. A reconciliation an AI prepared still needs a reviewer, and the reviewer needs to see why the match was made. If you can't trace it, you can't sign off on it.

Accounts payable and receivable

Payables is usually the first stop because the volume is high and the checks are well defined: read the invoice, match it to the purchase order and the receipt, code it, route exceptions. The risk is paying the wrong amount or paying twice, so the design that works keeps exceptions with people and puts duplicate checks in front of the payment run. The vision-extract pattern explains how document reading behaves in practice.

Receivables splits in two. Cash application (matching incoming payments to open invoices) is mature and low drama. Collections prioritization, deciding whom to call first and what to say, is a prediction problem that rewards a clean payment history and consistent customer records.

Treasury

Treasury has the clearest need and the hardest bar. The 2026 AFP survey found cash and liquidity forecasting is both the top treasury priority (63%) and the most challenging activity (49%). AI is a natural fit because the inputs (bank balances, receivable and payable schedules, payroll calendars) are numeric and recurring. But a forecast that drives a funding decision has to be explainable, and it has to be backtested against what actually happened.

The same survey is a caution on readiness. Treasury policies on AI scored 2.9 out of 5, the lowest of any policy area, and 34% of respondents considered themselves effective in AI knowledge versus 50% who said it was important. Write the policy before you scale the tool.

Tax

Tax teams are enthusiastic individually and cautious as departments. Thomson Reuters found 43% of corporate tax respondents had used AI themselves, yet only 6% called their departments active users, and more than half rated their department's technology posture as chaotic or reactive. Practical uses are research support, summarizing rulings and notices, extracting data from supporting documents and classifying transactions for indirect tax. Because an error can land in a filing, keep a qualified preparer as the owner of every output, and keep sensitive taxpayer data inside approved tools.

Audit and controls

PCAOB staff found audit firms using generative AI mainly for administrative and research activities in 2024, with other uses seen as potential rather than routine. For internal audit and controls teams, the strongest use is testing the full population of transactions against a control instead of a sample, with exceptions listed and evidence attached. Policy and evidence summarizing is a second, newer use. In both, the PCAOB's observation applies: the person who uses the tool remains responsible for the result and the documentation. An AI tool doesn't dilute accountability.

Where to Start: A Simple Prioritization

Rank each candidate on three questions, each scored 1 to 3:

1. Value. How many hours or dollars does this touch each month?
2. Data readiness. Is the data complete, consistent and accessible without a clean-up project? The data readiness for AI guide covers the checks.
3. Error tolerance. If the output is wrong and nobody notices, how bad is it? Score 3 for low impact, 1 for high.

Add the scores. The best first projects score high on all three. In most finance teams that points to the same cluster:

• First wave: variance commentary drafting, invoice capture and coding, bank reconciliation matching, cash application. High value, structured data, a human reviewer already in the flow.
• Second wave: journal entry anomaly flagging, collections prioritization, driver-based forecasting, full-population control testing. These need cleaner history and more tuning, and they reward a first wave that already fixed the data.
• Later, with governance first: cash forecasting that drives funding decisions, tax research feeding filings, anything that initiates a payment. Not because they're off limits, but because the cost of an undetected error is high.

The choosing the right AI pattern guide gives the same logic from the technology side.

How to Run a Finance Use Case, Step by Step

1. Write the process as it runs today. Inputs, steps, approvers and the failure you see most often. If nobody can describe it, it isn't ready.
2. Set a baseline. Measure cycle time, touch time and error rate now, so you can show a change later.
3. Check the data. Pull a sample and look for missing fields, duplicate vendors, inconsistent account codes and stale master data.
4. Choose the human checkpoint. Name who reviews, what they see and what they can override. This is human-in-the-loop design, and it should be a real step with evidence, not a click-through.
5. Run in parallel first. Let the tool produce its output next to the manual process for a close cycle or two and compare results line by line. Some preparers told the PCAOB they ran new technology in a parallel environment before deployment, and the same discipline suits finance.
6. Limit permissions. Give the tool read access where read access is enough. Segregation of duties still applies: the tool that prepares shouldn't also approve.
7. Log everything. Keep the inputs, outputs and reviewer actions so an auditor can follow a transaction end to end.
8. Review and expand. After a few cycles, compare against the baseline, fix the data issues you found, and only then pick the next use case.

Common Pitfalls

Starting with the glamorous use case. Autonomous forecasting makes a good board slide. Invoice coding makes a better first project, because the volume is high and errors are easy to see.

Confusing a demo with your data. A tool trained on tidy invoices can stumble on your suppliers' scanned PDFs. Test on your worst month, not your best.

No owner for the output. If the AI drafts a flux explanation and no one is named as accountable, it goes into the close pack unread. Every use case needs a named person.

Skipping controls. Finance runs on segregation of duties, approvals and audit trails. A tool that bypasses them, even by accident, creates a finding. Add AI use cases to your AI risk register the way you'd add any new system touching the ledger.

Leaving policy for later. The AFP result is a warning here. Decide what data may go into which tools, who can approve a new use case and how errors are reported before adoption spreads informally.

Measuring only speed. Hours saved is easy to count. Track errors caught, reviewer corrections and exceptions too, or you'll miss a tool that's fast and wrong.

What to Read Next

If you're deciding between automation approaches, AI vs RPA in finance explains which jobs need a model and which just need a rule. For the groundwork, read finance data readiness for AI, and for the big picture, start with AI in finance. Pick one first-wave use case, name the reviewer, set the baseline and run it in parallel for a cycle. That's a better first step than any strategy deck.