AI in Finance

Imagine this note arriving in your inbox from the finance team, with a line at the bottom saying an AI assistant drafted it in under a minute.

The note is invented, but it is the kind of thing finance teams now produce every day. Would you forward it? It is clear, confident and sounds exactly like finance. Yet you do not know what data it read. You do not know whether hosting costs really rose or an invoice simply landed in a different month. And you do not know who decided that pausing hires was a recommendation this assistant should be making at all.

That small hesitation is the whole subject of finance AI. The technology is already in the building. In Gartner’s annual survey of finance leaders, the share of finance functions using AI rose from 37 percent in 2023 to 58 percent in 2024, then held at 59 percent in 2025. Nine in ten of those users reported only low or moderate initial impact¹. The question is no longer whether finance will use AI. It is how to get the speed without losing the thing that makes finance worth listening to: numbers that someone can defend.

Fluency is cheap; authority is not

Finance leaders tend to miss in one of two directions. Too timid, and AI only polishes meeting packs while the close still takes a week. Too trusting, and a system that writes good sentences ends up changing records or releasing payments that nobody approved.

The stance between them fits in three words: draft, decide, control. AI drafts, detects and recommends. People with the authority to do so decide. And the controls that make finance trustworthy (review, approval, segregation of duties, an audit trail) travel with every use case instead of being added later. Each section below applies that pattern to one part of the finance job, and the chapter ends with the question that matters most to an executive: how much authority any given system should hold.

Finance is mostly reading, not arithmetic

Ask a room of executives what share of finance work is arithmetic and the guesses often run high. Spreadsheets solved the arithmetic decades ago. What remains is collecting information from billing, payroll and the ledger and reconciling it; interpreting what the numbers mean; forecasting; explaining why this month differs from last; assessing risk; and communicating all of it to people who must act. Almost all of that is reading, comparing, judging and writing.

That is why finance is fertile ground for AI. It also tells you which kind of AI fits which job, because “the AI” in a finance proposal can mean three quite different tools. As AI vs Machine Learning showed, generative and predictive systems do different work, and many finance problems are still best solved with plain rules.

Treat the three as different tools with different failure modes. A generative model fails by writing something plausible and wrong. A forecasting model fails by extrapolating a past that no longer holds. A detection rule fails by raising too many alarms, or too few. The controls you need depend on which failure you are buying.

Reporting: AI drafts, finance accounts

Reporting is where many finance teams start. A manager asks why support costs jumped in one region. Today an analyst searches, queries, builds a chart and writes an answer, and much of that time goes into finding things rather than thinking about them. AI shortens the finding. The workable pattern is a short chain: the question comes in, the system retrieves data from approved sources only, it drafts the analysis and commentary, and finance validates the figures against the ledger and adds what the data cannot know. Only then is anything issued.

Why so strict? Because the failure is fluent. When researchers built FinanceBench in 2023, a benchmark of questions about public-company filings, a leading model connected to a retrieval system answered 81 percent of questions incorrectly or refused to answer². Models have improved since, but the failure mode has not disappeared, and the hallucination research explains why it is built into how these systems generate text³. COSO, whose framework most companies use to design internal control, put it plainly in its February 2026 guidance: generative AI “can be confidently wrong”⁴. When the PCAOB’s staff talked to audit firms and public companies, the same message came back: invest, but keep human supervision and review of the outputs⁵.

The value is real too. In a field study of accountants on an AI-enabled platform serving 79 small and mid-sized firms, heavier AI use was associated with an 18 percent increase in weekly client support per standard deviation of use and a faster month-end close, and accountants intervened most where the AI reported low confidence in its own output⁶.

One design warning belongs here. A review step that approves hundreds of good drafts in a row trains people to stop reading. Lisanne Bainbridge called this one of the ironies of automation: the more reliable the system, the less practiced the human who must catch its rare failure⁷. Direct review effort where it counts, for example at every line above a materiality threshold, instead of asking someone to skim everything.

Forecasting: judged by the decision it changes

Forecasting is the second natural opportunity: revenue, expenses, cash, working capital, demand. The test for a better forecast is not its accuracy score. It is the answer to a simple question: what decision changes because of it? When to hire, how much inventory to hold, which investment to make, how much cash to keep on hand. A forecast that improves none of those decisions is a statistic, not decision support.

Models learn from history, and history cannot see the major customer negotiation, the planned price change or next quarter’s acquisition. So strong forecasting combines the model, explicit business assumptions and finance judgment, and records which is which. Be careful how that judgment enters. A study of more than 60,000 forecasts at four companies found that large, deliberate adjustments to statistical forecasts tended to improve them, while small tweaks often made them worse, and upward adjustments were especially likely to be wrong⁸. The lesson for a finance leader is to require a reason for every override and to watch for optimism.

AI also makes scenarios cheap. What if revenue falls ten percent, or the largest customer leaves? When a forecast moves, the executive question is why: volume, price, mix, region or a known event. If nobody can answer, you have a prediction, not decision support. How to read a forecast as a range rather than a single number is covered in AI in Prediction, Forecasting and Optimization.

Detection: more eyes, not a new judge

The third area is finance operations and controls: reconciliations, invoices, expenses and the search for errors and fraud. Here the case for AI is about time. In the 2026 edition of the largest global study of occupational fraud, the Association of Certified Fraud Examiners analyzed 2,402 cases. The typical scheme ran for 12 months before it was found, and more than half involved missing internal controls or controls that someone overrode. Speed of detection mattered enormously: frauds caught within six months had a median loss of about 40,000, while schemes that ran for more than five years cost over 1.1 million⁹.

Notice what that argues for. Tips remain the most common way fraud is found: 43 percent of the ACFE cases were uncovered by a tip⁹. Analytics and machine learning add more eyes that never tire, scanning every transaction rather than a sample. There is some evidence that this pays: in the same study, organizations using proactive data monitoring reported lower losses, an association rather than proof that the monitoring caused them⁹. The engine is usually statistics, machine learning and rules; generative AI is more useful for helping an investigator understand and explain a flagged case than for doing the flagging.

Two costs come with it. The first is false positives. Every false alarm takes an analyst’s time and can delay a supplier who did nothing wrong, and a detector that raises alarms all day destroys its own value. The second is the temptation to let the detector act. Flagging a payment for review and stopping it are different decisions, which is where the next section begins. The routing of invoices and documents through extraction and confidence checks is the subject of AI in Document and Data Processing.

Authority is a ladder you climb on purpose

Finance runs on a separation that predates computers. One person prepares, another reviews, someone with delegated authority approves, and the system executes. COSO’s internal control framework names segregation of duties as a point of focus for control activities¹⁰, and the Institute of Internal Auditors’ three lines model builds independent assurance on top of it¹¹. That separation is a large part of why auditors, boards and investors trust your numbers. AI should not quietly collapse it.

The useful way to think about any finance AI proposal is as a position on a ladder of authority.

On the first rung the system reads approved data and answers questions. On the second it drafts commentary, forecasts and flags for a person to act on; most finance AI should live here. On the third it takes small, reversible actions below a threshold, such as holding a duplicate invoice, with every action logged. The fourth rung, approving and executing consequential actions such as payments, journal entries and regulatory filings, stays with a named person who holds the authority to sign.

COSO’s 2026 guidance makes the same distinction in its own vocabulary. Among the capabilities it asks organizations to control separately are posting, where the AI writes to the books, and judgment, where it makes evaluative calls⁴. Each rung up therefore needs deliberate design: access scoped to what the task requires, transaction limits, segregation between the system that proposes and the person who approves, and a record of the input, data source, output, user, approval and time. If you cannot reconstruct what happened, you cannot defend it to an auditor. The general case for limiting what autonomous systems may do is made in Agentic AI and Autonomous Actions; in finance the ladder is simply that principle wearing a controller’s badge.

Analyze is not authorize

One picture helps keep the ladder in mind. A library card gets you into the reading room: you can read every ledger, compare every year and explain what you find. A bank signature card lets you move money. A finance assistant with read-only access holds a library card. A system with write access, one that can change a record or release a payment, holds a signature card, and its impact is in a different league.

Many proposals blur the two. A helpful analyst is demonstrated, and a few releases later it has become a payment engine. So ask which card a system holds, and treat any upgrade from one to the other as a new decision. A library card is not harmless either: it can still expose salaries or margins to the wrong person, and a misleading analysis can still drive a bad decision.

Story: the close that learned to explain itself

The following is an illustrative composite, not a documented case. It is built from the patterns above.

A software company that sells subscriptions to other businesses closes its books each month in about a week, close to the 6.4-day median that APQC’s benchmark of some 2,300 organizations reported¹². After the close, a small team of analysts writes the commentary for the leadership pack. They pull figures from the billing system, the customer database and the ledger, compare periods line by line and write the explanations by hand: why billings moved, why hosting costs rose, why one region is behind plan. By the time leaders read it, many of the decisions it should have informed have already been made. The analysts spend their time explaining the past.

The company redesigned the close, not just the tools. The system now reads only authorized data from the ledger and billing. It flags every variance above a materiality threshold that the controller sets, and drafts the first explanation with links to the underlying transactions. Analysts review every material line and add what the data cannot know, such as a pricing change or a delayed deal. The controller signs. The system holds a library card, not a signature card.

In one of the first months, the draft said billings had dropped because customers were leaving. It read perfectly. An analyst checked: no customer had left. A large multi-year contract had moved from monthly to annual invoicing. Billing timing, not churn. Had that sentence reached the leadership pack, it could have triggered a retention campaign aimed at a problem that did not exist.

The commentary now arrives earlier, and the analysts spend their time on causes and next steps. But the lesson the finance team took from the redesign was not the time saved. It was the caught error. The review step was not a courtesy. It was the design.

What this means for leaders

Finance AI rewards ambition in drafting, forecasting and detection, and caution in authority. Four lessons follow. First, place every proposal on the ladder before you discuss its features, and default to the drafting rung. Second, ask which of the three tools is actually in use, because the failure you must control depends on it. Third, judge forecasting by the decisions it changes and detection by how much faster it finds problems, not by accuracy claims in a demonstration. Fourth, budget for review as part of the system, aimed at what is material, rather than as an afterthought that slowly turns into a rubber stamp.

Check yourself

1. Finance is rules-based, so most of it can be automated end to end.MythFact
2. If an AI system is read-only, it carries no risk.MythFact
3. A forecast should be judged mainly by the decisions it improves.MythFact
4. Small judgmental tweaks to a statistical forecast reliably improve it.MythFact
5. Fraud caught early costs far less than fraud caught late.MythFact
6. An AI tool whose errors people can fix will always save money.MythFact

Reflection: where does your line sit?

What comes next

In finance, a wrong answer costs money and credibility, and most of it can be reversed. The next chapter, AI in Human Resources, turns to a function where AI touches decisions about people’s jobs and careers, and where the cost of a confident mistake falls on an employee rather than a ledger.