Blog
Written by
Naman Mathur
Published on
Key takeaways
Generative AI is ready for the close but under-used there: match proposals, journal drafts, flux commentary and remittance reading all run in production with a named reviewer.
We ran three general-purpose AI agents on eight real reconciliation datasets 10 times each. On the hardest, one agent's wrong matches varied between 7 and 302 by run.
Treat every AI output as a control: a named reviewer, a logged decision and evidence your auditor can test.
Plan the next 12 months by quarter, starting with bank reconciliations and ending with intercompany, and measure each step against your baseline close days.
Most finance teams have put AI to work everywhere except the close. Gartner's 2025 survey of 183 chief financial officers (CFOs) and senior finance leaders found knowledge management (49%) and accounts payable (AP) automation (37%) were the most common uses, and only 7% reported a high or very high impact. In our experience, the close is where impact is easiest to prove, because the work recurs every period and has an objective right answer.
On a broader Gartner measure that counts plans as well as live use, 84% of finance organisations have implemented or are planning to implement AI. The decision in front of a controller this year is which close tasks to put into production and which to keep in pilot. The standard advice, to roll out a general assistant and see what sticks, gets that order backwards.
This article answers how generative AI is used in accounting and finance today, where it is mature and where it is still experimental, what the risks are, and what to automate in the next 12 months.
How Is Generative AI Being Used in Accounting and Finance Operations Today?
Generative AI in accounting and finance is used mainly to draft and propose work on top of ledger data. It writes variance commentary, drafts journal entries, proposes transaction matches and reads remittance advice. Deterministic systems still calculate the numbers, and an accountant approves each output before it reaches the ledger.
Generative AI is software, usually built on a large language model (LLM), that produces new text or structured data from a prompt. Agentic AI adds an agent that plans steps, runs enterprise resource planning (ERP) queries and finishes a task for a human to review.
Where Finance Teams Use AI Most Today
Adoption of AI in accounting and finance is broad but has stalled. The Gartner 2025 finance AI survey of 183 finance leaders found 59% reporting the use of AI in their finance function. That share jumped from 37% in 2023 to 58% in 2024, then barely moved.
Gartner ranks error and anomaly detection third, at 34% of adopters.
The close is missing from that top three, and that is the opportunity. Gartner notes in the same release that the biggest impacts arrive once AI reaches the production stage. Close tasks recur every period and have an objective right answer, which makes them the shortest route from pilot to production.
Where Generative AI Is Mature in the Close
Mature means the use case runs every period with a named reviewer. The audit trail shows what the AI proposed and what the reviewer changed. Four close tasks meet that bar today.
Use case | What AI does | What stays human | Maturity |
|---|---|---|---|
Reconciliation and matching | Proposes matches across bank statements and subledgers | Investigating and approving the unmatched remainder | In production at high volume |
Journal entries | Drafts recurring entries against policy and routes them for approval | Approval before posting | In production for recurring entries |
Flux analysis | Drafts variance commentary that cites transaction IDs | Business context and sign-off | In production, with team-added context |
Cash application | Extracts remittance data and matches payments to open invoices | Short pays and unclear payers | In production for clean payments |
Source: Stacks implementation practice, 2026.
Reconciliation and Transaction Matching
In our implementations, matching is the most established production use. An agent reads CAMT.053 (the ISO 20022 bank-to-customer statement) or MT940 (the older SWIFT statement format) bank statements alongside subledgers such as Stripe, Adyen, Chargebee or the accounts receivable (AR) ledger. It then proposes matches for a preparer to accept.
With automated account reconciliations, the agent matches clean items and routes the rest to an exception queue with its reasoning attached. Our case study on how Volt reached 97% auto-matching reports over 95% matched in the first close and 97% by the third, by January 2025 [VERIFY: January 2025 date for Volt's third close against the Volt case study]. Its monthly close fell by three days (33%).
The audit question sits in the remainder. Your auditor will test how the unmatched items were investigated and approved, so the exception queue needs the same rigour as the matches.
Journal Entry Drafting and Posting
With journal entry preparation agents, recurring entries such as prepayment releases and reclassifications are drafted against your accounting policy. Each draft goes to an approver before it posts, which keeps segregation of duties intact.
At Volt, the Excel Assistant reduced manual journal steps from 37 clicks to a few. That matters most on high-volume recurring entries, where every click is a chance to post to the wrong account.
Flux Analysis and Variance Narratives
Flux commentary is the most genuinely generative use in the close. Explaining a variance means tracing it to the transactions behind it before anyone writes a word. On its AI flux analysis product page, Stacks says its Flux AI Agent removes "1–3 days of manual prep work and narrative writing".
The agent queries the ERP deterministically, then drafts commentary that cites the transaction IDs behind each movement. Supermetrics runs 17 flux reports a month this way. The team reviews each draft and adds the business context a ledger cannot supply.
Cash Application
Cash application is where generative AI reads unstructured documents. With automated payment application, an agent extracts remittance details from PDFs and emails and matches each payment to open invoices.
Exceptions, such as a short pay or one transfer covering six invoices with no remittance, reach the AR team with the agent's reasoning attached. When you see a demo, ask for the exception view first, because that is where your team will spend its time.
Where Generative AI Is Still Experimental
Generative AI in finance and accounting is still experimental wherever the answer depends on judgement or the model works without deterministic data underneath. That covers judgemental accruals and estimates, technical positions under International Financial Reporting Standards (IFRS) 15 and 16, consolidation judgements and forecast narrative without reconciled numbers. Fully autonomous posting with no reviewer belongs on the same list.
The reason is consistency. In our transaction matching test, we ran three general-purpose AI agents on eight real reconciliation datasets, 10 times each. On the hardest dataset, the same agent produced between 7 and 302 wrong matches depending on the run.
No controller would trust a preparer with that spread on the bank reconciliation. The practical rule is to let the LLM propose and explain, and let deterministic logic decide whether two items match.
What Are the Risks of Generative AI in Finance?
The main risk is a plausible wrong number: output that reads well, ties to nothing and changes between runs. The answer is to treat every AI output as a control, with an evidence trail and a named person who signs off.
The Financial Reporting Council (FRC) has set out the same risks for auditors. Its FRC generative and agentic AI guidance, published in March 2026, lists "Hallucinations – information that has been fabricated by the LLM" among output deficiencies. Omissions, distortions, faulty reasoning and inconsistencies sit on the same list.
The guidance is written for audit firms' technical teams, but it works as a checklist for your own controls. Its mitigations include human review and rules-based checks, and it warns that giving an orchestrating agent more autonomy may carry greater risk.
For EU entities, the EU AI Omnibus timelines confirm that the AI Act's Annex III high-risk rules apply from 2 December 2027. Close automation isn't among the Annex III use cases, but groups with EU subsidiaries should confirm scope with counsel for anything touching employees or individual customers' credit. [VERIFY: legal review of Annex III wording]
Before your next audit, document each AI-assisted task: which steps are deterministic, who reviews the output, how rejections are logged and where the evidence sits.
Why "Start With a Chatbot" Is the Wrong First Step
The standard advice is to roll out a general-purpose assistant, run a few pilots and let use cases emerge, which produces plenty of activity and, in our view, explains much of Gartner's 7% figure.
A chatbot answers questions, which is knowledge management, the top use case in Gartner's survey. Searching the policy manual faster does not shorten a close. Impact shows up when AI does recurring, high-volume work in production.
Pilots also fail for a structural reason. A pilot run beside the real close rarely has a named reviewer or an evidence trail, so it never proves the one thing your auditor will test.
Start with work that recurs every period and has an objective right answer, such as bank reconciliations and recurring journals. Teams that tried robotic process automation (RPA) will recognise why: rules-based bots stall on exceptions. Our piece on agentic AI versus RPA covers what changes when software can work an exception and explain it.
What Should a Finance Leader Plan to Automate in the Next 12 Months?
Plan the year in quarters, starting with high-volume work that has an objective answer and finishing with tasks that span entities. Measure your close in calendar days before you start, so every quarter has a number to beat.
The American Productivity & Quality Center (APQC) month-end close benchmark puts the median at 6 calendar days, with the fastest at 5 and the slowest at 10 (N=10,201). APQC measures from the initial trial balance to agreed consolidated statements, weekends included, so measure yours the same way.
Quarter | What to automate | What to measure |
|---|---|---|
Quarter 1 | Baseline close days, then bank reconciliations | Calendar close days and auto-match rate |
Quarter 2 | Recurring journals and cash application | Hours spent posting journals, share of payments applied untouched |
Quarter 3 | Flux narratives | Share of general ledger (GL) accounts with drafted explanations, reviewer edit rate |
Quarter 4 | Intercompany and close orchestration | Intercompany differences at close, tasks finished on schedule |
Source: Stacks implementation practice, 2026.
The order matters: each quarter builds on data reconciled in the one before. When you evaluate vendors, ask four questions:
Which steps are deterministic, and which does a model generate?
How are rejected proposals recorded, and can your auditor see them?
What does the reviewer see when the agent is unsure?
Does it connect to your ERP and subledgers directly, or rely on exports?
Why Choose Stacks for Generative AI in the Close
Stacks builds agentic AI for the enterprise close, with deterministic ledger data underneath and agent-drafted work on top. Agents pull live data from your ERP and prepare the work as it arrives. Your team reviews the exceptions and signs off.
The platform runs close management workflows with preparer and reviewer roles across a multi-entity scope. Alongside them sit account reconciliations, transaction matching, the Payment Application Hub [VERIFY: product name], AI flux analysis and the Excel Assistant.
Stacks connects natively to NetSuite and [VERIFY: other native ERP connectors, e.g. Xero, Sage, Microsoft Dynamics 365 Business Central], so there is nothing to migrate. It holds System and Organization Controls (SOC) 2 Type II and ISO 27001:2022 certification and complies with the General Data Protection Regulation (GDPR). It also enforces segregation of duties, with audit logs on every action.
FAQs for Generative AI in Accounting
Will Generative AI Replace Accountants?
No. It takes over preparation such as matching, journal drafting, first-pass commentary and remittance reading, while accountants keep the judgement and the sign-off.
What Type of AI Is Used in Accounting?
Accounting tools combine machine learning for matching, optical character recognition (OCR) for documents, rules engines for deterministic checks and LLMs for drafting text. Agentic AI joins these into multi-step tasks with a human review point.
Can Generative AI Speed Up the Month-End Close?
It can when it runs in production on recurring close work. Nivoda's close fell from more than 20 days to 12–16 days, most recently 12 in [VERIFY: month and year].
Is Generative AI Accurate Enough for Reconciliations?
Yes, when the model proposes and explains while deterministic logic makes the final match. Left to match on its own, a general-purpose AI agent gives different answers on different runs.
What Are the Biggest Implementation Challenges?
Gartner attributes the 2025 adoption plateau to complexity, data and talent challenges. In practice, the agent needs live connections to the ERP and to subledgers such as Stripe, Adyen, Coupa or payroll.
See How Stacks Maps to Your Close Process
Book a 30-minute walkthrough. We'll discuss your close process, show you which tasks run automatically, and outline the next step from there.
"Stacks has transformed how our finance team operates... it's saved us time and reduced frustration."
Graham B., SVP of Finance at Volt
Trusted by fast-growing companies including: Datavant, Future, Cleo, Motorway, Nivoda, Yoto and Epidemic Sound.
