AI for accounting stopped growing. Here's what works.

Ninety-two percent of accountants already use AI, says a survey Karbon ran on its own audience. Agentic finance adoption is up six hundred percent, says one from Wolters Kluwer. The headline numbers on AI for accounting all point up and to the right, and the ones that get quoted hardest were collected by companies selling accounting software to the people they surveyed.

Gartner's 2025 AI in Finance survey asked 183 CFOs and senior finance leaders the same question three years running. Thirty-seven percent used AI in the finance function in 2023. Fifty-eight percent in 2024. Fifty-nine percent in 2025.

That is not an explosion. That is a curve that climbed hard for a year and then went flat, which is a much more useful fact if you run a finance team and are deciding what to do next quarter. Something stopped the second wave, and it wasn't interest.

What can AI for accounting actually do right now?

AI for accounting reliably does three things: it reads documents and pulls structured data off them, it compares two sets of records and flags what doesn't match, and it answers questions about where a number came from. Everything else on the vendor slide is either a wrapper around those three or still a demo.

That maps to what finance leaders report using it for. In Gartner's 2025 survey the top uses were knowledge management at 49%, accounts payable automation at 37%, and error or anomaly detection at 34%. Document reading, matching, and lookup. In that order.

Notice what is not on that list. Nobody's AI is closing the books, setting accruals, or making a judgment call on revenue recognition. AI for accounting absorbs the work that sits underneath the judgment, and there is a great deal of it.

The tasks where it holds up

• Invoice and receipt capture: pulling vendor, date, amount, PO number, and line items off a PDF that arrives in a different layout from every supplier.
• Two and three-way matching: invoice against purchase order against goods receipt, with the mismatches routed to a human instead of the matches routed to one.
• Bank and account reconciliation: proposing matches on transactions that don't tie cleanly by reference number.
• Anomaly flags: a duplicate invoice, a vendor bank detail that changed last week, an expense coded to a department that has never used it.
• Answering "why is this number what it is" by tracing the entry back through the documents behind it.

The tasks where it doesn't

Anything requiring an estimate a person has to defend. Accrual judgments, reserve levels, revenue cut-off calls on a messy contract, anything an auditor will want reasoning for. AI can assemble the evidence for those decisions faster than any analyst. It cannot own them, and no finance leader should let it try.

Why did finance AI adoption stop growing?

Adoption plateaued because the easy half of the market already bought, and the second half hit the same wall in the same place: the tool worked in the pilot and never became a process anyone depended on. Interest was never the constraint. In Deloitte's Q4 2025 CFO Signals survey of 200 North American finance chiefs, 87% said AI will be extremely or very important to their finance department in 2026, and only 2% said it won't matter.

So you have near-universal intent for AI for accounting sitting on top of an adoption number that moved one point in a year. That gap is the whole story of this category, and we've written about the same pattern showing up across functions in the AI adoption gap.

The reason it bites finance harder than most is that finance work is periodic. A marketing automation that breaks gets noticed the same afternoon. A close routine that breaks gets noticed at the next close, three weeks later, after it has already written wrong numbers into a ledger somebody signed.

What stalls it

The most common failure is an ownership one. Someone has to retrain the extraction model when a supplier redesigns its invoice, adjust the matching rules when procurement changes a PO format, and re-authenticate the ERP connection after IT rotates credentials. Nobody is hired for that, so it lands on the controller, on top of the close.

The second is scope. Vendors demo on tidy invoices, and production runs at an 18.4% exception rate in Ardent Partners' 2025 benchmark. Exceptions are where the finance labor actually lives, so a pilot that skips them proves very little.

The third has nothing to do with the model at all. Nobody could show the auditor a control, so the project got switched off by someone who never evaluated the technology.

Where AI for accounting pays for itself.

Accounts payable is where the unit economics are legible enough to put in a board deck. Ardent Partners' 2025 State of ePayables research, based on 204 AP professionals, puts the average all-in cost of processing a single invoice at $9.84, up from $9.40 in the prior edition, with an average processing time of 8.2 days. Worth knowing that this research is sponsored by AP software vendors, which is exactly the caveat this article opened with. It survives the objection because Ardent publishes its sample size and its method, and the same benchmark has run for years on consistent definitions.

Best-in-class AP teams in the same benchmark run 79% below that on cost and 79% faster on cycle time, with exception rates 47% lower than everyone else. Run those numbers against your own invoice volume and you have a payback calculation that survives contact with a CFO.

The second job is reconciliation, where the value sits entirely in the tail. NetSuite or Sage Intacct already matches everything that ties cleanly by reference number. What eats an analyst's evening is the remainder: a truncated memo field, a two-day settlement lag, a customer who paid three invoices with one wire and rounded down.

The third is document lookup, the highest-reported use in Gartner's survey and the least discussed. When an auditor asks for support on a journal entry from March, someone currently goes digging. That retrieval is a genuine time sink and it automates cleanly, because being wrong is cheap: the human sees the source document and judges it.

A word on the AP ceiling, because it gets oversold. Touchless processing rates stall well short of universal, and we broke down why in AI accounts payable automation and the 33% ceiling. Plan for the exceptions to stay, and staff the design around them.

What does AI for accounting cost to keep running?

Price it per transaction, not per seat. At Ardent Partners' benchmark of $9.84 to process an invoice against roughly $2.07 for best-in-class performers, a team running 3,000 invoices a month carries roughly $29,500 in monthly processing cost, of which about $23,000 is the theoretical gap to best-in-class. That is the number worth chasing, and it is also the number every vendor quotes back at you.

Gartner's 2025 forecast has better than two in five agentic AI projects scrapped before 2027 is out. Cost creep, business value nobody could pin down, and risk controls that didn't hold are the three reasons it names. The same analysis put the number of vendors doing real agentic work at roughly 130, out of thousands making the claim. Most of what is marketed to finance teams under that label is workflow software with a new sticker.

What that calculation leaves out is the running cost, and it is the line nobody puts in the business case. A subscription you can approve and an integration you can scope are both bounded numbers. Maintenance is not: every Coupa release, every Bill.com API change, every supplier that redesigns its invoice template puts hours back on somebody's calendar, and that somebody is usually the controller who is also closing the month. That unpriced line is what turns a working pilot into an abandoned one, and it's the same dynamic we costed out across functions in the hidden cost of low AI adoption.

The staffing math makes it worse rather than better. AICPA's 2025 trends data, reported in the Journal of Accountancy, shows 55,152 accounting degrees awarded in 2023-24, down 6.6% year over year, with master's programs off about 15%. Spring 2025 enrollment rose 12.4%, so the pipeline is refilling, but those students are four to six years from being useful to your close. The people you have now are the people you have.

Your auditor will ask who approved it.

Before any AI touches your ledger, decide who signs off on its output, where that approval is recorded, and how you'll reproduce the decision eighteen months later. Vendor risk sections talk about bias and black boxes. Your auditor will ask which ITGC the thing falls under and who owns the approval.

A named human approves anything that posts. The AI drafts the entry, proposes the match, codes the expense. A person with the authority approves it. That approval is the control, and it has to be a click somebody makes, not a setting somebody toggled once in March.

The evidence trail is captured as the work happens, not reconstructed afterwards. For every decision your reviewer will want the source document, what the model pulled off it, what it proposed, who approved it, and when. This is what turns into your PBC list. If you are assembling it from screenshots in week two of fieldwork, it was never a control.

Segregation of duties applies to software too. An agent that can both create a vendor record and approve a payment to it is a fraud path with better uptime than a human one. Split those permissions the way you would split them between two employees, and expect a SOC 1 Type II reviewer to test exactly that.

If you're a SOX filer, none of this is new work. It's the control environment you already run under 404, extended to a participant that reads invoices faster and never gets bored. If you're not, your first serious audit or diligence process will ask for it anyway, and building it in later costs more than building it in now.

How should a 50-500 person finance team start?

Pick one routine that runs on a schedule, costs a known number of hours, and has an obvious right answer. Instrument it, run it alongside the manual process for one full cycle, and compare. Then decide.

Concretely, that means invoice capture and coding before month-end close, and bank reconciliation before anything touching revenue recognition. The tell for a good first candidate is that a new hire could be taught it in an afternoon and would still find it dull in month three.

Then answer the question the vendor demo skips: when this breaks in March, who fixes it? If the honest answer is your controller, you have bought a second job for the person with the least slack in the building.

That question is the reason Uplift works the way it does. You describe the routine in the words you'd use to brief a new analyst, and we build the agent, run it, and keep it running as NetSuite ships a release, a supplier changes its invoice layout, or a bank feed moves. Nobody on your team opens a canvas, wires a node, or gets paged when a vendor changes an invoice template. You can see how that plays out per function on our team breakdown.

The teams still getting value from AI for accounting in year two aren't the ones that picked better software. They're the ones who never made the close calendar depend on somebody finding time to maintain it.