Corporate finance has acquired a curious enthusiasm for artificial intelligence.
CFOs who normally scrutinize every spending request are now scrambling to make room for AI in their budgets. The excitement is understandable. The results, rather less so.
In June, Gartner reported that 84% of finance organizations had implemented or were planning to implement AI. A mere 7% reported a high or very high impact. For a profession that prides itself on measuring returns, this is an awkward set of figures.
Nor is enthusiasm in short supply. In Deloitte’s fourth-quarter 2025 survey, 87% of CFOs considered AI very or extremely important to their departments’ operations in 2026. Perhaps there’s a little corporate FOMO at play, with the fear of falling behind competitors proving more persuasive than the returns themselves.
But accounting is a less accommodating playground for experimental technology than, say, marketing or customer service. An AI-generated advertisement that misses the mark can be discarded. A journal entry that hallucinates an expense is a more troublesome proposition.
Still, caution can explain only so much. Companies have spent decades acquiring software to automate accounting, only to retain an impressive appetite for hiring people to keep the machinery running.
Finance departments have become accomplished at managing complexity. Whether they’ve become equally good at eliminating it is another question.
The Accountant’s Stubborn Workload
Few professions have been promised liberation from drudgery quite as frequently as accounting.
Each generation of financial technology has arrived with assurances that the paperwork would finally disappear. Much to the disappointment of automation’s prophets, accountants have continued to find plenty to do.
Consider the month-end close, the venerable ritual of finalizing the books. APQC’s research, covering more than 11,000 organizations, puts the median time required to complete monthly consolidated financial statements at six calendar days.
That is six days to establish what happened the previous month, despite much of the activity having already been recorded electronically.
Much of that time goes into work automation leaves unfinished. It could be a payment that refuses to match an invoice, an expense recorded in the wrong period, or two subsidiaries with conflicting intercompany balances. Each discrepancy needs to be investigated and resolved before the books can be closed.
For businesses with multiple entities, financial consolidation adds another headache, as accountants must reconcile different currencies, accounting treatments and intercompany balances.
Don’t get me wrong, some of this work is unavoidable. Financial statements require scrutiny, and no sensible controller would sacrifice accuracy for a faster close.
Yet there’s something faintly absurd about employing highly trained accountants to spend their days hunting down information that’s already sitting somewhere in the company’s systems.
For all its sophistication, “modern” accounting software still leaves an awful lot for humans to figure out.
A different kind of accounting assistant
AI agents offer a plausible way out of this predicament.
Unlike conventional automation, which generally follows a prescribed sequence, agents can investigate transactions, consult supporting records, and decide which permitted step to take next.
Suppose a customer pays three invoices in one transaction, deducting a credit note mistakenly applied to another account. The payment doesn’t match the outstanding invoices, leaving an accountant to sift through the records. An AI agent could investigate, gather the evidence, and propose a correction for approval.
Such work is hardly glamorous, but neither is it cheap. Every hour spent investigating discrepancies is an hour of skilled accounting labor, and reducing that burden could save companies considerable time and money.
Some companies are already showing what’s possible. Consider Graphite, a technology company whose finance team once spent ~20 days closing the books. Its finance leader, Rafael Tihanyi, had concluded that even three additional hires wouldn’t solve the problem. After switching to DualEntry, an agentic ERP startup, Graphite brought its close down to a single day, without the additional hires.
The problem with giving machines the keys
AI’s occasional flights of fancy have already landed professionals in hot water.
In 2023, two New York lawyers were sanctioned after submitting fictitious court cases generated by ChatGPT. The cases came complete with convincing legal citations. Unfortunately for the lawyers, none of them existed.
The consequences are less entertaining in corporate accounting. While a fabricated legal citation can embarrass a lawyer, an improperly posted accrual can distort reported profits and mislead investors.
The evidence suggests even today’s most capable models deserve a healthy dose of skepticism. DualEntry’s Accounting AI Benchmark evaluates leading models across 101 accounting tasks, from transaction classification to accounts payable, reconciliation and month-end close.
As has been seen over time, performance varies considerably across these activities, making blanket claims about AI’s accounting abilities difficult to defend.
An agent might perform admirably on routine classifications yet stumble over an unusual customer payment or month-end adjustment. Treating both tasks as equally reliable would be an odd way to run a finance department.
Then there’s the question of oversight. Asking accountants to scrutinize every automated decision defeats the purpose of automation, especially when reviewing the work takes almost as long as doing it. Yet handing agents unrestricted authority over the ledger would be an equally dubious proposition.
The sensible approach is to give agents room to investigate discrepancies, gather evidence and prepare adjustments, while keeping consequential decisions subject to established accounting controls.
As Santiago Nestares, co-founder of DualEntry said: “If an agent has enough context to do the work itself, it can handle that in the background, tee up the result, and bring you in when judgment or approval is actually needed.”
An agent might recommend a journal entry, for instance, but posting permissions, approval requirements and period locks should determine whether it ever reaches the books.
This is where the underlying financial system becomes especially important. Agents need reliable records, enforceable permissions and an audit trail that explains how each decision was reached.
The emerging agentic ERP approach brings agents into the accounting system itself, where their work can be governed by the same controls that apply to human accountants.
Finance gets a taste of its own medicine
The prospect of running a larger business without continually expanding the accounting department ought to interest plenty of CFOs.
Growth has traditionally meant more financial complexity and, sooner or later, more people to manage it. But AI is already making that old assumption look increasingly shaky.
That doesn’t mean accountants are heading for extinction, far from it. Expanding internationally still brings unfamiliar tax rules and reporting obligations, while financial judgment remains stubbornly difficult to automate.
But companies may become less willing to accept rising finance headcount as the inevitable price of growth.
It may also be time to reconsider what makes a productive finance department. Closing the books quickly is one measure, but how much human effort does the process consume?
A controller overseeing thousands of transactions with relatively little manual intervention may be running a more efficient operation than one managing a larger team through the same workload.
Of course, the history of enterprise technology is littered with promises that failed to survive contact with reality. AI may disappoint in ways its most enthusiastic advocates have yet to imagine.
But the possibility of managing considerably more financial activity without continually expanding the payroll is becoming difficult to ignore.
