Agentic AI in Accounting: Capabilities, Limitations, and the Path to Adoption

Countsy, a California-based firm that provides outsourced accounting and CFO services to hundreds of venture-backed start-ups, processes up to 3,500 invoices a month, a workload that illustrates the growing relevance of agentic AI in accounting. Until recently, each one required roughly five minutes of human attention. An accountant would open the PDF, read the vendor name and line items, look up the vendor in NetSuite, assign the correct GL codes, check for a matching purchase order, and route the invoice for approval. Multiply that by 3,500 and the firm was spending approximately 300 hours a month on data entry alone.

Today, 78% of those invoices are processed without anyone on Countsy’s team reviewing them. The firm adopted Vic.ai, an AI-powered accounts payable platform, and the time savings have been redirected into higher-value client work: performance analysis, KPI dashboards, advisory. Processing time per invoice fell from five minutes to under two. Coding accuracy reached 95%. Countsy’s founder, Mairtini NiDhomhnaill, described the shift as structural:

“Implementing autonomous AP has made us more efficient and provides us comfort as our clients grow their businesses quickly and increase invoice volume. We’re now able to handle an influx of invoices without adding tons of manual work to our accounting staff’s plate.”

This is agentic AI in accounting in practice. Not a chatbot that answers questions about lease accounting standards. Not a macro that auto-populates last year’s numbers. Something fundamentally different: software that can take a pile of unprocessed invoices and work through them the way a competent AP clerk would, making judgment calls about coding, flagging what looks wrong, and moving clean transactions through to posting, without being told what to do at each step.

What Makes It Different From Previous Automation

Most accountants have used some form of automation. Rules-based tools like robotic process automation follow scripts: if an invoice arrives from a known vendor, apply this GL code and route to this approver. These systems do exactly what they are programmed to do. They cannot handle an invoice they have not seen before, interpret an unfamiliar line item, or decide that something looks unusual. When data does not fit the script, it drops into an exception queue for a human to resolve. Research from Ardent Partners found that 62% of AP professionals say these exceptions are their biggest daily headache.

Generative AI added language comprehension. Tools like ChatGPT can explain the rules around ASC 842 or draft a client memo. But they are reactive: they answer when you ask. They do not open your inbox and start working through the invoices.

Agentic AI in accounting does. It receives an objective rather than a script and determines the steps itself. When a vendor invoice arrives, the system reads the document using computer vision, extracts the vendor name, invoice number, amount, due date, and individual line items, then predicts the correct GL account, department, and cost centre for each line based on how the firm has coded similar invoices in the past. It checks for a matching PO. If everything lines up and the system’s predictions are confident, the invoice moves straight to the approval workflow and posts to the ledger. The accountant never sees it. If something does not line up, the system escalates, just as a well-trained junior would. An invoice significantly above the PO amount gets flagged. A new vendor with no history gets routed for review. The system handles the 75 to 80% of transactions that are routine, freeing the human for the 20 to 25% that genuinely require judgment.

The Capabilities Across the Practice

Agentic AI in accounting is being deployed across the full range of professional services work, not just AP.

Accounts payable is the most mature use case. Vic.ai, trained on more than one billion invoices, operates at 97 to 99% accuracy across its client base and continuously learns from user corrections. HSB, a Swedish housing cooperative that manages 6,000 properties and processes 1.5 million invoices a year, reached 96% coding accuracy in its first month on the platform and cut processing time per invoice from over two minutes to 45 seconds, saving an estimated 25,000 staff hours annually. Ramp, a New York-based financial operations platform, has taken a different approach by embedding agents directly into its corporate card and spend management product. Its Accounting Agent, launched in February 2026, auto-codes every transaction the moment it occurs and handles accruals, reversals, and ERP reconciliation in real time. Separately, Ramp’s AP agents, launched in late 2025, flagged over $1m in fraudulent invoices in their first 90 days by cross-referencing payment history, vendor details, and invoice patterns. Lauren Feeney, controller at Perplexity, described the impact of Ramp’s system:

“We’re not choosing between moving quickly or getting it right. The AI handles routine coding in real time and keeps our books audit-ready, so we close faster every month without the scramble.”

The month-end close is where the cumulative effect is most dramatic. Traditionally, the close takes five to fifteen working days: gathering data, reconciling bank statements against the GL, posting accrual entries, chasing missing invoices, preparing the trial balance, running variance analysis, and routing everything for approval. Agentic AI compresses this by handling the preparatory work continuously rather than in a batch at month-end. KPMG launched its Ignite Financial Close Companion in April 2026, built with Google and Workday, which follows a firm’s own close checklist step by step: organising data, interpreting results, flagging anomalies in the trial balance, and drafting entries for review, while posting nothing without human approval. A joint study by MIT Sloan and Stanford found that accountants using AI cut 7.5 days off the monthly close and produced reports with 12% more detail. Accountants using AI also supported 55% more clients per week.

In audit, tools like DataSnipper and Trullion extract information from contracts, link source documents to workpapers, and surface anomalies earlier in the cycle. Trullion’s agentic assistant Trulli, launched in mid-2025, handles document analysis and the interpretation of complex standards such as ASC 842 and IFRS 16. Rather than an auditor spending hours cross-referencing a lease agreement against the disclosure schedule, the agent reads both documents, identifies discrepancies, and presents them for review.

In tax, agentic AI in accounting is moving beyond form preparation into advisory territory. Traditional automation populates returns with data from last year’s filing or a connected payroll system. An agentic system goes further: it analyses a client’s financial data, reviews current regulations, identifies tax savings opportunities, and drafts a client communication outlining them. Thomson Reuters’ CoCounsel platform, which reached one million users in February 2026 across 107 countries, is developing a fully agentic version for tax and audit that independently researches regulations, retrieves authoritative sources, and delivers structured work product with citations.

Who Benefits Most

A recent MIT/Stanford study conducted this year led by Jung Ho Choi of Stanford GSB and Chloe Xie of MIT Sloan, analysed hundreds of thousands of transaction entries across 79 small and mid-sized firms and surveyed 277 accountants. The headline productivity gains were striking: Accountants using AI supported 55% more clients per week, cut 7.5 days off the monthly close, and logged more billable hours because the technology converted previously non-productive time into client-facing work. But the gains were not evenly distributed. Experienced accountants extracted significantly more value from the tools than junior ones. The reason comes down to something very practical: how people respond to confidence scores.

Every agentic system attaches a confidence indicator to its outputs. When it codes an invoice to a GL account, it signals how certain it is. A senior accountant who has coded thousands of similar invoices by hand over a career has an internal benchmark. They know what a transaction from a particular vendor type should look like. When the system produces a high confidence score on a coding prediction that matches their instinct, they let it through. When the score is high but the coding feels wrong, perhaps a marketing expense categorised as cost of goods sold, they intervene. They treat the system as a junior colleague whose work they review with calibrated trust.

Junior staff, the researchers found, lack that internal benchmark. Some accept AI outputs at face value, even when the system’s own confidence scores flag uncertainty, allowing errors to flow through to the ledger unchecked. Others go the opposite direction, second-guessing predictions that are correct, manually overriding accurate coding and losing the efficiency gains. “AI helps with multitasking,” Choi noted. “Accountants have to pull information, connect bank transactions, track vendors. AI assists with that setup, which means they can serve more clients, more efficiently.” But the study also provided early evidence of a risk the profession has not fully grappled with: AI-generated errors flowing through human-in-the-loop accounting systems when the human in the loop does not have the experience to catch them.

The implication cuts both ways. For firms with experienced staff, agentic AI in accounting is a genuine force multiplier. For firms relying heavily on junior teams, the risk of over-trust is real and measurable. And for the profession as a whole, it raises the most uncomfortable question in the entire debate: if the technology rewards the judgment that comes from years of hands-on experience, what happens when the hands-on experience itself disappears?

The Limitations

Accuracy is uneven. For high-volume, structured work like invoice coding, error rates are low. But for complex, judgment-heavy tasks, the technology is far less reliable. Research published by METR in January 2026 found that current AI models can complete a task that would take a human roughly five hours with only 50% reliability. A survey of 128 CFOs and controllers by Accounting Seed found that 35% cite fear of errors and hallucinations as their primary concern. Agentic AI in accounting is excellent at processing. It is not yet reliable at reasoning through ambiguous situations.

Data quality is a prerequisite, not a byproduct. Agents built on fragmented systems with inconsistent GL coding amplify existing problems rather than solving them. The Wolters Kluwer Future Ready Accountant report found that high-growth firms are 53% more likely to have highly integrated technology and 38% more likely to be fully cloud-based.

Regulation has not caught up. No framework governs how AI-generated audit evidence should be treated, who bears liability when an agent makes a material error, or what firms must disclose to clients about the role autonomous systems play in their engagements.

The talent pipeline is at risk. The Big Four have collectively invested at least $9bn in AI while cutting graduate hiring. UK accountancy graduate listings fell 44% year on year when compared with the 24/25 financial year. KPMG slashed its intake from 1,399 to 942. Deloitte reduced by 18%. EY by 11%. A Stanford study found hiring for entry-level, AI-affected roles fell 16% in two years. The work those graduates did, coding transactions, preparing workpapers, reconciling accounts, chasing missing documents, is the work agents now handle. The staffing pyramid that defined the profession for decades is being compressed: fewer juniors at the base, more technology in the middle, roughly the same number of partners at the top. Industry consultants project 15 to 20% fewer entry-level positions by 2027 at firms that fully adopt AI. If the formative years at the base of that pyramid are automated, the profession risks producing a generation of accountants who can oversee AI but have never done the underlying work themselves. No one has demonstrated a credible alternative path to developing that same depth of professional instinct.

The Path to Adoption

For firm leaders considering agentic AI in accounting, the technology is ready. The question is whether the firm is. Based on the patterns emerging from early adopters, five foundations determine whether deployment succeeds or fails.

Data integration comes first. Agentic systems need connected, consistent data to function. Firms running separate systems for AP, billing, payroll, and the GL with inconsistent charts of accounts will not get meaningful results from autonomous software. The first investment is not in AI. It is in connecting the systems you already have and standardising how data flows between them.

Start with high-volume, rules-driven processes. AP is the most common entry point for a reason: invoices are structured, coding patterns are repetitive, and the cost of manual processing is easy to quantify. The month-end close is a natural second step. Audit and tax work, where judgment calls are more frequent and the consequences of errors are higher, should follow only after the firm has developed confidence in how the technology performs and how its people interact with it.

Define the human-in-the-loop model explicitly. The most effective deployments build clear escalation rules: what the system can do autonomously, what requires review, and what must always be handled by a qualified professional. Firms that skip this step end up with staff who do not trust the system and override it constantly, negating the efficiency gains, or staff who trust it too much and miss the errors that slip through.

Invest in governance before you need it. Audit trails, explainability, data security, and clear accountability for AI-assisted outputs are not optional. Every action the system takes should be logged: what data it ingested, what predictions it made, what confidence scores it assigned, and what a human reviewed or overrode. Firms that treat governance as an afterthought will find themselves exposed when an error surfaces or a regulator asks how a number was produced.

Train your experienced people first. The MIT/Stanford research is clear: senior professionals extract the most value from the technology. Investing in their comfort and fluency with agentic tools generates faster returns than onboarding the entire firm at once.

Where the Profession Stands

AI adoption in accounting firms leapt from 9% in 2024 to 41% in 2025. The global AI accounting market is estimated to hit $10.87bn in 2026, growing at roughly 44.6% annually. Gartner projects that by 2028, a third of enterprise software will include agentic capabilities. The adoption curve for agentic AI in accounting shows no sign of flattening. Jason Marx, chief executive of Wolters Kluwer Tax and Accounting, described the competitive reality:

“Firms leveraging AI are working faster, with greater precision and confidence, transforming how strategic decisions are made and how client value is delivered.”

Thomson Reuters’ Future of Professionals Report found that firms with a clear AI strategy are three to four times more likely to see revenue growth. Those without one risk falling irreparably behind within three years. KPMG has already leveraged its own AI capabilities to negotiate a 14% fee reduction from Grant Thornton, its auditor. Fee pressure of this kind will become the norm, not the exception.

But the profession’s most consequential challenge is not competitive. It is generational. The MIT/Stanford research showed that agentic AI in accounting amplifies the value of professional expertise. The more judgment an accountant brings to the table, the more they get from the technology. That is the case for optimism. The case for concern is that the profession is simultaneously dismantling the training ground that produced that expertise. The firms that solve this tension, finding new ways to develop professional judgment in a world where the grunt work is automated, will not just survive the transition. They will define what the accounting profession becomes next.