Artificial intelligence has become increasingly familiar within corporate finance. Controllers and accounting teams are using tools such as ChatGPT, Claude, and Gemini to research accounting questions, summarize information, analyze files, and prepare initial drafts. The larger question is whether these same tools can be trusted to perform recurring accounting work that affects the general ledger.
That distinction between using AI for knowledge and using AI for execution was central to a recent Controllers Council and Maxima webinar featuring Yogi Goel, Co-Founder, CEO, and CFO of Maxima. Goel has spent more than two decades working across accounting, banking, strategic finance, and enterprise leadership, including experience at EY and Rubrik.
During the AI for Corporate Accounting: A Controllers Guide webinar, Goel examined the limitations of general-purpose AI for accounting, the controls required when AI touches financial processes, and the role purpose-built accounting agents may play in the financial close.
Most Accounting Teams Are Still Early in Their AI Adoption
An opening poll provided a useful picture of how accounting organizations are currently approaching AI. Thirty-nine percent of attendees reported using AI for preliminary evaluation, ad hoc research, and analysis, while 35 percent had implemented AI in one or more workflows. Only 12 percent said AI was integrated across accounting and finance operations.
Those results reflect a finance profession that has become familiar with AI while still working through how it should be applied to controlled accounting processes.
Goel encouraged accounting professionals to think about AI as a progression. At the simplest level, a chatbot can answer an accounting question or assist with a one-time technical accounting memo. AI can then advance toward spreadsheet analysis, recurring preparation work, complex high-volume processes, and eventually broader execution across accounting workflows.
The difficulty increases considerably as AI moves further along that progression.
Accounting Knowledge is Different from Accounting Execution
Generic AI can be useful when the assignment is isolated, informational, or relatively low risk. An accountant might ask how a particular expense should generally be treated, request help analyzing a spreadsheet, or research an unfamiliar accounting issue.
Executing a monthly accrual is different. The system needs company-specific data, prior-period information, accounting policies, supporting evidence, accurate calculations, approvals, and a record of what occurred.
As Goel explained, “knowledge and execution incredibly different.”
He cited research that tested advanced general-purpose AI models against 160 real-world accounting tasks developed by experienced accounting professionals. The tasks included reconciliations, data entry, variance analysis, schedules, and accruals. According to the presentation, the models achieved approximately 56 to 59 percent accuracy when attempting the work independently.
For accounting departments, consistency matters because one incorrect entry can carry forward into subsequent periods. An error in a prepaid balance in January, for example, may affect February and later periods unless it is identified and corrected.
That makes a system’s ability to explain accounting concepts considerably different from its ability to prepare work that a controller can confidently review and an auditor can reperform.
What Accounting AI Needs Beyond an LLM
For AI to participate meaningfully in recurring accounting processes, Goel identified several capabilities that must surround the underlying language model.
First is reliable access to company data. Accounting work rarely depends upon a single system. Relevant information may reside in the ERP, purchasing platform, banking system, CRM, payroll application, HR platform, credit card system, or other sources. Without complete transactional and master data, an AI system may lack the context required to prepare accurate work.
The second requirement is deterministic calculation and validation. Controllers need to understand how a number was calculated and verify that the same process produces the same result when it is repeated.
Continuity also matters. Accounting depends heavily upon historical information. Prior-period balances, previous classifications, amortization schedules, and established treatments influence current-period work. A temporary chatbot session may not retain that context in the manner required for recurring accounting processes.
Finally, accounting AI requires appropriate controls. Access restrictions, segregation of duties, approval records, and audit trails remain necessary regardless of whether a person or an AI agent prepares the underlying work.
The Auditor Still Needs Evidence
Auditability was another substantial part of the discussion.
An accounting AI system cannot simply arrive at the correct answer. Finance teams need to demonstrate where the source data originated, how calculations were performed, who prepared the work, who reviewed it, and how the resulting entry relates to the supporting information.
Goel summarized the accounting mindset with a familiar saying from his auditing experience: “we trust in God, but the rest, everything we must audit.”
For AI-supported work, that means calculations should be reproducible, source transactions should remain traceable, journal entries should connect to their supporting evidence, and review activity should be documented.
Human review remains particularly important. As Goel noted, “the evidence of a human review is incredibly important.”
Without sufficient evidence, accounting teams may end up performing the work again manually. Any anticipated efficiency can disappear quickly if employees or auditors must recreate calculations because they cannot verify how the original result was produced.
A Practical Checklist for Evaluating Accounting AI
Controllers evaluating AI for production accounting should therefore look beyond the model itself. The webinar identified several questions worth considering before allowing an AI system to participate in recurring financial processes:
- Can the system maintain reliable integrations with the ERP and other source systems?
- Does it work with current company data, including changes to accounts, cost centers, subsidiaries, and other dimensions?
- Are calculations performed through deterministic methods that accounting professionals can understand and reproduce?
- Can the system apply company-specific policies, thresholds, materiality levels, and approval requirements?
- Are human review gates incorporated throughout the workflow?
- Does the system preserve detailed, retrievable audit trails for prior periods?
These considerations become increasingly important as AI moves from research and analysis into processes that ultimately affect the books.
Seeing an Accounting Agent in Practice
The webinar also included a demonstration of Max, Maxima’s accounting agent.
Goel described Max as an agent designed to perform preparation work while accountants retain responsibility for review and approval. The system can pull source information, validate data, build workpapers, draft outputs, run checks, and route completed work for review with supporting evidence and an audit trail.
During the demonstration, Max prepared an accrual workflow. The agent retrieved purchase order information, invoices posted to the general ledger, and invoices awaiting approval. It then prepared accrual calculations, created supporting schedules, performed checks, and assembled journal entry backup.
The calculations remained formula-driven so that reviewers could inspect how amounts were determined. After the work was prepared, a human reviewer could examine the workbook, supporting documentation, and proposed journal entry before approving it for posting to the general ledger.
The example illustrated an important principle for controllers considering AI: automation does not remove the review process. Instead, it can change where accounting professionals spend their time within that process.
Where Controllers May Want to Start?
When attendees were asked which close workflow, they would first assign to an agent, 34 percent selected bank and cash reconciliations, while 21 percent chose prepaid and amortization schedules. Another 12 percent selected flux commentary, while 28 percent were not yet ready to hand a close workflow to an agent.
The preference for reconciliations and schedules is understandable. These activities are recurring, structured, and generally involve less judgment than many technical accounting decisions.
That provides a practical starting point for organizations evaluating accounting agents. Rather than attempting to automate an entire close immediately, controllers can identify repetitive preparation activities with established procedures, measurable outputs, accessible source data, and clear review requirements.
The Controller’s Role Will Continue to Matter
One audience member asked what skills controllers and accounting teams need as agents will assume more preparation work.
Goel pointed to accounting judgment, the ability to triangulate information, detailed knowledge of accounting treatments, and influence across the organization.
Those responsibilities become more consequential when technology handles a greater portion of mechanical preparation. Controllers still need to recognize unusual activity, assess whether accounting treatment is appropriate, identify information that may be missing, and understand the financial implications of decisions made elsewhere in the business.
They may also have more time to work with sales, operations, finance, and other functions before business decisions create difficult accounting consequences.
Autonomous Accounting Still Requires Human Oversight
The webinar concluded with an important qualification concerning autonomous accounting.
Goel cautioned against interpreting the term to mean a completely hands-free finance function. “autonomous accounting today is a promise and not a reality,” he said.
His preferred model today is closer to a co-preparer relationship. AI can perform a substantial portion of repetitive preparation, while accountants continue to review the work, exercise judgment, resolve exceptions, and maintain responsibility for the financial results.
For controllers, this distinction may provide a useful framework for evaluating the next generation of accounting technology. Generic AI remains valuable for research, drafting, analysis, and other individual assignments. Production accounting requires considerably more structure, including reliable data, company-specific context, deterministic calculations, access controls, human approvals, and durable audit evidence.
The question for accounting leaders is therefore becoming more specific. Rather than asking whether AI belongs in accounting, controllers can begin determining which work can be delegated, what evidence must accompany that work, and where professional judgment must remain firmly with the accounting team.
To see the complete discussion, product demonstration, audience polling, and Q&A, watch the full webinar here.
About the Sponsor
Maxima AI is the AI-native agentic accounting system of work built for enterprise accounting teams. Their agent pulls data from your ERP, applies your policies, and prepares journal entries, reconciliations, flux analysis, and financial close with audit-ready accuracy so accountants review and approve. The result: faster closes, fewer errors, and full control. Learn more at www.maxima.ai
