Accounting AI leaders say autonomous agents aren’t there yet
During the July 8 Earmark webinar “AI Agents for Accountants: What’s Working Right Now”, host Blake Oliver, who also serves as CEO of Earmark and co-host of The Accounting Podcast, asked what — in theory — should have been the easiest question of the session.
“What are the AI agent workflows that firms are successfully deploying now?” Oliver asked. After a pause, a panelist deflected the question to panelist Toni Witt, co-founder and CEO of Sweet, a company that helps accounting firms develop artificial intelligence strategies and custom AI agents.
Witt’s points and the subsequent discussion moved through examples that CFOs and firms alike are already putting to work, including inbox management, Excel automation, dashboard creation, reusable AI skills and, ultimately, workflow standardization.
Panelists described AI handling spreadsheet transformations, organizing email, generating reports and helping accountants complete recurring tasks. As the conversation progressed, examples of accounting workflows fully owned by AI agents remained questionably limited throughout a session devoted to the topic.
This conversation echoes a broader trend emerging across finance. While the hype around agentic AI is there for those who sell it, many of the CFOs thinking of implementing it are skeptical — introducing the technology through structured, repeatable workflows before expanding into more complex tasks.
Recent research from Bain & Company found only 7% of surveyed organizations have fully autonomous AI agents running in production, while Gartner has argued that data context remains one of the biggest barriers to reliable agent performance. Those themes surfaced repeatedly throughout the panel.
Firms are starting with predictable work
Witt’s answer suggested accounting firms are finding the earliest success where work already follows defined rules. He pointed to spreadsheet-heavy processes that accountants perform every day — things like transforming point-of-sale exports, validating formulas, preparing “workpapers” and converting raw data into standardized templates.
“Within the firm, one of the lowest-hanging opportunities is Excel-heavy workflows,” Witt said. “If you have staff working with formulas, importing CSVs into Excel, exporting data and converting it into templates, those workflows are a really good place to start.” He also described firms using AI to organize inboxes, generate dashboards for clients and create reusable “skills” that standardize recurring work across teams.
That measured approach resembles how larger finance organizations are deploying agentic AI, too. Earlier this year, Marie Myers, CFO of Hewlett Packard Enterprise, the global IT arm of Hewlett Packard, told CFO Dive the company deliberately began with one strategic operational process before expanding AI into accounts receivable, accounts payable and forecasting.
Building the foundation before expanding
Twyla Verhelst, vice president of industry relations and community at Karbon, a practice management platform for accounting firms, said on the panel that many firms are moving beyond simply experimenting with AI tools and are now figuring out how to incorporate them into the way work gets done. That transition, she said, takes time because introducing new technology doesn’t immediately change how an organization operates.
“This will take time to get the time back,” Verhelst said. “It’s an adoption curve.” She hinted that firms need to embed AI into their workflows before the technology can consistently deliver the efficiencies many organizations expect.
Her comments align with recent Gartner research, which argues that AI agents require far richer context than traditional enterprise data typically provides. According to their research, organizations that build semantic layers around their data are expected to improve AI accuracy while reducing costs, suggesting the biggest barriers to agentic AI may lie in enterprise data infrastructure and not in the capabilities of the tool itself.
Colleen Oates, head of indirect go-to-market at Gusto, a payroll and HR software provider that works with accounting firms, said on the panel that many professionals, including herself, are still becoming familiar with the basic building blocks of agentic AI before deciding where it fits within their organizations.
“I have felt a sense of overwhelm when it comes to how to start and how to scale,” Oates said, when asked about her experience in building autonomous agents. “I want to understand the elements that live inside of an agent, I want to know the components of it, and I want to know how to build it.”
Her comments suggested that, even as interest in agentic AI accelerates, many accounting professionals are still working to understand the technology itself before determining how to deploy it across their firms. She hinted that finding resources that teach how these things work, though plentiful, can be challenging.
Recent research from Bain points to a similar conclusion. While companies continue increasing AI budgets, the consulting firm found most organizations have yet to achieve their expected cost savings, with inadequate data integration emerging as the most commonly cited obstacle.
As a result of their research, Bain recommended that companies begin by automating repeatable, high-value workflows before attempting broader autonomous deployments, an approach closely resembling the examples discussed during the panel.
Judgment remains with the CFOs
The panel’s clearest answer came near the end of the discussion during the audience Q&A. An attendee asked which accounting close tasks AI agents currently own from start to finish without human review. “None at all,” Oates replied.
She said accounting data, client behavior and transaction activity remain too inconsistent to remove professionals from the review process.
“I think it’s going to be a while before we feel comfortable saying, ‘Let’s pull a human out of this entirely,'” Oates said. “I think it’s 75% to 80% of that workload that is easier and more consistent and happens more efficiently, and then the 20% to 25% that we still need to check or review.”
Again, her comments mirror what finance executives have been telling CFO.com in recent months. Karen Walker, CFO of cloud security provider Sysdig, said there is nothing she would “full stop delegate without oversight,” while Gina Goetter, CFO and COO of Hasbro, an international toy and entertainment holding company, said one of the first responsibilities she would hand to an AI clone would be administrative approvals, opting to keep the strategic decision-making for herself.
By the close of the session, a pattern had emerged of panelists repeatedly returning to examples involving spreadsheet automation, meeting recordings and other recurring types of work. The discussion illustrated that accounting firms, like CFOs, are finding some practical uses for AI today, but also showed how the role of fully autonomous agents in accounting remains an open question, even among the practitioners building and deploying the technology.
