Autonomous Agents Explained: Benefits, Risks, and Real-World Applications
EY plans to scale from 150 AI agents serving 80,000 tax professionals to 100,000 agents by 2028 as firms like PwC expand agentic AI platforms.
For an accounting professional, “autonomous agents” is the technology term most worth understanding in 2026. It appears in software release notes, in firm strategy slides, and in the trade press almost weekly. Beneath the hype is a specific category of software that is already changing what accountants do, how firms price their work, and which tasks remain reserved for human judgment. This guide is written for working accountants who want a clear, practical understanding of what the technology is, how it differs from the AI tools you have already used, where it is being deployed in the profession today, what it does well, what it does badly, and what to do about it.
What autonomous agents actually are
An autonomous agent is software that pursues a goal on its own across multiple steps. To break that down, an agent does four things, repeatedly, in a loop: it perceives information from its environment, reasons about what to do next, acts through other software, and adapts based on the result.
The reasoning is usually handled by a large language model, the same family of system behind ChatGPT and Claude. The acting is done through application programming interfaces (APIs), which are the standard way software systems read from and write to each other. In an accounting context, those APIs connect the agent to your general ledger, your bank feeds, your tax preparation platform, and any other system holding the data the agent needs.
The single most important idea to grasp is that “AI” in the accounting profession now refers to three quite different things, and mixing them up leads to bad decisions about what to buy, what to trust, and what to control.
| Type | What you give it | What it does |
|---|---|---|
| Traditional automation | A fixed rule | Follows a script; breaks if anything unexpected happens |
| AI copilot | A prompt | Does that one thing and stops; reactive by design |
| Autonomous agent | A goal | Works through all the steps needed to achieve it, calling on tools and data as it goes |
That progression is the conceptual ladder worth memorising. Each rung adds capability and removes a layer of human direction. A bank-feed rule that codes Starbucks transactions to meals-and-entertainment is automation. Microsoft Copilot answering a single question is a copilot. An agent told to “close the books for May” will pull the bank data, post the journal entries, run the reconciliations, flag the exceptions, and hand you a clean trial balance to review, without being prompted at each step. The umbrella label “agentic AI” simply means software built around these autonomous agents.
Why this matters for your profession right now
The reason the largest accounting firms have moved so aggressively into autonomous AI is not simply enthusiasm for technology. Compliance work remains the foundation of accounting revenue, and autonomous agents significantly reduce the labour cost of producing it. EY announced in 2025 that 80,000 tax professionals would gain access to 150 AI agents through its EY.ai platform, with plans to scale to 100,000 agents by 2028 following more than $1 billion in annual AI investment. PwC has also launched its own “agent OS” platform to integrate AI agents across enterprise and client workflows as major firms push to automate labour-intensive compliance tasks.
The implication worth understanding clearly is this: when the marginal cost of a reconciliation or a routine return falls toward zero, the price clients will pay for that work falls with it. EY’s Raj Sharma has described a shift from hourly billing toward a “service-as-software” model, where clients pay for outcomes rather than time. Junior roles are being reshaped accordingly. At KPMG, juniors are being trained as “managers of agents” rather than as doers of first-pass work.
In other words: the work that used to fill your week is becoming a commodity. The work that used to fill a senior accountant’s week (judgment, interpretation, advice) is what your firm now sells.
Where autonomous agents are being used in accounting today
The clearest way to see what these systems do is to walk through a month-end close at a mid-market firm that has deployed them.
On day one of the close, an agent pulls bank statements through a feed, matches transactions against the general ledger, and posts the straightforward reconciling entries. Where it cannot find a confident match, it stops and hands the exception to an accountant with a one-line proposed treatment. The human decides. The agent records the decision and uses it to handle similar cases later.
On day two, an accounts-payable agent ingests vendor invoices, performs three-way matching against purchase orders and goods-received notes (checking that what was ordered, what was received, and what is being invoiced all agree), codes each invoice to the chart of accounts, and routes the approval. Anything ambiguous, an unusual amount, a new vendor, a coding question, escalates to a human reviewer.
On day three, the agent prepares recurring journals, accruals, and intercompany eliminations, surfacing only items that fall outside expected patterns. On day four, a separate agent runs variance analysis against budget and prior year and drafts a commentary that the controller edits rather than writes from scratch. By day five, the human work (sign-off, judgment calls on cut-off and provisioning, disclosure decisions) looks much as it always did. What has changed is the forty hours of preparation that used to come before it.
Vendors building this stack are no longer theoretical. Pilot launched a fully autonomous AI bookkeeper in February 2026, capable of running accrual-basis books end to end. FloQast markets “autonomous accounting” software built around the close. Karbon is rolling out agents for data entry, onboarding, and workflow coordination. In tax, Thomson Reuters has built agentic tools for year-round research and compliance. Goldman Sachs, CNBC reported, is co-developing agents with Anthropic to handle accounting for trades and client onboarding.
One caveat is worth carrying with you. Ellen Choi, an accounting-focused AI consultant, told Accounting Today in February 2026 that current systems handle simpler individual tax returns and single-entity bookkeeping well, and struggle with the multi-entity, judgment-heavy engagements that define senior practice. Treat that as the operating envelope today, not the long-term ceiling.
What can go wrong
There are four risks of autonomous agents to understand, in the order they will reach you.
Data quality is the risk closest to your desk. An agent is only as good as the systems it can read and write to. A chart of accounts with inconsistent naming, duplicate vendor records, and journal entries that exist only in someone’s spreadsheet does not become useable when you add an agent. It becomes a faster source of bad output. Cleaning the data and integrating the underlying systems is the prerequisite, not the optional extra.
Hallucination is the risk in the headlines when it comes to autonomous agents. Large language models work by predicting the most plausible next word, sentence by sentence. They have no internal sense of whether something is true, only whether it sounds right given everything they have seen. The result is that they can produce confident, fluent, completely fabricated information with exactly the same tone as accurate information. In October 2025 Deloitte Australia agreed to partially refund a roughly AU$440,000 (US$290,000) government contract after a report it had delivered was found to contain invented academic citations and a fabricated quotation from a Federal Court judgment, errors traced to undisclosed use of generative AI. The lesson is not to avoid the technology. It is that the review process you apply to an agent’s output must be designed to catch confident falsehoods, which is a different control than the one designed to catch honest mistakes. Spot-checking sources, not just conclusions, is the practical implication.
The audit trail is the risk regulators care about. Control frameworks such as SOC 2 (the standard that governs how service organisations protect client data and financial systems) expect privileged actions to be attributable to an accountable person. “The agent did it” does not satisfy that. Any agent in production should keep a reasoning trace: a step-by-step record of what it saw, what it considered, and why it acted, for every autonomous decision.
Regulation is the risk on the calendar. Under the EU AI Act, the bulk of obligations take effect on 2 August 2026, with full enforcement powers, although a recent amendment may push some high-risk finance and employment rules to December 2027. Either way, continuous monitoring, technical documentation, and demonstrated human oversight are becoming legal expectations rather than best practice.
Questions worth asking
If autonomous agents come up in a meeting at your firm, these are the questions that matter.
If your firm is considering deploying them:
- What data is the agent reading and writing to, and who owns that data?
- What happens when it gets something wrong, and who is accountable?
- Can we see a log of every decision it made and why?
- What must it escalate to a human, and who is that human?
If you are reviewing output that an agent has produced:
- Have the sources behind any cited figures actually been checked, or just the figures themselves?
- Does anything here require a judgment call that the agent could not have made reliably?
The second list matters because the most important skill when working alongside these systems is not knowing how to use them. It is knowing what to be sceptical about when you are handed their output to sign off on.
Accountio.
The go-to weekly newsletter for accounting professionals. Trusted by 10,000+ industry leaders to deliver the tech trends and insights that matter most. Join them today.
