AI is increasingly capable of doing more than assisting accounting teams. It can identify discrepancies, suggest classifications, prepare actions, and in some cases execute them directly within accounting systems.
This creates a governance problem that simple approval workflows cannot solve: when software performs part of the accounting process, who is responsible for the decision?
The answer cannot be the system itself. AI processes data and operates within defined boundaries, but it does not carry responsibility for financial outcomes. That remains with the business and the people who define what the system is allowed to do, when it must be reviewed, and how its outputs are validated.
This division of responsibility was central to a recent eWeek Trend Talk sponsored by QuickBooks Online. Corey Noles, AI expert and managing editor of The Neuron, and Accounting and QuickBooks expert Max Crampton-Thomas examined how businesses can introduce automation without losing the judgment, context, and accountability that sound financial management requires.
As AI takes on more operational authority, finance teams need a middle ground between manual control and full automation. That middle ground is graduated autonomy: assigning different levels of authority to AI based on the risk of the task and the system’s demonstrated reliability.
Automation authority should be earned
Not all accounting tasks carry the same level of risk. Categorizing a low-value, routine transaction is very different from approving a large payment or making an adjustment that materially affects financial statements. A single automation rule cannot safely cover both.
AI can therefore operate at different levels. It may simply flag an issue, recommend an action for human review, prepare an entry for approval, or execute an action within strict limits. Which level is appropriate depends on the potential impact of an error, whether the action can be reversed, and how consistently the system performs in that specific workflow.
Crampton-Thomas captured this risk-based approach clearly in the Trend Talk. “If I put it simply, the harder a mistake would be to undo, the less authority I would give the system on its own,” he said.
This creates a progression of trust. A system might begin by suggesting classifications while people review every recommendation. If it demonstrates consistent performance over time, the organization may allow it to handle more of that narrow workflow while continuing to escalate uncertain or higher-risk items.
Importantly, trust is not a general judgment of whether an AI tool is accurate. It is specific to a task, dataset, and operating context. Strong performance in one area does not justify broader authority elsewhere.
The goal is not maximum automation. It is appropriate automation aligned with risk.
Approval is not the same as oversight
The idea of a “human in the loop” is often treated as sufficient control. But an approval step alone does not guarantee meaningful oversight.
For review to be effective, the reviewer must understand what the system is recommending, what information informed that recommendation, and what the consequences of an error would be. The reviewer must also have the time and authority to challenge the output.
Without these conditions, approval can become routine. When AI recommendations appear reliable and are embedded in familiar accounting workflows, employees may begin accepting them automatically. The checkpoint remains, but meaningful scrutiny disappears.
As Crampton-Thomas explained, “The reviewer needs to understand why the item reached them and see the evidence behind the recommendation. If they can't tell what their approval will actually change, that approval step creates a false sense of security.”
This is especially important in accounting, where context matters as much as pattern recognition. A transaction that looks unusual may be an error, or it may represent a legitimate business decision the system cannot fully interpret. A vendor may appear under different names, while a change in revenue or expenses may reflect circumstances that are not visible in the underlying data.
Effective oversight therefore requires more than inserting a person after the AI. Finance teams must define who is accountable for review, what evidence is provided, and which situations require escalation.
Review effort should also match risk. Low-value, reversible transactions do not require the same attention as high-impact or unusual financial events. Human judgment should be concentrated where it can materially affect the outcome.
Exception management becomes a core finance skill
AI does not simply accelerate accounting work. It changes what people spend their time reviewing.
In manual processes, teams may inspect thousands of routine transactions to find a small number of errors or anomalies. AI can process more of that predictable activity and surface the items most likely to require attention.
This does not eliminate judgment. It concentrates on it.
It also introduces a new responsibility: determining what qualifies as an exception. If thresholds are too broad, teams may become overwhelmed with alerts. If they are too narrow, important issues may pass without review.
Finance teams must decide which patterns, values, and uncertainties justify escalation. Transaction size matters, but so do reversibility, frequency, supporting evidence, and potential financial impact.
In this model, AI reshapes accounting work into a smaller, higher-value set of decisions where human expertise matters most. Exception management becomes part of financial control, not merely operational cleanup. Teams must ensure the system is surfacing the right issues, not simply fewer of them.
Accountability cannot be delegated to software
As AI moves from recommending actions to executing them, accountability can become less visible.
Even if a person does not initiate each individual action, someone must remain responsible for the framework that allows those actions to occur. That includes defining thresholds, approving the system’s authority, monitoring its performance, and intervening when its behavior changes.
Traceability is therefore essential. For any material action, the organization should be able to reconstruct what data the AI used, why it acted, what it did, and whether human review occurred.
Without that record, automation may increase speed while reducing control. Transactions may appear correct even as the organization loses the ability to explain how they were produced.
Responsibility does not require people to execute every step manually, but it also cannot disappear into the system.
Finance needs clear decision rights for AI
The core challenge is not whether AI should be used in accounting. It is how much authority the technology should receive and who remains accountable for its actions.
This requires explicit decision rights. Finance teams must define which tasks AI can perform independently, which require approval, which must remain under human control, and what evidence is needed before the system’s authority expands.
Graduated autonomy provides a practical structure. It allows organizations to automate repetitive, lower-risk work while preserving human judgment where context and financial consequences matter most.
That distinction also keeps automation connected to a business outcome. Crampton-Thomas summarized the objective near the end of the Trend Talk, “The goal is not to remove people from accounting. Really, it's to free them from repetitive work so they can focus on decisions where their experience actually matters.”
AI can identify, recommend, prepare, and execute actions. It cannot own the outcome.
As AI becomes more embedded in accounting, the strongest control will not be universal human approval. It will be a clear governance model in which automation is earned, oversight is meaningful, exceptions are well managed, and accountability always has a human owner.