|
EXPLAINER |
AI FOR ACCOUNTANTS · PART 12 OF 16
Agents versus automation: the close-time difference that matters
Automation follows a script; an agent reads the situation - and that gap shows up fast when a reconciliation gets messy.
The word "agent" has started appearing everywhere in accounting software marketing, usually alongside "automation," as if the two are interchangeable. They are not. Automation is a rule that fires when a condition is met - post this journal entry when the bank feed matches, flag this invoice when the amount exceeds a threshold. An agent is something different: it reads a situation, decides what step to take next, and can change course based on what it finds. That distinction is not semantic. It changes how you supervise the work and what happens when something goes wrong.
Think about a standard bank reconciliation. A rule-based automation can match cleared items, calculate the outstanding balance, and drop a variance line into a workpaper. It does that the same way every month. If the variance is zero, fine. If the variance exists but is explainable by a known timing item, the rule still just flags it - a human reads it and moves on. An agent, given the same task, might notice the variance, pull the prior month's workpaper to check whether that timing item recurred, look at the GL detail for similar amounts, and draft a note explaining what it found. It made three decisions you did not script. That is useful. It is also where things can go sideways if the agent's scope is not defined.
The failure mode with agents is not that they break - it is that they do something plausible that is quietly wrong. An agent working a revenue reconciliation might decide that a contract modification in the client file explains a discrepancy and close the item without flagging it for review. The logic is reasonable. The conclusion might be wrong. With automation, the failure is visible: the rule misfires or does not fire. With an agent, the failure can look like a completed workpaper. This is why 'scope' is the operative word, not 'script.' You are not telling the agent every step to take; you are telling it which files it can touch, which conclusions it can make on its own, and which ones require a human sign-off before anything moves.
The practical implication at month-end close is that automation and agents should live in different lanes. Automation handles the deterministic work: matching, routing, flagging by threshold, populating templates from a structured data source. Agents handle the interpretive work: reading a memo, summarizing a contract, drafting an explanation of a variance, or triaging a queue of open items. The boundary is where judgment enters. If the next step requires reading context and choosing among possible responses, that is agent territory. If the next step is always the same given the same input, that is automation territory. Mixing them up - expecting an agent to be perfectly consistent, or expecting automation to handle ambiguity - is where most close-process breakdowns happen.
The control mechanism for agents is scope definition, not prompt engineering alone. A prompt tells the agent what to do; scope tells it what it is allowed to touch. In practice this means specifying which workpapers an agent can write to versus read from, whether it can mark an item resolved or only flag it for review, and whether it has access to the prior-period file or only the current one. Firms that have gotten this right tend to treat agent scope the way they treat user permissions in their GL system - role-based, documented, and reviewed when the workflow changes. That framing - permissions, not instructions - is the most useful mental model for keeping agent-assisted close work auditable.
WORKED EXAMPLE
In practice
A staff accountant is closing the revenue reconciliation workpaper for the month just ended. There is a variance between the subledger and the GL. The client file includes a contract modification memo from earlier this month.
What came back. The agent identified a variable consideration clause in the memo that could plausibly defer recognition of the variance amount to the following period, listed three pieces of supporting documentation it would need to confirm the treatment, and drafted two sentences of potential footnote language. One of its reasoning steps assumed the modification effective date matched the period close date, which the accountant confirmed was incorrect by one week - the effective date was after period end, changing the analysis.
How it was checked. The accountant compared the agent's stated effective date assumption against the signed modification memo and found the discrepancy before the item was marked reviewed.
A constructed example. The prompt is usable as written; the figures show the shape of a result, not a measured one.
WHEN TO USE IT
| WHEN NOT TO
|
WHAT TO TAKE FROM THIS
| Define agent scope as permissions - which files, which conclusions, which sign-offs - not just as a prompt. | |
| Keep deterministic steps in automation and interpretive steps in agents; blurring that line creates silent errors. | |
| Any item an agent marks resolved should have a documented rationale the reviewer can inspect before sign-off. |
SPONSORED
QUESTIONS THIS ANSWERS
What is the difference between an AI agent and workflow automation in accounting?
Automation executes a fixed rule when a condition is met. An agent reads context, decides among possible next steps, and can change course mid-task. The agent introduces judgment; automation does not.
How do I control what an AI agent does during a close?
Define scope as permissions: which workpapers it can write to, which conclusions it can finalize without review, and which actions require human sign-off. Treat it like GL user permissions, not like a to-do list.
Where do AI agents fail in accounting workflows?
They fail silently - producing a plausible but wrong conclusion that looks like a completed workpaper. The risk is highest when the agent can mark items resolved without a required human review step.
SOURCES
Where this comes from
What the accounting job market is actually asking for.
GO DEEPER
Go deeper
IN THIS SERIES
Previously: Why the same prompt gives a different answer twice
Next: What an AI audit trail should contain (coming)
An explainer, not a study: it carries no statistics on purpose. Examples are illustrative.
How accountants are using AI, automation and smarter workflows to close faster, audit cleaner, and free up time for real work.
Accounting Stack · Audit Friendly Data · Accounting & Finance Jobs
Audit Friendly · modernaccounting.ai