What AI agents are actually doing in finance
AI agents in 2026 are effective at repetitive, pattern-based work: categorizing transactions based on vendor and amount patterns, drafting vendor payment reminders, generating first-pass variance commentary from raw numbers, flagging anomalies in close data. These are tasks where the answer is usually one of a small set of options and the signals are consistent.
The gap between what AI can do and what AI should do in finance is wider than the marketing suggests. A tool that categorizes 95% of transactions correctly sounds great until you realize the remaining 5% includes the transactions that would actually change your reporting. Small, repetitive, low-value transactions are easy for AI. Larger, context-dependent transactions are exactly where the errors concentrate.
In practice, the teams getting value from AI agents treat them like a junior accountant. The AI does the first pass, a human reviews, corrections flow back into the training data, accuracy improves. The teams struggling are the ones who tried to remove the human review layer entirely to save time, and are now discovering the errors months later during audit prep.
The best AI-assisted finance setups we have seen use AI for three specific workflows: categorization of repetitive transactions, first-draft variance commentary on routine line items, and anomaly flagging. These three together save 10-15 hours a month at most mid-size companies. The setups that struggle try to automate revenue recognition, accruals, and judgment-heavy reconciliations.
Where they still fall short
Judgment calls still need humans. Should this contract be booked as revenue this quarter or next? Is this $12K expense a one-time project or a recurring commitment? Does the margin compression reflect mix shift or price pressure? The data is only half the answer - the other half is context about the business that AI does not have access to.
Revenue recognition is the canonical example of where AI judgment falls short. A contract that has three deliverables with different recognition timing, or a contract where some milestones are met and others are not, requires reading the agreement, understanding the commercial intent, and applying accounting standards. AI can assist, but a controller still has to make the call. Getting this wrong creates audit issues that take months to unwind.
Accruals are the other common trap. Accrue too much and you under-report current-period profit. Accrue too little and you over-report. The decision depends on information that lives in emails, Slack messages, and conversations with department heads. Unless the AI has access to all of that context and can interpret it, the accrual decision is staying human for the foreseeable future.
Classification of unusual or one-time items also needs human judgment. A $50K payment to a consulting firm might be professional services OpEx, a one-time restructuring charge, or a capitalizable project cost depending on context. AI cannot read the intent behind a transaction. A controller who has been in the meetings can.
How to introduce AI into close operations
Start with categorization and reconciliation assistance, not reporting. Let the agent propose categorizations for the controller to review, not post entries directly. Build a feedback loop so errors get corrected and the system learns your chart of accounts. Most companies see meaningful time savings within two months, but only if someone is actively tuning the feedback.
A 90-day pilot structure that works: month 1, let the AI run categorization in parallel to your existing process, compare the output. Month 2, promote AI to the primary with human review on everything. Month 3, narrow the review to items flagged as uncertain or above a materiality threshold. After 90 days you have data on accuracy, time saved, and where the AI struggles specifically in your chart of accounts.
The feedback loop is the part most teams skip. When the AI miscategorizes something, most teams fix it in the current period and move on. Unless that correction gets fed back into the system somehow, the same error happens next month. Make the correction workflow a first-class part of the process - even if that just means a shared document of "AI missed this, here is why" reviewed monthly.
Change management matters more than the technology. A finance team that has been doing things manually for years needs time to trust AI-generated output. Roll out slowly, show early wins, and let the team see the errors get caught by the review layer. Trying to force adoption in month 1 usually produces resistance that slows the rollout rather than speeds it.
The role of the controller changes
When AI handles first-pass categorization and reconciliation, the controller spends less time on transaction-level work and more on review, variance analysis, and judgment calls. This is generally a net improvement - the controller is more valuable doing high-judgment work than matching transactions. It does mean the role skillset is shifting toward interpretation and communication.
Finsightic handles accounting, controller oversight, and fractional CFO work for growing companies. Fixed monthly pricing, no long-term contracts.
The controller's job does not go away with AI. It moves up the value chain. Less time on "was this categorized correctly" and more time on "what does this quarter's margin compression tell us about pricing." That is the version of controller work boards actually want to see, and most companies have not had enough of it because controllers were drowning in transaction-level review.
The secondary effect: controller hiring criteria shift. The technical accounting depth still matters, but the communication and judgment skills matter more. Controllers who can write clear variance commentary, explain quarterly results to non-finance stakeholders, and partner with the CFO on strategic decisions are more valuable in 2026 than controllers who are excellent at ticking reconciliations.
For smaller companies, AI is changing the economics of hiring. You can now get controller-grade output from a combination of AI tools and a part-time senior accountant for less than the cost of one full-time mid-level hire. This makes proper financial oversight accessible to companies that previously could not afford it.
Take the free Financial Health Score →