For the last three years, when someone said their finance stack had AI, they usually meant a chatbot sitting on top of a dashboard. Ask it a question about the numbers, and it would answer. Ask it to summarize a report, and it would summarize. Ask it to explain a variance, and it would explain. Genuinely useful. Also genuinely bounded. The AI could tell you what happened. It could not do anything about it.

That is changing, and it is changing faster than most finance leaders I talk to have registered.

The First Wave Was Conversational. The Second Wave Is Operational.

I have watched this shift happen across the platforms I work with, and the pattern is consistent enough that I think it deserves a name. The first wave of finance AI was built to answer questions. The second wave is built to execute work.

The distinction matters more than it sounds like it should. A conversational AI layer sits next to the finance function and makes it faster to understand what already happened. An operational AI layer sits inside the finance function and does part of what used to require a person. Those are different products solving different problems, and most vendors have not been honest with buyers about which one they are actually selling.

I see the operational version showing up in specific, concrete places now. Contract ingestion systems that read a signed order form and structure the billing terms without a person transcribing it. Collections workflows that monitor invoice status, send reminders, and apply incoming payments against the right invoice without a person matching them. Reconciliation processes that flag genuine exceptions and clear the routine matches automatically, rather than surfacing every line item for a human to review. Revenue recognition engines that identify performance obligations directly from contract structure and keep the revenue waterfall current without someone rebuilding a spreadsheet every month.

None of that is a chatbot. All of it is finance work that used to require a person doing it by hand.

The Proof Point That Changed How I Think About This

I wrote recently about Rillet crossing a billion-dollar valuation, and buried in the coverage of that raise was a detail worth pulling out on its own. One of Rillet’s customers reportedly runs a two-billion-dollar revenue business with a three-person finance team, using AI agents to do the work that would traditionally require dozens of people.

I do not think that number is representative of most businesses today. I think it is representative of where the ceiling actually sits once you stop treating AI as an assistant and start treating it as an agent that does the work directly. That is the real shift. Not that AI got smarter at summarizing. That AI became trustworthy enough, in specific and bounded domains, to actually transact.

What “Doing the Work” Actually Requires

Here is where I want to be careful, because there is a version of this article that turns into hype, and hype is not useful to anyone making a real decision about their finance stack.

An AI agent that drafts a journal entry is not the same as an AI agent that posts it. An AI agent that flags a probable duplicate invoice is not the same as an AI agent that decides not to pay it. The gap between drafting and deciding is where the actual engineering work lives, and it is also where the actual risk lives.

The platforms doing this well have built explicit boundaries around what the agent can do autonomously and what requires a human to approve. Tabs, for instance, runs routine invoice generation, payment reminders, and cash application through its agents automatically, but routes exceptions back to a human reviewer rather than forcing every judgment call through automation. Campfire’s approach to the close leans on a similar principle: automate the reconciliations that are genuinely mechanical, and surface the ones that require actual judgment. NetSuite’s embedded AI capabilities are earlier in this evolution, more focused on anomaly detection and process automation within existing workflows than on autonomous execution, which is a fair and honest place for a platform with NetSuite’s install base and regulatory footprint to be.

The finance leaders getting real value out of this shift are the ones who understand that distinction. The ones getting burned are the ones who bought the marketing claim that everything is now automated, without asking where the actual boundary between automation and human judgment sits.

The Governance Question Nobody Is Asking Loudly Enough

If AI agents are moving from summarizing your finance data to actually transacting on your behalf, the governance question gets more serious, not less.

A chatbot that gives you a wrong answer wastes your time. An agent that autonomously applies a cash payment to the wrong invoice, or closes a period before an adjusting entry lands, or auto-approves a duplicate payment, creates a real financial control problem. The audit trail requirements do not go away because the work is being done by an agent instead of a person. If anything, they get more important, because the person who used to catch a mistake by manually doing the work is no longer in the loop by default.

I ask every finance leader I work with who is evaluating agentic AI capability the same question: what happens when the agent is wrong, and how would you know before it became a problem? If the vendor cannot answer that question with specificity, about audit logs, about exception thresholds, about the actual controls sitting behind the automation, that is a real gap, not a minor detail to sort out later.

SOC 2 compliance and encryption are table stakes at this point. They tell you the platform is secure. They do not tell you whether the agent’s decision-making is governed the way your controller would want it governed if they were watching every transaction personally.

What to Actually Evaluate

If you are looking at a platform that claims agentic AI capability, I would stop asking whether it has AI. Every platform will tell you it does. Ask three narrower questions instead.

What decisions is the agent allowed to make without a human reviewing them first, and what is the actual list, not the marketing description? What triggers a handoff to a human, and how is that threshold set, by the vendor or by you? And what does the audit trail look like six months after the agent made a decision, if you needed to reconstruct exactly what happened and why?

The answers to those three questions tell you more about whether a platform’s AI is real than any demo will.

Where This Actually Lands

I think the honest way to think about this shift is that AI agents are moving down the stack, from the layer where you ask questions about the numbers, into the layer where the numbers actually get produced. That is a genuinely different capability, and it is arriving faster than most finance functions have built the governance muscle to handle responsibly.

Go look at your own finance stack this week. Find the parts where a person is still manually translating between what an AI tool tells them and what actually happens in your system of record. That gap, wherever it sits for you right now, is where agentic AI is going to land next. Better to understand the boundary on your own terms than to discover it the first time an agent gets something wrong.