Every finance team has run the same experiment by now. Point a capable model at a real question, watch it produce something that looks right in seconds, and then feel the small hesitation before you put that number in front of anyone who matters. The demo is easy. The number you can defend is not.
That gap is what we came together to talk about. On August 12 we pulled a table of finance leaders together at the Hotel Chelsea for The Finance Leaders' Table, the Agentic Edition. No slides, no product pitch. Just dinner and a real conversation about what actually shifts in the finance function once agents move past basic automation, and what that means for the people leading through it.
A few threads held the whole night together. Here is what stayed with us.
Automation has been quietly running in finance for years. A rule fires, a journal posts, a report refreshes. It is reliable precisely because it does not think. It follows the rule you defined, every time, and it never adapts on its own.
Agents are a different kind of help. They do not just execute a step you scripted. They take on whole jobs (pull the actuals, reconcile the sources, refresh the forecast, draft the variance story) and carry them to a finished answer. The people at the table were clear-eyed about where that already earns its keep: the close moves faster when the reconciliation and the first-draft commentary are done before anyone sits down. The forecast stays current instead of going stale between cycles. Anomalies surface while there is still time to do something about them.
The consensus was not that agents replace the work. It was that agents change which work is worth a human's attention. Coordination becomes the new manual task, and the analyst's job moves up the stack toward the questions that actually need judgment.
The most useful part of the evening was the disagreement.
Nobody in the room wanted an agent quietly making the final call on a number headed to the board. But nobody wanted to hand-check work that a system does perfectly well either. The interesting question is not "AI or humans." It is where the line sits, process by process.
A rough map emerged over dinner. The more a task is governed by a clear, stable rule, the safer it is to let the machine run it end to end. The more a task carries audit risk or lands in front of leadership, the more a named human needs to own the output before it goes. General AI is probabilistic. Finance is deterministic. The teams getting this right are not deploying more tools. They are deciding, deliberately, which tool belongs where, and building the human sign-off into the workflow rather than bolting it on after.
If agents take on more of the assembly, what is left for the finance leader?
More, not less, was the answer around the table. The job tilts further toward the parts that were always the point: framing the right questions, pressure-testing the assumptions behind a plan, and translating the numbers into a decision the business will actually make. Someone put it well: the value was never in producing the report. It was in knowing whether the report was right and what to do about it.
That raises a real question about skills, for leaders and for their teams. If juniors no longer cut their teeth on manual reconciliation, how do they build the instinct for when a number smells wrong? A few people were already rethinking how they train, using the time agents give back to put analysts on harder problems sooner rather than fewer of them.
Here is where the conversation kept landing, no matter where it started.
An agent is only as good as the data it is given. A model built on fragmented, ungoverned data will give you a confident answer, and you will have no way to know whether it is right until it is too late. Building on ungoverned data is a faster way to build the wrong thing.
Several people made the same point from different angles: the tools they already trust, including assistants like Claude, get dramatically more useful when they are working on top of a clean, governed financial layer instead of guessing. That layer is what gives a general model the chart of accounts, the business logic, and the permissions it needs to stop pattern-matching and start retrieving. It is the difference between an answer that looks right and an answer that opens, all the way down to the transaction that produced it.
This is the whole idea behind Cube as the Agentic Finance Layer. Unify the data from every source system into one governed foundation. Let the agents run planning, analysis, and reporting on top of it. And keep every figure traceable back to the source, so the answer holds up when someone asks where it came from. Data goes in messy. It comes out decision-ready.
Finance-grade AI isn't about the model. It's about the data underneath it.
Thank you to every finance leader who joined us. Conversations like this one, off the record and unhurried, are the fastest way we know to figure out what this shift actually asks of us. We will be setting the next table soon.
If you want to see what the governed layer looks like on your own numbers, book a demo. Bring a question from your own close. We'll run it through Charlie in a live sandbox and open the answer down to the transaction.