There is no shortage of predictions about what AI will do to financial planning. What's rarer is observation: what finance teams actually do with an AI agent once it's sitting inside their planning environment, connected to their live model, during a real close or a real forecast cycle.

Our customers tell us. The patterns are consistent, and they invert most of what this category's marketing says.

Pattern one: they ask it to build, not to explain

The expectation is that finance teams use AI to summarize: explain this variance, write this commentary. That phase exists, and it's short.

The dominant use, and it becomes dominant within weeks, is construction. Write this formula. Debug this calculation. Tell me why this logic returns the wrong number. Teams start by asking the agent to explain the model they have, and as soon as they trust the answers, they ask it to help build the model they want. FP&A leads tell us the agent has become the first stop for formula logic they'd previously have escalated to a solutions consultant or a power user, which changes who can build and how fast.

Explaining is reading. Building is writing. The second is a much harder job for an AI, and it's where the demand goes almost immediately. Any evaluation that only tests the summarization use case is testing the phase teams grow out of first.

Pattern two: the second biggest job is trust work

The next largest category of questions isn't analysis. It's verification. Why do these two figures disagree. Where did this number come from. Which of these versions is the one that counts.

The questions look mundane and they are anything but. When two views of the same metric diverge, the useful answer isn't a guess or a summary. It's a trace: back to the source entries, showing that the two figures measure subtly different things, corroborated by the underlying data. That's what turns an AI answer into something a finance professional will actually act on.

Sometimes the trace goes further than any person would. In one pattern we've now seen more than once, a reconciling item that a team had hunted for manually was resolved because the agent connected it through a text memo on a journal entry, a reference buried in millions of rows that no analyst would realistically sift to find. Not a calculation, a correlation in text. That class of answer only exists when the agent can reach the transaction level of the live model.

Two more behaviors turn out to matter as much as the answers. The first: when a question is ambiguous, the right response is to ask for clarification before searching, not to guess and answer confidently. Users consistently call this out as the thing that builds their confidence, because it prevents the assumption errors they'd make themselves. The second: when a user asks for analysis on a version that's still mid-load, the right behavior is to decline and say why, because a summary built on partial data is confidently wrong. An agent that always answers is less trustworthy than one that knows when it shouldn't.

This is the unglamorous center of AI in finance: not predicting the future, but proving where the present came from. Finance teams verify. An agent that can't show its work gets checked manually every time, which erases the point of asking.

Pattern three: the frontier is what-if, and most teams aren't there yet

The strategic questions, what would it take to add this much to EBITDA, which lever moves margin fastest, are the smallest share of real usage. Not because the capability is missing, but because trust is sequential. Teams graduate: first explain, then build, then verify, then simulate.

The what-if layer is where the strategic value concentrates, and what customers describe says most teams are still two steps before it. Which means the vendors selling autonomous strategic insight are selling the last step to buyers who haven't been given a reason to trust the first three.

What this says about the AI-in-planning debate

The marketing narrative across this category leads with insight generation and autonomous analysis. What customers actually report leads with construction and verification. The What is cheap. The Why, the How, and the What-If are where the work is.

And those are exactly the questions that require the agent to operate on the live model, with lineage down to the source transaction, rather than on an extract or a summary. An agent can only trace a number to its source if the model it reads is the model the number lives in. It can only debug a formula it can see. It can only decline to use half-loaded data if it knows the load state.

One design choice underneath all of this: the language model doesn't do the math. Large language models get lost in large datasets and are unreliable calculators, so in Fintastic the agent's job is to reason about which slice of the model matters, and every calculation and comparison runs in the platform's deterministic calculation engine. The agent narrows, the engine computes. That's why an answer can take a moment, a single question can involve many reasoning cycles against the live model, and it's also why the answer arrives with a trace instead of a guess. None of that survives a nightly sync into a separate AI layer.

The architecture isn't a detail under the AI story. It is the AI story.

See it answered live

This Thursday, August 20, our Head of Solutions Anthony Losurdo is taking exactly these questions live: why planning is still slow, complicated, and inaccurate in the year everyone got AI, and what the Why, the How, and the What-If require from the system underneath. Live Q&A.

Save your seat