Most platforms hand AI a simplified copy of your data. Fintastic points it at the live model and lets the engine that runs your plan do the arithmetic.
Why did this move?
idleWorks out what the question needs. No data touched yet.
Returns only the slice of the model that question touches.
The deterministic engine that runs your plan does the arithmetic.
Gross Margin variance, Live Actuals against the 2026 Board Approved Budget.
Runs inside your permissions, on your live model, end to end.
You can trace a number that looks wrong. The harder problem is the one that looks right.
The line read a tenth of what it should. Nothing was miscalculated and nothing was out of balance. The view was wrong, and only a reader who understood how the model was assembled could have seen it.
The report agrees with itself. That's the problem.
Records arriving from the source system without a product code had been landing on a catch-all member, while the forecast sat on the real ones.
Nothing was miscalculated and nothing was out of balance. The view was wrong.
Asked to write a management summary from that version, the Agent refused the column. Revenue came in at a little over half the neighbouring months, and costs came back sign-flipped.
It said so, sourced the month from live actuals instead, and wrote the summary.
Every example on this page is a real customer session, anonymized.
“In my finance career, it's rare to see an interconnected platform of this scale support iterative scenario planning without sacrificing speed or reliability.”
Not a label on a report. The postings underneath it, at transaction-line grain, with the reason attached.
Asked what drove the credits in one account, it returned every posting with vendor, date and invoice, and found they were accrual reversals, not real cost reductions.
Nobody had to write the check first. The Agent derives the relationship from the model's own structure.
Two verticals would not tie on a revenue report. An account categorized as a non-operating adjustment was landing in neither leg. Reconciled to the cent.
Your baseline is untouched. The scenario runs on a copy, in the same engine, with the same dependencies.
Over fifteen turns, a planner and the Agent shaped next year's rate curve together, finishing with model-ready weekly inputs.
The Agent builds it and shows you the result. Nothing enters your model until you approve it. Point it at a formula you already have and it reads the logic, not the label.
A co-pilot, not an autopilot.
Strip any one of them away and the AI is guessing.
Not an extract, an aggregate, or a flattened copy sitting a day behind. The granular truth is already there to be asked.
Fintastic reads dimensions and formula dependencies to find the part of the model the question touches, and serves the Agent only that subset. It never reasons over the whole dataset.
The Agent decides what to calculate. The same deterministic engine that runs your plan performs it, so the same question on the same version returns the same number.
The Agent runs as you. Two people can ask the same question and receive two correctly scoped answers.
It retains conventions, terminology and standing rules, including what a term like "bookings" means in your model.
The Agent reasons. The platform serves. The engine computes.
a schematic. cell counts are illustrative, not a measured ratio
Dimensions and formula dependencies decide what a question touches. The platform serves the Agent that subset and nothing else, at full granularity, inside the permissions of the person asking.
whole model → structure-aware lookup → the subset the question touches
A remembered number goes stale. A looked-up number never does.
The real test is not whether an Agent can produce an answer. It is whether that answer is accurate, reproducible, and grounded in the model.
Same model, version and permissions. The answer should be reproducible, because the calculation comes from the planning engine and not from the AI estimating a result.
A trustworthy Agent tells you why the model cannot support the request instead of producing a plausible answer anyway.
You will quickly see whether it understands the model and its dependencies, or is simply interpreting a report label.
We'll run these tests on your model.
A reasoning Agent for enterprise planning and analysis that works directly with the Fintastic model. It traces numbers to their drivers, explains and debugs formulas, checks reconciliation, builds model objects on request, and creates scenarios using the same model and calculation engine that run your plan.
No. The AI determines what needs to be calculated. Fintastic's deterministic engine performs the calculation, so the arithmetic is never the AI's to produce.
Yes, at full granularity, including its dimensions, formulas, versions and dependencies, rather than an extract, aggregate or separate copy. The Agent is served only the slice of the model each question touches, never the whole dataset.
Both. Ask it for a metric, a list, a dimension, a report or a formula and it builds one and shows you the result. Nothing enters your model until you approve it.
Not on its own. Scenarios are built and tested on a copy of a version, so your baseline stays unchanged, and any model change waits for your approval.
Yes. It runs within the permissions of the person asking. Two people can ask the same question and receive two correctly scoped answers. There is no elevated path around the access model already configured in Fintastic.
No. Your model data and business-specific memory are not used to train shared or foundation models.
Yes. Fintastic's MCP server allows compatible AI assistants to query model information under the same permission boundaries enforced in Fintastic.
Trace the number. Understand what drove it. Test what happens next. Find what the report missed.
Book a working session
