Forecasting with AI: how Claude and ChatGPT build and calculate your model — and which assumptions stay human work (2026)

Forecasting with AI: how Claude and ChatGPT build and calculate your model — and which assumptions stay human work

From evenings maintaining formulas to a model the AI builds and the data layer keeps filled — with the drivers finance turns itself.

Can AI create a forecast?

Not in the sense the question usually means — and that is good news. A forecast is not a prediction that falls out of a model somewhere, but a set of explicit assumptions calculated through consistently: this much revenue growth, this many hires from that month, that contract indexing per quarter. An AI that “predicts revenue” without those assumptions extrapolates a series and wraps it in confident language. For a trend line that suffices; for a forecast that a bank, investor or management team steers on, it does not.

‍

What AI cán do is precisely the part of forecast work that costs the most time. Building and maintaining the model: the structure, the formulas, the link between revenue, costs, personnel and cash flow. Filling the model: retrieving actuals, updating series, guarding the tie-out with the reporting. And querying the model: calculating what an assumption does, signaling deviations, showing the result’s sensitivity to every driver. In that work a language model is not just as good as a human — it is faster and more consistent at it.

‍

The division of roles is thereby the same as in creating management reports with AI: the AI builds, calculates and signals; finance chooses the assumptions and carries the story. The conceptual foundation — what a good forecast even is, from rolling forecast to latest estimate — sits in the main article on forecasting for SME CFOs; this article is about the AI setup on top.

And as with the reporting: the result is as good as the data beneath it. So that first.

Which data does an AI forecast need?

Three layers — and all three have to be right before the model is worth anything.

‍

The actuals as the foundation. Every forecast starts where realization ends: the most recent months determine the starting point of every series. Those actuals must be consolidated — cleaned of intercompany flows, across all entities — and sit in the own reporting structure, so the forecast lines tie one-to-one to the monthly report. A forecast in a different layout than the reporting is a guarantee of tie-out differences — and for a group, that includes the group picture, not just the addition per entity.

‍

The history as context. Assumptions are tested against series: seasonal patterns, the actual growth rate of recent years, how costs moved with revenue and headcount. Because the data layer holds the full history — not just the current financial year — the AI can put every assumption directly next to the multi-year course: “you are using 4 percent cost inflation; the past three years averaged 6, and last year was 7”.

‍

The existing forecast as reference. Through the Finstack MCP, next to the actuals the recorded forecast is available as well — including the budget, which in Finstack is a recorded form of forecast. So the forecast too comes straight from Finstack — the same source as the actuals, no separately supplied file — and with that, the comparison works today. How important that layer is shows when it is missing: in early practice we saw an assistant calculate a cash runway and note by itself that it was an indication, not a projection — because there was no forecast in the source to set it against. Exactly right reasoning, and exactly why the forecast belongs in.

‍

All three layers the assistant retrieves itself through the MCP connection: read-only, at aggregate level, current at the moment of asking. No exports, no copies of the model drifting out of step. How that connection works sits in connecting AI to your accounting software.

How do you build a forecast model with AI?

Not from scratch — from the model that already exists. Virtually every finance team has an Excel forecast model holding years of thinking: the build-up from revenue to cash flow, the cost lines, the logic per item. That model is the best blueprint the AI can get: have it rebuild the structure, and the result is a calculable model the team already understands.

‍

What that looks like in practice, we saw with a user who had Claude rebuild his existing Excel model into his own forecast application. The build-up shows what a mature model needs: tabs for P&L, balance sheet, cash flow and topline, a headcount plan and a cost overview per vendor, and a separate tab with assumptions. And per cost line an explicit driver: fixed with an inflation uplift, growing with headcount, growing with revenue, or manual input for items that resist capture. Personnel costs calculate through from the plan — a new hire in month three automatically means salary and employer costs from month three.

‍

That driver structure is the core of a maintainable model: every line has one explicit logic, and whoever disagrees with something knows exactly which knob to turn. The AI moreover guards the consistency — a revenue assumption automatically works through into the revenue-linked cost lines and the cash flow, without anyone dragging formulas.

‍

Important to keep clean: this model lives with the user, in their AI environment or spreadsheet — it is nót Finstack functionality. The data layer delivers what keeps the model filled and honest: the actuals, the history and the recorded forecast, always current through the connection. Exactly the division of labor that works: the thinking in finance’s model, the data from one source.

How do you work with scenarios and assumptions?

The biggest gain of an AI-built model sits not in the base scenario, but in what becomes possible after it: calculating variants without breaking the model. In a classic Excel model, a scenario is a copy of the file — with all the version misery that brings. In a model with explicit drivers, a scenario is a different set of assumptions on the same structure: a base version that stays fixed, and alternatives you keep and compare next to it.

‍

The working method taking shape in practice: keep one base scenario as the anchor — the forecast the management team steers on — and have the AI calculate variants next to it on request. What does half a year’s delay in the hires do to the cash position? What if revenue growth is not 15 but 8 percent? The answer is there in seconds, including the pass-through to costs and cash flow, and the base scenario stands untouched.

‍

The discipline sits in the assumptions themselves. Record them in one place — the assumptions tab from the example above is not a detail but a design principle — and give every assumption an owner and a date. Have the AI name, with every calculation, whích assumptions drive the result hardest: that sensitivity analysis is worth more to the meeting than the point figure of the outcome.

‍

And have the AI chállenge assumptions, not invent them. “Test this growth rate against the history” is an excellent instruction; “what will revenue be next year” is not. The difference seems subtle, but it is exactly the difference between a forecast with a story and an extrapolation with a bow on it.

How do you explain forecast versus actuals with AI?

This is the monthly heartbeat of the forecast process — and the place where the AI delivers most. Every month, as soon as the administrations are closed, the question stands ready: where does realization deviate from what we expected, and what does that mean for the rest of the year?

‍

Because the MCP delivers both the actuals and the recorded forecast, the assistant can answer that question directly at every level the reporting knows: per entity, per cost center, per line of the reporting structure. The working approach is the same as in the variance analysis: have the AI rank the deviations — largest first, with amount, percentage and the course of the series — and have it mark what looks incidental and what structural. The question finance teams literally ask here: “we planned thirty people and twenty-five are working — show me where that difference sits in the costs.” For an assistant with the right data, that is a question of seconds.

‍

The explanation stays human work here too — why the hires came later or a vendor invoiced earlier, the organization knows, not the dataset. But the difference with before is large: the analysis no longer starts with an afternoon of tying out, but with a ranked list in which the attention is already steered.

‍

The outcome of that monthly loop is the adjusted expectation: structural deviations work through into the assumptions, and the forecast shifts to a new latest estimate. That keeps the model a living steering instrument instead of an annual ritual.

How does the AI forecast fit into the monthly rhythm?

As a fixed step áfter the reporting draft round. The rhythm that works: the administrations close in the accounting package, the data layer syncs, the reporting analysis runs — and directly after it the forecast loop: actuals against forecast, ranking deviations, adjusting assumptions where needed, and setting the calculated latest estimate ready for the meeting. One whole with the rhythm from automating management reporting, not a separate process alongside.

‍

For whoever keeps the model in Excel or Google Sheets, the same loop exists through the other route in the Integrator pack: the 2-way sync automatically puts the actuals next to the forecast columns in the own model, and the AI assistant in the spreadsheet analyzes them there. Both routes work on the same data basis — which one fits is a question of where the team prefers to work, not of functionality.

‍

Sharing runs as with all steering information: the forecast that matters sits next to the actuals in the reporting environment, where management and investors look along with rights per user. The AI environment is the workshop; the reporting is the shop window. Whoever keeps the two separate prevents a half-calculated scenario accidentally making the meeting.

‍

And the rhythm reinforces itself: the same loop every month, a sharper fit between assumptions and reality every month — and a forecast that earns the trust that decisions rest on. Do plan the loop at a fixed moment, right after the closing day: a forecast updated halfway through the month steers on a picture that is already stale — and precisely the disappearance of the collection work makes that fixed moment feasible now.

What do you need to start — and what does it cost?

The same three building blocks as the reporting application. The data layer: a Finstack environment with connected administrations and a configured reporting structure — first connection in about 5 minutes, full setup typically within a day, from EUR 39 per month with no implementation fees, with a 14-day free trial. The AI access: the Integrator pack, with MCP rights for whoever may see the full consolidated figures. And a business license for Claude, ChatGPT or another MCP-supporting assistant, with training on company data off.

‍

Start with the existing Excel model as the blueprint — not with an empty canvas. Have the AI rebuild the structure and the drivers, record the assumptions on one tab, and run the first month in parallel: the AI model next to the trusted model, same actuals, same assumptions. Where the outcomes differ sits either a mistake or a hidden assumption that was never written down — you want to find both befóre switching, and that second category proves the most valuable find in practice.

‍

For the business case, do not count saved hours alone. The real gain sits in what does not happen today because it is too much work: scenarios that go uncalculated, assumptions untested against the history, a latest estimate updated only twice a year. With the model in the AI workshop and the data from one source, all of that becomes monthly work instead of project work.

‍

What else this landscape offers — and why the forecast model belongs living with finance itself — sits in the overview of the best reporting software for SMEs.

The 3 most common mistakes in forecasting with AI

Three patterns we keep seeing at teams using AI in the forecast process. Each costs reliability; each is preventable with the right setup.

Sources and provenance

• Finstack — Spreadsheet sync & Integrator pack: finstack.io/solutions/spreadsheet-sync
• Finstack — Reporting & insights (reporting structures, forecast next to actuals): finstack.io/solutions/reporting-insights
• Finstack — Pricing (from EUR 39/month): finstack.io/pricing
• Model Context Protocol — open standard for AI data connections: modelcontextprotocol.io
• Finstack user practice — AI-built forecast models (anonymized)

Last reviewed: 1 September 2026 · Next review: December 2026