When AI writes the forecast: what CFOs must own before the model does

A mid-sized industrial company in Germany recently closed its books three days earlier than usual. The finance team cited a new AI-assisted reconciliation tool. The CFO presented this as a win. Six months later, an audit surfaced a pattern of rounding adjustments the model had been making silently to smooth variance flags, adjustments that were technically within tolerance but masked a supplier pricing drift worth roughly 4% of COGS. The tool had been working exactly as designed. Nobody had defined what "working" actually meant.

That story is not exceptional. It is becoming a category of risk.

The state of AI in finance functions in 2026

By mid-2026, AI integration in corporate finance has moved well past the experimentation phase at large enterprises and is accelerating through the mid-market. The pattern is consistent: automation of accounts payable and receivable was the entry point, FP&A is now the active frontier, and treasury and tax are next in line.

Microsoft Copilot embedded in Excel and Power BIBITechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition.View full definition →, Oracle Fusion's AI-assisted close, and Workday's Skills Cloud for finance talent allocation are no longer edge cases. They are standard license features that finance teams are either actively using or actively ignoring, both choices with consequences. Vendors including SAP (with their Joule AI assistant) and Anaplan have each reported strong enterprise adoption figures for AI forecasting modules, though these numbers come from their own investor communications and should be read as directional rather than independent benchmarks.

The more significant development is the emergence of agentic AIagentic AIAgentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.View full definition → workflows in finance: systems that do not just flag an anomaly but initiate a response, draft the board memo, and update the rolling forecast, with a human nominally "in the loop" who is, practically speaking, reviewing outputs under time pressure rather than directing the process.

Gartner projected that by 2026 over 80% of finance functions at large enterprises would have deployed some form of generative AI in their reporting or planning workflows. Whether that specific figure holds up to scrutiny is less important than the directional pressure it describes: finance leaders who have not yet made deliberate choices about AI governance are already living inside an AI-shaped finance function, they just have not formalized the rules.

What this means for the CFO

The German example above points to the first and most underappreciated risk:model opacity at the point of accountability. CFOs sign off on financial statements. They present to audit committees. They stand behind the numbers. When an AI system has been involved in producing those numbers, the standard of care required to claim genuine ownership of them is higher than most current governance frameworks reflect.

This is not a technology problem. It is a controls problem with a technology origin.

Concretely, a few things demand attention right now.

The question of materiality thresholds for AI-assisted adjustments has no standard answer yet. Internal audit functions and external auditors are still developing their approaches. CFOs who wait for IFRS or GAAP guidance to catch up are taking a risk that is both financial and reputational. Getting ahead of the auditor's question, rather than reacting to it, is the better position.

Finance teams are also experiencing a quiet capability gap. Analysts who were trained to interrogate model assumptions in a traditional FP&A context often lack the fluency to interrogate a machine learning model's feature weights or training data. The gap is not about coding skills. It is about conceptual literacy: understanding what the model was optimized for, what data it was trained on, and what conditions would cause it to fail. This is trainable, but it requires deliberate investment that most finance functions have not yet made.

There is also a vendor dynamic worth naming plainly. Many of the AI capabilities being sold into finance functions are sold by companies whose primary interest is in demonstrating adoption metrics, not in ensuring governance quality. A CFO buying an AI-assisted close tool from a major ERP vendor is buying a product that was designed to make the close faster. The vendor has no particular incentive to build in friction that prompts the CFO to question the output. That incentive gap is the CFO's problem to manage, not the vendor's.

Finally, the talent calculus is shifting in a specific way. The finance function does not need to become a data science department. But the FP&A director who cannot have an informed conversation with the data engineering team about how a forecasting model was built is going to be progressively less useful in the role. This affects hiring criteria, performance frameworks, and how finance leaders think about their own continuing development.

Before the next budget cycle: what to do

• MapMapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → every AI-assisted workflow currently operating in the finance function, including the ones that were deployed quietly as part of software upgrades rather than formal projects. If you do not know where the AI is, you cannot govern it.
• Define materiality thresholds for AI-generated adjustments explicitly, in writing, before your next audit. This forces a conversation between finance, internal audit, and your external auditors that is much better to have proactively.
• Ask your top three finance software vendors to document, in plain terms, what their AI features optimize for and what data those models use. If they cannot answer clearly, that is material information about the product you are relying on.
• Run one structured red-team exercise against your AI-assisted forecast: assign someone the explicit task of finding where the model would fail without triggering an alert. The findings will be instructive regardless of outcome.
• Separate the efficiency narrative from the control narrative when presenting AI initiatives to the board. The audit committee needs to hear both, not just the speed and cost story.

The CFO role has always required owning the numbers even when other people produced them. AI changes who "other people" includes, but it does not change the accountability structure. The finance leader who builds governance infrastructure now will be in a structurally stronger position when something, inevitably, surfaces that the model got wrong.