The CFO in AI

Section IThe Productivity Claim Arrives at Finance

A new category of enterprise software promises to compress the monthly close from five days to one, to replace three analysts with one, to produce a board-ready variance report from a natural-language prompt. The claims are not all wrong. Some of them will prove out. The discipline is in deciding which ones, in which sequence, at what cost, and with what governance.

That discipline belongs to finance.

The CFO is not the right person to decide which large language model has the most compelling benchmark scores, or which vector database integrates most cleanly with the existing data warehouse. Those are technical questions with technical answers. But the CFO is exactly the right person to decide what problem the organisation is actually trying to solve with AI, what the cost of solving it conventionally would be, what governance is required to ensure the AI-generated output is trustworthy, and whether the capital allocation is disciplined.

This is a new kind of role. It requires the CFO to hold simultaneously the traditional control-and-oversight function and a new role as capital allocator for a technology category that is moving faster than any capital allocation framework was designed to handle. Neither posture alone is sufficient.

Section IIThe AI-Native Finance Function

"AI-native" is overused, but it points at something real. A finance function built from scratch today — without legacy systems, without the accumulated technical debt of twenty years of ERP customisation, without headcount structured around manual reconciliation — would look materially different from a finance function built in 2005 and progressively modernised.

The AI-native finance function has several structural characteristics that distinguish it from its predecessor:

• Continuous data, not periodic consolidation. The traditional finance function operates on a monthly rhythm because the close cycle requires human time that is finite and expensive. An AI-assisted function can maintain a near-continuous view of financial position. Automated reconciliation runs overnight. Variance flags surface in real time. The close compresses because many of the tasks are automated out of existence.
• Structured and unstructured data, simultaneously. A human analyst works through structured data and supplements it with qualitative context. An AI model can process structured financial data alongside contract repositories, customer feedback streams, operational logs, and external market data in a single analytical pass — a richer picture, produced faster, with explicit citation of the sources that drove each conclusion.
• Exception management, not transaction processing. In a well-implemented AI finance function, the human analyst is not the first pass through the data. She is the review layer for exceptions the automated system has flagged, the judge of whether a flagged anomaly is a control failure or a business event, and the author of the qualitative narrative that no automated system can yet produce with the appropriate judgement and ownership.
• Speed of reporting. The board that waits three weeks for month-twelve management accounts is not reading the same business that existed when those accounts were prepared. Faster financial reporting is not a luxury. It is a governance improvement.

Section IIIAgentic FP&A and the Continuous Forecast

Financial planning and analysis has historically operated on two timescales: the annual budget, which is a commitment and a control document, and the rolling forecast, which is an update. Both are periodic. Both are backward-looking in the sense that they are compiled from data that is already historical by the time the forecast is published.

Agentic AI — systems that do not merely answer questions but execute multi-step analytical tasks autonomously — changes this. An agentic FP&A system, connected to live operational data, can run a continuous forecast that updates as the underlying drivers change. A hotel booking rate that has been running 12% above plan for three consecutive weeks updates the revenue forecast automatically. A supplier cost that has moved with a commodity index updates the margin forecast without a human having to rebuild the model.

The CFO must be clear-eyed about what this capability does and does not provide. It provides speed and consistency. It eliminates the manual assembly work that consumes analyst time at forecast revision. It catches driver movements that would previously have been visible only in the next period's actuals. What it does not provide — and what no automated system currently provides — is judgement about whether the observed trend will continue, whether there is a structural explanation for the movement, or whether the business should change course. That judgement remains human.

The governance of a continuous forecast is also non-trivial. If the forecast updates automatically and the board is presented with a different number each time it meets, the planning function has created noise, not signal. The CFO must design the governance of automated forecasting outputs with the same care she would apply to a human analyst's work: what is the trigger for an updated forecast to be presented to management? What is the confidence threshold below which an automated update should be held for human review? Who is responsible for the output if a forecasting agent makes an error?

Section IVThe Cost Curve of AI and the New Gross-Margin Maths

The economics of AI deployment are genuinely unusual compared to conventional enterprise software. The costs are front-loaded — model selection, fine-tuning, data preparation, integration, governance build-out — and the benefits are back-loaded and often diffuse. The productivity improvement in the finance function does not show up as a single line in the P&L; it appears gradually in headcount leverage, in reduced external consultancy, in faster decision-making.

The CFO must resist two failure modes in evaluating AI investment:

• Undercosting the denominator. Vendor pricing for AI models — cost per token, cost per API call, cost per seat — is only one component of the total cost of ownership. Compute costs for inference at scale are significant and non-linear. Data infrastructure costs — the pipelines, the vector stores, the embedding workflows — are often underestimated. Integration costs into existing systems are routinely larger than anticipated. And the ongoing governance costs — the human review of AI outputs, the model retraining, the incident response when a model produces a materially wrong output — are rarely in the initial business case.
• Attributing too much to AI. When a process improves after an AI deployment, the causal attribution is rarely clean. The process may have improved because the deployment forced a workflow redesign that was long overdue. The headcount reduction may reflect a recruitment freeze coinciding with the deployment rather than productivity from the AI. The CFO who presents AI ROI numbers to the board without acknowledging these attribution challenges is presenting a number that looks more certain than it is.

There is also a structural effect on gross margin that deserves attention. In a business with a high volume of data-intensive operations — pricing, personalisation, fraud detection, supply chain optimisation — AI can compress the variable cost per transaction materially. The cost curve of AI is falling faster than most enterprise software categories. A marginal cost that is material today may be negligible in three years, but the capital and capability investment required to reach that point is real and required now. The CFO is the person who decides whether and when to make that investment.

Section VAI Governance, Controls, and Audit

The governance of AI outputs is the finance function's most significant new challenge. Not because AI is uniquely untrustworthy, but because the consequences of trusting an incorrect AI output without adequate review can be severe — in regulated financial reporting, the consequences are material.

• Model risk management. The banking sector has developed a mature framework for model risk management — the documentation, testing, validation, and governance of quantitative models used in financial decision-making. The same framework, suitably adapted, applies to AI models used in the finance function. Every AI model used to generate financial outputs should be subject to a model risk framework that specifies who validated it, against what benchmark, with what ongoing monitoring cadence.
• Hallucination and data provenance. Large language models can produce confident, well-formatted, plausible-sounding output that is simply wrong. In a finance context, a hallucination that goes undetected can become a material misstatement. The CFO must insist on AI systems that cite their sources, that are constrained to operate on verified data rather than general training data, and that surface uncertainty rather than suppress it.
• Access control and data governance. AI systems given access to sensitive financial data — board materials, M&A analysis, unreported earnings estimates — create data governance risks that go beyond conventional IT security. A model that has ingested confidential information can, in some architectures, surface it in response to queries from users who should not have access. The access control framework for AI systems must be at least as rigorous as the framework for the underlying data.
• Audit readiness. The external auditor will, increasingly, ask whether AI-generated outputs have been subject to human review before reliance is placed on them. The finance function must be able to demonstrate that the review process exists, is documented, and is applied consistently. This is the basic discipline of ensuring the organisation owns its financial outputs, not the model.

Section VIRebuilding the Finance Organisation Around AI

The finance organisation that most efficiently harnesses AI is not the one that deploys AI tools on top of its existing structure. It is the one that rethinks the structure in light of what AI can and cannot do.

The tasks that AI handles well — high-volume, rule-based, data-intensive, pattern-recognition — map closely to tasks historically performed by junior finance staff. Reconciliation, data entry, format conversion, trend identification in large datasets: these are well within the capability of current AI tools. The implication, which finance leaders need to manage carefully, is a compression of the traditional analyst pipeline. If AI performs the work that would have been done by two junior analysts, the career pathway for those analysts — which ran from reconciliation to variance analysis to business partnering — is disrupted.

The finance function that manages this well will do two things simultaneously: deploy AI to handle the high-volume transactional work, and invest deliberately in developing the human capabilities that AI cannot replicate — judgement, relationship, qualitative synthesis, and accountability for decisions. The analyst who was hired to reconcile intercompany accounts must become the analyst who judges whether a reconciling item represents a control failure or a timing difference; who can explain a variance to a business partner who did not study accounting; who can challenge an AI-generated forecast because something in the business context makes the model's assumptions wrong.

This is a more demanding job than the one it replaces. It requires hiring differently, training differently, and measuring performance differently. The CFO who manages this transition well will have a leaner, more capable finance function in three years. The CFO who does not will have a finance function that deployed AI on top of an unchanged structure and achieved neither the efficiency gains nor the quality improvement.

Section VIIThe CFO as AI Capital Allocator

The investment required to build an AI-capable organisation — infrastructure, tooling, talent, governance, change management — is not trivial, and it is not a one-time expense. It is a sustained capital allocation decision that must be made under uncertainty, with imperfect information, in a technology category where the best practice today may be obsolete in eighteen months.

The CFO as AI capital allocator must perform several functions that are not traditionally in the finance job description.

• Technology literacy, not technology expertise. The CFO does not need to understand how a transformer architecture works. She does need to understand the difference between a model that runs on proprietary data within the security perimeter and one that sends data to a third-party API; why a fine-tuned model may be more accurate for a specific task than a general-purpose one, and why that accuracy comes with a maintenance cost; when a pilot that worked well at small scale will break at large scale. This is not deep technical expertise. It is informed executive judgement.
• Milestone-gated investment. The AI project that is approved as a single large capital commitment — build the full thing, then see if it works — is a risk management failure. AI investments should be structured as milestone-gated programmes: a proof of concept with defined success criteria, a pilot with defined scale parameters, a production deployment with defined governance and monitoring. Each gate is a decision point. The CFO's role is to design those gates and enforce them.
• Build-buy-partner discipline. The choice between building AI capabilities in-house, buying them from vendors, and partnering with specialised providers is a strategic decision with long-term implications. In-house build is justified where the AI application is a genuine source of competitive advantage. Buy is appropriate where the problem is common and the solution commoditised. Partner is appropriate where the capability is specialised and incentives are aligned. The CFO must insist this analysis is done rigorously for each material AI investment.
• The counterfactual. For every AI investment, the correct comparator is not "what does this cost?" but "what would we spend to achieve the same outcome without AI?" If the answer is "we would not achieve the outcome at all" — because the data volume is too large or the pattern too complex for human analysis — the AI investment is not a productivity trade-off, it is an enabler of new capability. The capital allocation framework must distinguish between these cases.

Section VIIIA Framework for the AI CFO

The CFO's relationship with AI is not passive. She is simultaneously the investor, the controller, and the beneficiary. The following principles guide that role.