The finance function, rebuilt AI-native

There are two ways to approach AI in the finance function. The first is to treat it as a better tool for existing tasks — AI that drafts the variance commentary, categorises expense claims, reads supplier invoices. This is the automation frame. It is real, it produces savings, and it is not what this essay is about.

The second is to ask what the finance function would look like if designed from scratch today, with AI as a native capability rather than an add-on. That is the redesign frame. It is harder, more disruptive, and produces substantially more durable advantage.

Section IWhat AI changes about the speed of information

The conventional finance calendar was built around the assumption that useful financial information is produced at defined intervals: monthly management accounts, quarterly forecasts, annual budgets. The interval is set by the speed of the process — the minimum time required to gather, process, reconcile, and review the underlying data.

When AI is applied to the data pipeline — ingesting transactions continuously, applying classification automatically, running reconciliations in real time, flagging anomalies as they appear — the calendar constraint starts to dissolve. The transaction layer is always current. Management accounts become a query against a ledger that is already up to date, not a reconstruction exercise run at period end.

The value is not primarily the speed — it is the decision-making quality that continuous information enables. A CEO receiving management accounts fifteen days after period end is making decisions against information that is six weeks old by the time the prior period has been fully processed. A CEO with a continuous view is making decisions against current information. That is a different quality of management, and the finance function that enables it is playing a different role.

Section IIWhat AI changes about the allocation of human attention

The conventional finance team spends a significant proportion of its time on work that is, in process design terms, non-value-adding: gathering data that already exists, re-entering it into systems that should talk to each other directly, reformatting outputs that should be generated automatically, chasing approvals that a workflow system could route.

AI eliminates most of this. Document processing — reading supplier invoices, extracting relevant fields, posting the journal, matching to the purchase order — is a category that AI handles at scale and with reliability approaching or exceeding human accuracy for structured documents. Expense categorisation, bank reconciliation, intercompany elimination in a group consolidation, VAT return preparation from classified transaction data — all candidates for automation, not because they are unimportant but because the judgment content is low relative to the volume.

What remains when the low-judgment work is automated is the high-judgment work: interpreting what the numbers mean, identifying what is changing and why, deciding what to do about it, communicating it to the board and to investors, and maintaining the controls framework that makes all of the above trustworthy. This is where a skilled finance professional's time is most valuably spent, and it is the category that AI cannot perform independently.

The AI-native finance function has fewer people in data processing roles and more in analysis, interpretation, and governance roles. It produces more outputs, of higher quality, at greater frequency. The transition is uncomfortable and requires significant investment in tooling and process redesign. It is also, within a few years, unavoidable for any finance function that wants to remain competitive.

Section IIIThe controls question

The finance function's control environment is not incidental to its work — it is a large part of the reason the work has value. Financial statements are credible because they were produced by a system with controls: segregation of duties, authorisation hierarchies, reconciliation disciplines, audit trails, access restrictions.

AI introduces new risks to the control environment that must be managed explicitly. The most significant is systematic model error: an AI system classifying transactions or generating journal entries makes determinations based on patterns in training data. Those patterns can be wrong — systematically, in ways that are difficult to detect precisely because the errors are consistent rather than random. A random human error is caught by a competent reviewer; a systematic AI error propagates until someone asks why the model's output no longer corresponds to economic reality.

The control framework for AI in finance therefore requires: documented decision rules for every automated process so that a reviewer can understand and test the logic; exception reporting that surfaces the transactions the model was uncertain about rather than burying them behind a confident-looking classification; periodic back-testing against human review of a random sample; and clear escalation rules for transaction types or amounts exceeding a defined materiality threshold. The human in the loop is the governance of AI, not the absence of it.

Section IVAI as a capital allocation tool

The CFO's role in AI extends beyond the finance function's own operations. As AI investments grow across the business — in infrastructure, custom models, data engineering — the finance function is increasingly the decision-maker on capital allocation to those investments.

AI investments are structurally unusual. Upfront costs are visible: compute infrastructure, engineering time, data preparation, model training. Returns are often diffuse and difficult to attribute: a faster close, better forecast accuracy, lower transaction error rates. Standard NPV analysis, built for cash flows with predictable timing and attribution, does not model these returns well. The finance function that is AI-native has a significant advantage here: having observed the economics of AI deployment on its own processes, it has better-calibrated intuitions about where returns are real and where they are theoretical — and that calibration, applied to investment decisions across the rest of the business, is a capability no external benchmarking can replicate.

Section VWhat the rebuilt function looks like

An AI-native finance function has a different shape: fewer people in transaction processing, more in analytical and governance roles; an integrated technology stack where data flows automatically between systems; reporting that is continuous rather than periodic; and a control framework maintained as actively as any other operational process.

Building it requires sustained investment and a willingness to change workflows that have been in place for years. The finance teams that make that investment now will have a significantly more capable function in three years than those still automating at the margin.

The task is not to use AI better. The task is to build differently.