AI and the CFO seat

Every AI initiative is born owned by enthusiasm. A sponsor finds budget, a pilot is framed as learning, and for a quarter or two that framing is honest: small spend, unknown outcome, costless exit. The decision arrives later, and rarely reaches the agenda in these words: who owns this money now that it recurs? The subscription renewed. The integration has a maintenance line. Two people spend most of their week on prompts, review and data plumbing. The experiment has acquired the one property experiments are not supposed to have — permanence.

Our position: AI spend moves under the CFO's discipline the moment it recurs in the cost base. Not when the technology matures, not when the organisation feels ready — when it recurs. From that moment it deserves what every other recurring commitment gets: a business case with a named benefits line, an amortisation view of anything capitalised, benefits tracking against the case, and kill criteria exercisable at renewal. Treating recurring AI spend as perpetual innovation budget is not open-mindedness; it is how pilots become permanent costs, because innovation budgets are judged on activity and learning, and no line judged on learning ever fails a renewal.

The standard-setters are further ahead than most operating models. The IFRS Interpretations Committee concluded in March 2019 that a customer's right to access supplier software hosted in the cloud is typically a service contract, not a software intangible. In April 2021 it finalised the harder question — the configuration and customisation costs poured into making a SaaS platform fit the business — and the answer is symmetrical.

US GAAP reaches a similar structure by another route.

The detail matters less than the shared message: this spend is ordinary. It is classifiable, schedulable, amortisable. Nothing in the accounting literature recognises a category called innovation that suspends the discipline.

The oversight side has moved too. The Center for Audit Quality's July 2024 resource places generative AI squarely inside the financial-reporting perimeter, with question sets for audit committees on governance, data, deployment, monitoring, fraud and the regulatory environment. The profession's oversight bodies already treat AI spend and AI use as auditable, questionable, ordinary — a subject for the audit committee's standing machinery, not a protected species.

What the standards and the guidance do not supply is the trigger. That is our judgment, and we state it as a test: recurrence. The second renewal of a subscription, the first budget cycle in which the spend appears without being re-argued, the first headcount whose job description assumes the tool exists. Any of these ends the experiment, whatever the deck says. The measurable symptom of failure is a cost line that grows year on year while its benefits case is still described as learning.

In the corridor the ownership question has a geography. The AI stack of an Abidjan or Dakar operating company is typically contracted by the holding: group licences negotiated in Paris or Geneva, priced in euros or dollars and allocated down through recharges. The local CFO sees an allocation, not a decision. Kill criteria, where they exist, are exercisable at a renewal negotiated two time zones away, for a bundle in which the opco's usage is a rounding line. The costs are real in local currency; the benefits are supposed to appear in local operations that nobody instrumented.

Evidence cards

The discipline has to be built twice, and the second build is where it fails. A group business case proves nothing about Abidjan, so a local benefits line and a local baseline are needed or nothing is tracked. Local statements are drawn under a different accounting framework than the group's IFRS reporting, so the same spend can look like an asset in one book and an expense in the other. A holding board that has never asked how the group's AI recharges land in the opco's accounts is not overseeing AI spend; it is distributing it.

Sources and limitations. The accounting facts rest on the IFRIC agenda decisions of March 2019 and April 2021 under IAS 38 and on FASB ASU 2018-15, both verified against the standard-setters' texts on 2026-08-02; the oversight fact rests on the CAQ's July 2024 publication, a practice aid and graded accordingly. No consultancy material is cited. The recurrence test is the firm's position, graded as interpretation: it is not a measured result, and no client cost or benefits data appear here because none has passed our evidence-release gate. Boundaries are explicit: this note is not accounting advice on any specific arrangement — capitalisation outcomes are fact-dependent and belong with the entity's auditors; the corridor observations on group recharges are operating reasoning, not statistics; and the position does not argue that research spend must show quarterly returns, only that recurring operating spend must stop hiding behind the word innovation.

1. The group financial controller lists every AI line appearing for a second budget cycle and names who approved its change of status. If a line moved from innovation to run-rate without a decision, the move happened by default, and default is not a governance mechanism.

2. The CFO reopens the balance-sheet treatment of each recurring line: what was capitalised, under which standard, over what life, and what was expensed. The test is whether the treatment reflects the applicable interpretation rather than the sponsor's preference for a smaller P&L hit.

3. Each sponsor of a recurring line delivers a benefits schedule naming one reporting line, a baseline and a date. Where the answer is a narrative about capability building, apply the recurrence test and fund it, if at all, as a dated experiment.

4. The contract owner writes kill criteria into the next renewal and confirms that the entity bearing the cost can exercise them. A kill right held where the cost is not felt will not be exercised.