The real question is not whether you use AI, but who, within your organisation, is accountable for it.
An organisation is judged not by the technology it buys, but by the discipline with which it answers for it. Artificial intelligence lays this truth bare. Many leaders treat it as a tool: it is installed, it is used, and a gain in productivity is expected of it. Yet a tool does not create legal liability, does not produce decisions that can be held against the organisation, and does not reproduce bias at scale. AI does. That is why the real question is not whether you use AI, but who, within your organisation, is accountable for it. This article sets out a durable way of asking that question and answering it, independently of passing fashions and of whatever regulations happen to be in force at a given moment.
From tool to subject of governance
Let us begin with a distinction that I believe to be fundamental. A tool performs a task defined by the person who holds it. A subject of governance, by contrast, commits the organisation beyond the intention of its user: it produces effects that must be capable of being anticipated, traced, justified and corrected. A calculator is a tool. A system that sorts job applications, steers a lending decision, drafts an opinion or interacts with a member of the public is a subject of governance, because it shifts decision-making and responsibility.
This shift is not philosophical; it is highly concrete. When a regulatory framework provides, as in Europe from 2026, for penalties of up to €15 million or 3% of worldwide turnover for breaches relating to general-purpose models and transparency, and up to €35 million or 7% for prohibited practices, AI ceases to be a line of expenditure and becomes a line of risk on the balance sheet (source: Pôle d'excellence cyber, obligations applicable from 2 August 2026; Toute l'Europe). The amount matters less than the principle: a quantified exposure turns a technological curiosity into an internal control matter. And a principle of internal control does not expire when the law changes. It carries over.
Figure 1From expenditure line to risk lineFrom expenditure line to risk line
Why internal control is the right mental model
The usual reflex is to hand AI over either to the IT department or to the legal department. Both are necessary; neither is sufficient. IT knows how to keep the machine running and legal knows how to characterise the risk, but neither manages, day to day, the chain that runs from actual use to accepted responsibility. Professionals in financial governance already know this chain by another name: internal control.
Internal control rests on a simple and robust idea: you can only control what you can describe, assign, prove and correct. A spending process without an audit trail is not a controlled process, however talented the people involved. Exactly the same applies to AI. The talent of teams is no substitute for a control framework. That is why I propose treating AI not as yet another IT project, but as a new internal control scope, with its own map, its identified owners, its evidence and its stop mechanisms.
You can only control what you can describe, assign, prove and correct. The talent of teams is no substitute for a control framework.
The PACT framework: four dimensions of accountability
To make this approach operational, I summarise it in a framework I call PACT, for Perimeter, Authority, Control, Traceability. It is not yet another standard to add to the pile. It is an analytical lens for leaders who want to know, in four questions, whether their organisation genuinely governs its AI or merely uses it.
The framework
PACT: four dimensions of accountability
1
Perimeter
The first dimension, Perimeter, answers the question: what exactly are we doing? It means mapping the actual uses of AI across the organisation, including informal ones, those that nobody has approved but everybody practises. An AI whose use is unknown is the first blind spot of governance. Mapping does not mean drawing up a technical inventory; it means linking each use to a business process and to a purpose.
2
Authority
The second dimension, Authority, answers the question: who decides, and who is accountable? Each mapped use must have a named person, not an abstract department, who can explain why the system is used, what it produces and what it must not do. Authority turns a diffuse technology into named accountability. Without it, responsibility is diluted until it vanishes on the day a problem arises.
3
Control
The third dimension, Control, answers the question: can we stop or correct it? It is both the most neglected dimension and the most decisive. A system that can be neither suspended nor rectified quickly is not governed; it is endured. Reversibility is to AI what the circuit breaker is to an electrical installation: you hope never to use it, but without it everything else is illusory.
4
Traceability
The fourth dimension, Traceability, answers the question: can we prove it? It covers the retention of data, decisions and versions. Governing AI means being able to reconstruct after the fact how a result was produced, from which data and with which settings. The transparency required by regulators is simply the public face of this internal requirement: labelling machine-generated content and telling a member of the public that they are talking to a system are forms of evidence made visible. An organisation that can already prove its case to itself has nothing to fear from external scrutiny.
Figure 2The PACT framework and its four questionsThe PACT framework and its four questions
A use of AI that satisfies all four dimensions is under control. A use that lacks even one of them is a vulnerability waiting to be exposed.
Examples in context: maturity shows in mechanisms, not in rhetoric
Recent history offers useful illustrations, provided they are read as cases in context rather than as scoops. In 2026, several states showed that AI governance is decided upstream of the technology. In Senegal, a first cohort of more than 3,500 education staff, out of 3,909 enrolled, was certified in digital skills and artificial intelligence before any large-scale rollout of tools (source: allAfrica). The transferable lesson is timeless: you cannot govern a system that its users do not understand. Human competence is a prerequisite for control, not an optional extra.
At the level of the State, India has turned its digital public infrastructure into an instrument of public finance, with an instant payment system processing more than 21 billion transactions in the month of January 2026 alone, and a considerable cumulative volume of direct benefit transfers (source: Press Information Bureau, Government of India). What a leader should take from this is not the figure, which will date, but the mechanism: data becomes a management asset when its perimeter, accountability, traceability and reversibility are held together. The day one of these dimensions gives way, the asset turns into a liability.
Figure 3Two cases in context, one mechanismTwo cases in context, one mechanism
Practical lessons for leaders
A few principles for action follow from this framework, and they hold true whether in 2026 or several years from now. First, treat the arrival of a new regulatory obligation not as a legal constraint to be absorbed, but as a test of maturity: it does not create the risk, it makes it visible. Second, refuse to outsource responsibility entirely: a provider can supply the technology, but can never take on governance in your place. Finally, bring AI within the remit of existing control and audit bodies rather than creating yet another committee: effective governance strengthens existing arrangements rather than multiplying them.
Concluding thoughts
Artificial intelligence does not require organisations to invent a new form of governance. It asks them to extend to a new subject a discipline they already know, that of internal control, with its requirements of perimeter, authority, control and traceability.
The fundamental question is therefore not a technological one. It is whether your organisation can still answer for what it does, at a time when a growing share of what it does is produced by systems that never sleep. Technology will keep changing. The capacity to answer for it, however, will remain the true mark of a serious leadership team. One question remains, which every leader can ask today: of the systems we already use, how many would satisfy all four dimensions if the audit took place tomorrow morning?
Sources of the figures
Penalties under the EU AI Act, €15 million or 3% and €35 million or 7% Pôle d'excellence cyber
A budget adjustment should not be judged by the balance achieved, but by the capacity for action it leaves intact. Three filters to assess its quality.