Governing what we do not fully control: internal control put to the test by decision-making technologies
Perspective, Decision-making technologies
PACTPerimeter, Authority
Control, Traceability
Never use what you cannot answer for. A test of management maturity, not a technical test.
Every organisation eventually adopts tools that are faster than it is. First the spreadsheet, then integrated enterprise software, and now automated decision engines. With each wave, the same vertigo returns: technology enters processes faster than the organisation can explain what it does. The problem is therefore not new. What is changing is the degree of autonomy of the tool and the speed at which a decision can be taken without any human knowing precisely why. This article proposes a lasting framework for managing this grey area, in which an organisation uses a tool it does not fully understand without giving up on being accountable for it.
A confusion to dispel: controlling a tool is not the same as understanding it
Let us begin with a distinction that I believe to be fundamental. To understand a tool is to know how it works on the inside. To control a tool is something else entirely: it means being able to say where it acts, what it decides and who answers for it, and to prove it. These two requirements do not overlap. A chief financial officer has never needed to know how to program the accounting software in order to stand behind its financial statements. In the same way, an organisation does not need to understand the inner workings of a statistical model in order to govern it. It needs to keep a firm hold on it.
I define control of a decision-making tool here as the verifiable capacity to answer for its effects, regardless of its internal complexity. This definition shifts the burden. It does not ask the organisation to become an expert in technology. It asks it to remain an expert in accountability. And that competence, unlike the other, cannot be delegated.
Figure 1Understanding the inside, controlling the effectsUnderstanding the inside, controlling the effects
Why the issue is becoming pressing: a case in point from 2026
A recent regulatory signal illustrates this shift. In 2026, the European regulation on artificial intelligence (the AI Act) reached a new stage. From 2 August of this year, transparency obligations and a market surveillance system became enforceable, backed by a penalty regime that is no longer merely symbolic: up to €35 million or 7 per cent of total worldwide annual turnover for prohibited uses, and up to €15 million or 3 per cent for providers of general-purpose models. Sources: the penalty regime is set out in Article 99 of the Regulation, available at https://artificialintelligenceact.eu/article/99/, the regime for model providers in Article 101, at https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-101, and the implementation timetable is detailed by the Direction générale des entreprises (French Directorate General for Enterprise) at https://www.entreprises.gouv.fr/decryptages-de-nos-experts/le-reglement-europeen-sur-lintelligence-artificielle-publics-concernes.
I cite this case as an illustration, not as the subject. The principle it reveals is more enduring than the text itself. Sooner or later, a regulator always ends up asking an organisation to prove that it is in control of what it uses. The amounts change, the acronyms change, the requirement does not. An organisation that learns to answer for its tools before it is compelled to do so turns a constraint into an advantage. One that waits for the first inspection discovers its own workings in a hurry.
Figure 2On 2 August 2026, the requirement acquires a priceOn 2 August 2026, the requirement acquires a price
The PACT framework: four proofs for keeping hold of a tool that is not fully understood
To make this control operational, I propose a four-point grid, which I call the PACT framework. It does not measure the technology. It measures the organisation's capacity to answer for it. Each letter corresponds to a proof that the organisation must be able to produce calmly and in advance, before a third party asks for it.
The framework
PACT: four proofs to produce
1
Perimeter
Perimeter. Where does the tool actually act? The first governance failure is not the tool's error, but ignorance of its presence. Many organisations use automated decision-making in recruitment, relations with service users or customer scoring without keeping an inventory of it. The proof of perimeter is a map: the list of places where a tool influences a decision.
2
Authority
Authority. Who decides and who answers for it? A tool bears no responsibility. Responsibility remains human, always. The proof of authority consists in naming, for each mapped use, the person who approves it, the person who can suspend it and the person who is accountable for it. A use with no name attached is an orphan use.
3
Control
Control. What mechanism checks the tool over time? A tool that is accurate at the outset can drift, because the reality it processes changes. The proof of control is the existence of regular checkpoints: a tracked indicator, an alert threshold, a periodic review. Without them, the control achieved on day one quietly expires.
4
Traceability
Traceability. What record is kept of a past decision? On the day a decision is challenged, the organisation must be able to reconstruct the basis on which it was taken. The proof of traceability is the retention of a minimal log: which version of the tool, which data, which decision, and when. It is the memory of accountability.
Figure 3The PACT framework: four proofs to produce in advanceThe PACT framework: four proofs to produce in advance
These four proofs require no advanced technical skill. They require management discipline. That is precisely the business of internal control, a long-established function that decision-making technologies do not make obsolete but bring back to the fore.
What this changes for performance management: three transferable lessons
The first lesson is a reversal of the usual order of priorities. Faced with a new tool, the common reflex is to start with the technology and think about control afterwards. I propose the opposite. Start with a one-page map, even before deployment. Knowing where the tool is placed always comes before the art of tuning it.
The second lesson concerns the allocation of roles. Control of a decision-making tool is not the business of the IT department alone. It is shared between those who use it, those who check it and those who answer for it. Entrusting the governance of a tool to those who installed it amounts to asking a party to be its own judge. Segregation of duties, a founding principle of audit, retains its full force here.
The third lesson is strategic. Control of a tool is not merely a protection against penalties. It is a choice of sovereignty. Deciding which tools enter the organisation, with what data and under which jurisdiction, means deciding on one's own dependence. An organisation that can answer for its tools chooses its dependencies. An organisation that cannot simply endures them.
Never use what you cannot answer for.
An open conclusion
The real question raised by decision-making technologies is not whether we understand them. We have never fully understood our most powerful tools, and yet we have learnt to govern them.
The question is whether we remain able to answer for them as they gain in autonomy. It is a test of management maturity, not a technical test. The organisations that pass it will not be the most technologically advanced. They will be those that have kept alive a requirement both simple and exacting: never use what you cannot answer for.
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.
The P.R.E.T. grid (Performance, Risk, Échéance or time horizon, Transparency) offers an analytical framework to assess the consistency of executive remuneration, in both the private and public sectors.