How boards can gain from uncertainty
Surviving disruption is no longer enough. Jean-Jacques Henchoz outlines five ways that boards can learn, adapt, and grow stronger through volatility....
Governance
Trias Algorithmica: Why leaders must separate the powers of AI
Whether artificial intelligence will deliver on its promise, and whether it can be trusted, preoccupies corporate leaders and boards. But a more fundamental question is being neglected: Who inside the organization actually has the authority to define the rules AI follows?
Rapid read:
- AI can concentrate rulemaking, execution, and judgment within one system or team, weakening corporate checks and balances.
- Separate authority over objectives, data, deployment, monitoring, and appeals. Give independent reviewers the information and power to challenge decisions and stop harmful systems.
- Boards should assess what AI is authorized to do, who controls each stage, and whether employees and customers can meaningfully challenge its decisions.
In November 2022, after a death in the family, Jake Moffatt went to Air Canada’s website to find out how the airline’s bereavement fares worked. The airline’s chatbot suggested that it was possible to purchase a ticket and retrospectively apply for a reduced bereavement rate within 90 days. There was one problem: the chatbot’s instructions were wrong. Air Canada’s actual bereavement policy did not permit the retroactive application described. When Moffatt later requested the refund, the company refused it. In February 2024, the British Columbia Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation. The tribunal concluded that the airline was responsible for information provided through its website, whether that information appeared on a conventional webpage or through a chatbot.
The amount at stake was small. The institutional problem was not: algorithmic rules were not aligned with corporate policy. One part of a company had established a rule, and an algorithmic interface communicated a different rule to the customer. The customer acted on the information given. Another part of the company then judged the customer’s claim against the official rule rather than the rule its own machine had presented.
The chatbot had not merely answered a question incorrectly. For Moffatt, it had interpreted corporate policy. Once algorithmic systems begin interpreting policies, they begin doing something more consequential than producing information. They begin exercising institutional power inside and on behalf of the firm. In this case, the same algorithmic interface concentrated several forms of power: a rulemaking power by determining in practice what the policy meant for the customer; a rule-execution power, by operationalizing that interpretation in the interaction; and a rule-adjudication power, by determining the customer’s rights under its interpretation of the policy in light of the circumstances and the corresponding obligation of the company.
Traditional systems of corporate governance are designed, in part, to prevent the danger that arises when different kinds of power are concentrated. A business unit does not normally get to prepare the company’s financial accounts and serve as the independent auditor of those accounts. A trader is not supposed to determine the firm’s risk limits, execute trades without constraint, and retrospectively decide whether the resulting exposure was acceptable. In regulated banking, model development and model validation are governed through differentiated responsibilities and effective challenge. Boards oversee management precisely because ownership, execution, and oversight are not assumed to be the same function.
Yet as companies deploy algorithmic systems, a new governance threat is emerging. Powers that traditional corporate structures were designed to separate can first recombine inside the same organizational structure and then become fused in the algorithmic architecture itself. Responsibility for AI development and governance can remain heavily concentrated around technical and product development functions while cross-functional challenge and independent scrutiny remain weak or absent. Where responsibilities for product development, safety, AI governance, and oversight are insufficiently separated, a single product organization may concentrate enough power to determine what a system should optimize without being subject to independent challenge or enterprise-wide review of the broader consequences. Its engineers may select the data, design the model, and establish the thresholds at their own discretion. That same product organization may deploy that very algorithmic system within the business unit to which it belongs, determine the metrics, select the benchmarks, define what constitutes success for purposes of evaluation, interpret its performance, decide whether an incident is material, and judge whether someone affected by its decisions deserves reconsideration, with no other authority empowered to challenge that judgment.
A 2025 PwC survey found that 56% of executives said first-line teams, including IT, engineering, data, and AI, led their organizations’ responsible AI efforts. A 2026 benchmark study of 500 senior legal and executive leaders across industries provides a sharper picture of that concentration. IT or technology teams participated in AI governance at 80% of organizations, compared with only 35% for legal and compliance, while just 11% described legal and technical teams as highly collaborative. The same study also found that governance was far more common while AI systems were being developed and tested than at the points where organizations decide whether to deploy them or assess them once in operation; 72% applied governance during development and testing, compared with only 37% at deployment readiness and 44% during post-deployment monitoring.
The concentration does not stop at the organizational level. Once those concentrated decision-making powers are translated into an algorithmic system and that system is granted operational authority, the organization has effectively delegated its concentration of power to the algorithm. The algorithmic powers of rulemaking, execution, and judgment can then collapse into a single chain of command, placing all three powers in the same hands. In corporate governance, we have spent decades building controls to prevent precisely this kind of concentration. In AI, we often call it product ownership.
Before organizations ask whether AI is trustworthy, they must address an initial question of legitimacy:Â under what governing arrangements is the exercise of algorithmic power legitimate in the first place, and under what conditions should an algorithmic system be authorized to exercise that power on behalf of the firm? The distinction is between trustworthiness of output and legitimacy of the authority exercised; the former does not by itself validate the latter.
Trias Algorithmica: separating algorithmic powers
The three powers are distinguished by their institutional functions, even when those functions are not separate technical operations. In practice, a threshold can embody both rulemaking and judgment; monitoring can lead to new rules, and a recommendation can shape action without formally executing it.
But examining each function separately focuses the organization on how authority is distributed around those functions. Separation may be organizational, procedural, technical, or some combination of the three.
1 – Algorithmic legislative power: the ability to encode the rules
Every advanced algorithmic system contains choices about what should be optimized, what should be measured, which errors matter most, what evidence counts, what behavior triggers intervention, and what the system is permitted to do. A recruitment system needs criteria by which applicants will be ranked. A fraud system needs a definition of suspicious behavior. A pricing model needs to determine which signals matter and how aggressively prices may respond. An employee-performance system needs to decide what constitutes satisfactory work. A customer service agent needs to know which commitments it is empowered to make.
Other choices look like purely technical decisions. A threshold establishes how much confidence is sufficient to block a transaction. A loss function privileges one type of error over another. A recommendation objective determines whether revenue, retention, conversion, engagement, customer satisfaction, safety, or some combination of these constitutes success.
These may be engineering decisions. But they are not merely engineering decisions. They distribute attention, suspicion, risk, opportunity, workload, and even employment itself. A corporate rule does not cease to be a rule because it is expressed as a parameter rather than a policy manual.
2 – Algorithmic executive power: the capacity to turn those encoded rules into action
A forecasting model contained in a data scientist’s notebook has limited institutional power. But connect it to procurement, pricing, hiring, customer communications, or any functional system that enables enterprise processes, and its outputs can begin to trigger or determine real-world corporate action with potentially life-altering effects.
Everyone sees the difference between an algorithm permitted to recommend a price and one permitted to change the price. Likewise, an AI that identifies a suspicious transaction is different from one empowered to block the customer. Execution transforms rules into operative power governing business reality. The relevant resource is not only compute. It is also permission given through API access, transaction authority, or workflow integration.
3 – Algorithmic judicial power: the authority to determine whether the system has behaved acceptably
Who determines that the model remains accurate enough? Who decides that a flagged transaction was actually fraudulent? Who reviews a rejected candidate? Who decides whether a model failure justifies stopping deployment? And, crucially, who hears the appeal? All of this is judgment. When the algorithm itself becomes the effective final authority on those questions, algorithmic judicial power has effectively been concentrated in the system, bringing with it a higher risk of consequential harm. This function also judges whether employees, customers, suppliers, counterparties, or other actors subject to it have complied with its rules.
Establishing the legitimacy of algorithmic power requires separating these three powers. What is needed, then, is a governance architecture capable of determining how that power should be divided, balanced, and restrained so that the three powers do not fuse unchecked, making failures less controllable, less reversible, and therefore more consequential.
The institutional problem precedes today’s foundation models.
When the automation becomes the decision
In 2023, the US Equal Employment Opportunity Commission announced the settlement of a case against iTutorGroup. According to the EEOC, the company had programmed its application software to automatically reject female applicants aged 55 or older and male applicants aged 60 or older. More than 200 qualified applicants were rejected. The companies agreed to pay $365,000 and provide other relief.
The technology was not GenAI. Indeed, testimony before the EEOC specifically noted that it was better understood as automated screening than AI. That makes the case more instructive, not less.
The institutional problem precedes today’s foundation models. A rule had been encoded. The system could apply it automatically. And the result of the application became the employment decision. The rulemaking, execution, and judgment had compressed into software. This is algorithmic power in its simplest corporate form.
The lesson is not that companies should simply put a human somewhere in the loop. A human who routinely approves an algorithmic recommendation may add labor without adding counterpower. Meaningful separation requires a point in the process at which someone has the information, independence, incentive, and authority to say: No. This rule is unacceptable. Or: No. This output should not determine the decision.
Trias Algorithmica extends that lesson into the following governance principle: that the authority to define such rules, operationalize them, and review their consequences should not automatically collapse into the same chain of command.
Banking model-risk practice illustrates the need for differentiation of decision
rights.
Businesses are already experimenting with separation
Some companies have begun constructing institutional distance deliberately.
Banking model-risk practice illustrates the need for differentiation of decision rights. The 2026 interagency guidance emphasizes objective challenge, clear responsibilities and conflicts of interest, validation and monitoring, and governance commensurate with model exposure and materiality. It also makes a distinction executives should notice: model risk rises not merely with the technical characteristics of a model, but with the significance of the decisions its output influences.
That is almost exactly the principle Trias Algorithmica requires. The stronger the power delegated to the algorithm, the stronger the institutional counterpower around it should become.
The same pattern can be seen outside the boundaries of a single enterprise.
Amazon’s online marketplace business occupies multiple positions in its ecosystem: platform operator, retailer, logistics provider, and rule-setter for third-party sellers. In 2022, the European Commission made Amazon commitments legally binding that addressed, among other things, its use of non-public marketplace-seller data, Buy Box treatment, Prime eligibility, and logistics. The arrangements included an independent monitoring trustee and a centralized complaint mechanism for sellers and carriers.
Whatever one thinks about the underlying competition-policy dispute, its organizational logic is revealing.
Data access was separated from certain competing uses. Rules governing access were constrained. Monitoring was assigned to a distinct actor. A complaint channel created a mechanism for contestation. The architecture of business itself was being redesigned around separations.
Trias Algorithmica gives these scattered business practices a coherent governing logic.
Eight separations of algorithmic power
Trias Algorithmica is not a proposal to create an algorithmic legislature, executive committee, and supreme court in every corporation. Nor does every spreadsheet, forecasting tool, recommendation engine, or copiloted email require elaborate governance. The principle should be proportional to the algorithmic power at stake.
The more materially an algorithm shapes the firm’s strategic options, the stronger the case for preventing any one actor from controlling the entire chain of algorithmic power involved.
That principle leads to eight separations.
1 – Separate corporate goal setting from machine making.
The people capable of building a system should not automatically possess unilateral authority to determine what the system is legitimately for.
Consider an AI system for employee productivity. Engineers can optimize it for output per hour. Operations can optimize it for cost. Human resources may care about retention and development. Legal may identify employment risks. Management may have commitments regarding safety or culture.
Which objective governs? That is not a machine-learning problem. It is a leadership decision. For consequential systems, objectives and risk tolerances should therefore be owned at an organizational level commensurate with their impact. In some cases, that means a business executive. In others, an enterprise risk committee, AI governance council, or the board itself. Algorithmic legislative power begins by asking what we are authorizing this machine to optimize.
2 – Separate data stewardship from developer control.
Corporate data are not simply fuel for AI. They embody rights, commitments, provenance, strategic value, confidentiality, customer expectations, employee information, intellectual property, and sometimes regulatory restrictions. The fact that data are technically available does not mean they are organizationally legitimate to use.
Who may repurpose customer service transcripts for training? May employee communications become optimization data? Can information acquired through one commercial relationship be used to compete in another? These questions should not automatically fall to the team seeking maximum model performance. Data stewardship should possess its own authority, because availability is not authorization.
3 – Separate optimization from the boundaries within which optimization is permitted.
Optimization is extraordinarily useful because it relentlessly pursues an objective. That is also its governance weakness. A sales algorithm can maximize conversion. A scheduling system can maximize utilization. A customer-retention model can minimize churn. An autonomous purchasing agent can minimize acquisition cost. None of these objectives knows when enough is enough.
 The corporation therefore needs an institutional distinction between the team that makes an objective function perform better and the authority that determines the limits within which that objective may be pursued.
The capacity to build something and the authority to release it into consequential operations are different powers.
4 – Separate deployment authorization from design ownership.
The capacity to build something and the authority to release it into consequential operations are different powers. Knight Capital, which lost $460m in just 45 minutes due to a software error, demonstrates the extreme version of why this distinction matters. The consequences of the error were magnified because deployment gave the system immediate access to market execution at scale.
For high-impact AI, companies should therefore create risk-proportionate deployment gates independent enough to challenge the product organization. That may mean pre-deployment testing by a separate technical team, legal sign-off, model-risk validation, cybersecurity review, a business-risk owner, a responsible-AI function, or executive authorization.
For the highest-risk systems, authorization should include explicit limits on what the system can do once released, whether through permission boundaries, staged deployment, rollback capability or, increasingly, restrictions on which tools an autonomous agent may access. The fact that a team can deploy a system should not mean that it alone decides whether the enterprise should.
5 – Separate post-deployment monitoring from product success.
A system’s product dashboard cannot be the sole definition of reality. Revenue can rise while complaints rise. Productivity can improve while employee turnover deteriorates. The organization therefore needs monitoring signals that are not subordinated to the system’s optimization objective.
The National Institute of Standards and Technology (NIST) recommends continuing monitoring, periodic review, incident response, appeal, and override processes, and differentiated responsibilities around AI risk. In corporate terms, that means a second set of eyes and, sometimes, a second set of metrics. The team rewarded for maximizing the system’s success should not possess exclusive authority to define whether the system is succeeding.
6 – Separate contestability from algorithmic finality.
This may be the most immediate principle for companies. An employee downgraded by a performance system, applicant screened out by software, customer blocked by fraud detection, supplier suspended by an automated system, or business unit denied capital because of a risk score should have some meaningful route to challenge the result when the stakes justify it.
Not every algorithmic output needs a tribunal. But a button marked “human review” is not meaningful separation if the reviewer lacks the evidence, time, independence, or authority to overturn the system. The iTutorGroup case illustrates how directly encoded rules can become final decisions. Contestability requires a credible point of institutional contradiction. The algorithm can propose. It should not always be allowed to dispose.
For a company, compute is only one dimension of scale.
7 – Separate the capacity to scale from the authority to commit resources at scale.
For a company, compute is only one dimension of scale. The others are capital, transaction authority, customer reach, production capacity, access privileges, and organizational dependency.
As AI agents become capable of executing payments, buying advertising, placing orders, manipulating databases, provisioning infrastructure, negotiating with counterparties, or calling other software agents, this distinction becomes more important. A system capable of spending $100 is not institutionally equivalent to the same system capable of spending $100m.
Scale should be treated as delegated authority. Boards already understand limits on capital expenditure, credit exposure, trading authority, acquisition approval, and signing authority. Algorithmic systems need equivalents.
8 – Separate technological dependence from the practical inability to exit.
Algorithmic power does not reside only in models. It resides in the broader technological and organizational stack that surrounds them, including the infrastructure, data, systems, and processes on which they depend.
A company that cannot realistically replace a supplier because doing so would require rebuilding its models, moving enormous datasets, reconstructing workflows, retraining thousands of employees, and losing accumulated context has less bargaining power than its procurement contract may suggest. It has effectively surrendered a form of counterpower. Portability, interoperability, multi-provider architectures, data-exit rights, and credible fallback procedures are therefore not merely IT considerations. They are governance mechanisms. The ability to exit is one of the oldest forms of counterpower in commerce. Exit that exists only on paper is not much of a check.
Taken together, these eight separations do not prescribe a finished constitution for the algorithmic corporation. They establish a simpler governing principle: certain algorithmic powers should not be allowed to accumulate in the hands of a single actor. For a board, that may be a more productive starting point than determining whether the company has an AI ethics policy.
An AI product can increasingly contain policy, execution, surveillance, evaluation, and enforcement in the same technical architecture.
The case against an algorithmic constitution for business
The first objection is obvious. Companies already have governance. But AI is recombining powers that corporate governance has learned to separate elsewhere. An AI product can increasingly contain policy, execution, surveillance, evaluation, and enforcement in the same technical architecture. The traditional organization chart shows those functions as separate while software quietly reconnects them. That is the new problem.
A second objection is that separation will slow innovation. It probably will in some cases. But friction is not always failure. Financial controls slow transactions. Security review slows software releases. We accept these costs because speed is not the only legitimate corporate objective.
A more serious danger is bureaucratization. A company can create so many AI committees that responsibility disappears among them. Governance can become a ritual in which everyone signs a form and no one possesses enough authority to stop anything.
That is why Trias Algorithmica should be understood as functional separation, not committee proliferation. Another objection is that responsible-AI teams, model-risk functions, internal auditors, and boards already provide that challenge. Sometimes they do. Microsoft’s published governance structure and more than a decade of banking model-risk practice demonstrate that meaningful internal differentiation is possible.
But independence in name is not counterpower. What matters is not whether another function exists on the organization chart. It is whether that function can meaningfully say: No.
The boardroom assessment
Boards should resist the temptation to become AI engineering committees. Their responsibility is more fundamental: to understand where algorithmic power sits inside the enterprise and how it is constrained.
A boardroom assessment of algorithmic power should focus on three dimensions:Â authority, allocation, and counterpower.
Authority concerns the power delegated to the system. The board should understand which systems materially shape the treatment of employees, customers, suppliers, or counterparties; which systems can act rather than merely recommend; and which can commit capital, communicate policy, change prices, allocate work, approve or reject people, or interact with other enterprise systems.
Allocation concerns where decision rights sit across the algorithmic chain. It requires identifying who defines a system’s objectives, who determines the boundaries within which it may operate, who authorizes its deployment, who evaluates its performance, and who decides whether its outputs should become operative decisions.
Counterpower concerns the capacity for independent challenge. It requires determining who monitors what the system’s builders do not measure, who can reconsider or override its decisions, who can restrict its authority, and who ultimately possesses the power to stop it.
The purpose of this assessment is to reveal where algorithmic powers have become concentrated, particularly where the people who build a system are also those who define its purpose, authorize its deployment, evaluate its performance, interpret its failures, and determine whether it should continue to operate.
That is where algorithmic governance becomes a question of institutional architecture rather than technical oversight.
The emerging governance literature is already moving in this direction. NIST explicitly assigns responsibility for AI risk to executive leadership and describes potential roles for boards, management, audit, product, development, testing, procurement, and oversight functions.
Trias Algorithmica takes the next conceptual step. It says those are not merely roles around AI. They are counterpowers to algorithmic power. Once a board sees them that way, a different kind of governance conversation becomes possible.
For centuries, constitutional government has operated from a durable intuition: power is safest when it encounters power organized against it.
Business has repeatedly discovered the same principle in accounting, audit, financial risk, cybersecurity, compliance and corporate governance. AI does not make that principle obsolete. It makes it newly relevant to the operating architecture of the firm.
There will be no board meeting at which a corporation formally decides to surrender managerial power to algorithms. Algorithmic concentration will arrive much more quietly. Through useful copilots. Automated approvals. Better recommendations. Faster recruiting. Dynamic prices. Lower service costs. Systems that initially advise managers and gradually become the infrastructure through which management itself is exercised.
Trias Algorithmica begins from the proposition that companies should separate algorithmic powers before efficiency teaches them to stop noticing that they were powers at all.
Authors
Related
How should CEOs respond to the era of hyper-vigilance
CEOs face greater scrutiny than ever. Retreating from the speaking spotlight is no solution. IMD’s Jean-François Manzoni explains how CEOs should talk with trust and integrity....
Should your kids leave the family business before they lead?
Should kids leave the family business to lead later? Learn how outside-entrepreneurship strengthens leadership, ownership, and long-term renewal....
Honestly speaking? Your board is probably too vanilla
Modern corporate governance is built for a stable world that no longer exists. If your board members look diverse on paper but always agree in practice, then it’s time for a rethink,...
Audio available
Learn Brain Circuits
Join us for daily exercises focusing on issues from team building to developing an actionable sustainability plan to personal development. Go on - they only take five minutes.
Â
Explore Leadership
What makes a great leader? Do you need charisma? How do you inspire your team? Our experts offer actionable insights through first-person narratives, behind-the-scenes interviews and The Help Desk.
Â
Join Membership
Log in here to join in the conversation with the I by IMD community. Your subscription grants you access to the quarterly magazine plus daily articles, videos, podcasts and learning exercises.
Â