Trias Algorithmica: How Leaders Can Make Algorithmic Power Legitimate

by Hamilton Mann

As algorithms move from advising organizations to acting through them, leaders need to separate algorithmic powers. No single system should be able to concentrate the power to set the rules, execute them, and judge its own results. Independent counterpower capable of constraining the authority delegated to algorithmic systems is a condition of the legitimate exercise of algorithmic power.

Algorithms Acquiring Power Are Not New

In 2019, Deanna Amato discovered that the Australian government believed she owed it thousands of dollars when it seized her tax refund.

The debt had been generated through what became known as Robodebt, an automated welfare-compliance program that compared tax-office income data with social-security records. Rather than establishing what recipients had actually earned during individual two-week benefit periods, the system could average annual income across those periods and use that figure to calculate an alleged overpayment.

For Amato, the assumption was wrong. The resulting debt was nevertheless treated as real. The Federal Court declared the demand for payment invalid.

Once software can encode rules, operationalize them at scale, and produce conclusions that institutions treat as authoritative, the relevant question is no longer simply whether the algorithm is accurate. It is how much power has accumulated around it and who possesses the power to contradict it.

What’s New Is Algorithmic Power Becoming Normalized Authority

Fast-forward. In May, Senbin Hou walked into an Immigration and Customs Enforcement office near Seattle for a scheduled check-in. He left in custody.

Hou, a Chinese national whose immigration proceedings were still pending, had previously been released and enrolled in ICE’s Alternatives to Detention program. Part of that program required him to use a mobile application to complete biometric check-ins through facial recognition.

According to records maintained by ICE’s contractor, Hou had failed to complete 15 of those check-ins. Hou told a different story. He said that technical problems with the application’s facial-recognition function repeatedly prevented him from completing them, and that each time this happened, he contacted ICE to report the problem.

The system recorded noncompliance. Hou described technology failure. ICE acted on the former.

After what it described as a routine review of his record, the agency revoked Hou’s release and arrested him when he appeared voluntarily for an in-person appointment. For weeks, a database entry had become a deprivation of liberty.

In July 2026, a federal judge ordered Hou’s immediate release. The court found that his detention violated due process and emphasized the substantial risk of erroneous deprivation when the government could act on recorded biometric failures without first allowing the person affected to contest them.

Both Robodebt and the system governing Hou were algorithmic systems, but they represented very different generations of algorithmic technology. Yet while the technological sophistication had changed considerably, the governance structure around how algorithmic authority was exercised had not.

In both cases, a rule defined what counted as compliance; an algorithmic system translated messy real-world circumstances into a standardized administrative determination; that determination could trigger state action; and the individual subject to it lacked a meaningful opportunity to challenge the system-generated premise before serious consequences followed.

In effect, this concentrates algorithmic power across the chain from rule-making to execution and judgment.

Algorithmic Power Is Scaling Faster Than Counterpower

If we think this concentration of algorithmic power can have life-altering consequences only in the public sector, we are wrong. The private sector may demand even greater attention.

A 2026 Grant Thornton survey of 950 senior executives from midmarket and large organizations across ten industries, including banking, construction and real estate, technology and telecommunications, insurance, and energy, all sectors in which algorithmic power can materially affect people’s lives, found that 74% of organizations were already giving agentic AI access to their data and business processes. Five percent of organizations were already allowing AI agents to execute high-stakes decisions without human review. Yet 78% of senior executives lacked strong confidence that their organizations could pass an independent AI-governance audit within 90 days, and only 20% had a tested AI incident-response plan for an AI failure.

What those numbers imply when combined is striking. If 74% of organizations are giving AI agents access to data and business processes, while 80% lack a tested incident-response plan, then even under the most favorable possible distribution, and allowing for rounding in the reported percentages, at least about 53% of respondents must fall into both categories, potentially rising to about 74%. In a sample of 950 senior executives, that represents roughly 504 to 703 respondents whose organizations give AI agents access to data and business processes while lacking a tested failure-response plan.

Apply the same arithmetic to governance readiness: with 74% granting agentic access and 78% lacking strong confidence that they could pass an independent AI-governance audit, and similarly taking the reported percentages conservatively, at least about 51% of respondents must again fall into both categories, potentially rising to about 74%. In a sample of 950 senior executives, that represents roughly 486 to 703 respondents. These are not correlations inferred from the data, but the range of overlap mathematically implied by the reported percentages.

And the 5% permitting AI agents to make high-stakes decisions autonomously should not be dismissed as a small number. Applied to a sample of 950 senior executives, that represents roughly 48 respondents whose organizations allow algorithmic decisions to acquire a significant degree of operational finality. The greater that finality, the greater the risk that algorithmic power becomes difficult, costly, or even impossible to reverse. In the high-stakes contexts represented across these industries, that can mean decisions affecting whether someone receives credit, gains access to housing, retains access to a digital or communications service, qualifies for insurance, or receives or can afford energy.

The picture is clear.

The governance gap lies in the mismatch between the power organizations delegate to algorithmic systems and the governance infrastructure available to constrain that power.

The direction of travel is clear.

The central challenge is to govern the power acquired by algorithmic outputs, which should be inseparable from ensuring that meaningful counterpower exists before those outputs become power exercised over people’s lives.

The Powers of Algorithm We Mistake for Technical Functions

Every consequential algorithmic system sits within three functional forms of power.

The first is algorithmic legislative power: the ability to encode the rules. A threshold, objective function, ranking criterion, or system prompt may look technical. But these choices determine how attention, suspicion, opportunity, workload, capital, and risk are distributed. A corporate rule does not cease to be a rule because it is expressed as a parameter.

The second is algorithmic executive power: the authority to translate rules into action. Whether an algorithm recommends an intervention or is permitted to execute it directly by freezing an account, changing a price, advancing or rejecting a candidate, committing a payment, or triggering another operational process, it is the operational arm of algorithmic power. The critical resource is not computing power. It is permission.

The third is algorithmic judicial power: the authority surrounding the determination that an output, action, or person satisfies the relevant rules. It determines whether a fraud alert was justified, reviews rejected applicants, decides whether a model failure warrants suspension, hears appeals, and ultimately determines whether a prior decision is genuinely reconsidered or rubber-stamped.

Fused in algorithmic systems, these three powers can produce absolute authority over both people and the organizations that deploy them.

At a time where we are delegating even greater authority to algorithmic systems whose capabilities are becoming more sophisticated at a pace measured in weeks, extending the reach of concentrated algorithmic power into an ever-wider range of decisions, governing that power is not an option.

If history has taught us anything, it is that the separation of powers and the establishment of countervailing power are instrumental in preserving rights of appeal, protecting human agency, and ensuring the legitimate exercise of power.

Oversight Should Not Be Confused With Algorithmic Counterpower

Corporate governance already understands the principle of counterpower through separation of power.

Traders do not unilaterally set their own risk limits. Companies do not prepare their financial statements and independently audit them. Credit officers operate within authorities established elsewhere.

Yet, the algorithmic concentration of power often takes root in organizational concentration of power.

This occurs, for example, when a technical team defines an algorithmic system’s purpose, builds it, deploys it, evaluates its performance, interprets its failures, and decides whether its outputs should stand; in effect, cumulating functions analogous to policy-setting, risk management, safety, compliance, validation, incident review, and appeal can accumulate within the same chain of command.

Governance failures do not remain outside the technology; they compound within the algorithmic system itself. The organization concentrates authority over the system, while the system concentrates the algorithmic functions inadvertently delegated to it, through which its power is ultimately exercised.

Separation of algorithmic powers is therefore not simply a matter of inserting a human somewhere in the process or assigning AI oversight to another committee. It requires identifying where power has accumulated, organizationally and technically, ensuring that no actor controls the entire chain from rule-making to execution to judgment, while also subjecting each exercise of authority within that chain to productive friction and independent counterpower.

That means separating goal-setting from model building; data stewardship from model optimization; optimization from the limits within which optimization may occur; system design from deployment authorization; product success from independent monitoring; automated decisions from meaningful appeal; and technical capacity to scale from authority to create enterprise-wide exposure.

And the greater the power delegated to an algorithm, the stronger the counterpower organized around it must be.

The Leader Test For Governing Algorithmic Power

Boards and executive meetings do not need to become AI engineering committees. But they need to acknowledge that algorithmic systems do not simply enter the technology stack; they enter the organization’s authority stack. Leaders must therefore keep sight of how power is redistributed, including the authority granted to the algorithmic systems they hire. They must build their map on how algorithmic power reconfigures authority across the organization.
Three questions help provide one.

• What power is actually being delegated to the algorithmic system, and how far does that authority extend within the governance of the firm or institution?
• How is that power distributed around the system, and what prevents rule-setting, execution, and judgment from collapsing into the same hands?
• What independent counterpower can challenge, constrain, or stop algorithmic power when it exceeds its mandate or causes, or risks causing, harm?

That is the premise of Trias Algorithmica.

Ultimately, governing algorithmic power means acknowledging the social necessity of continually revisiting a simple question: what makes the power it exercises legitimate?