Capability Is Not Authority

Share
Capability Is Not Authority

Why AI governance must move closer to real-world action

Excerpt

As AI systems move from chat interfaces into enterprise workflows, infrastructure, and physical machines, intelligence is no longer the only variable that matters. The more urgent question is what the surrounding system has permitted AI to reach, modify, and control.


The AI Journal recently published a feature on SHIRO & Co.’s Authority Surface framework. The article captured the central distinction clearly: the risks created by an AI system do not arise from capability alone. They also depend on the authority that the surrounding technical and institutional environment gives that capability.

Read the feature in The AI Journal

This distinction becomes increasingly important as AI leaves the chat window.

A model can draft a message without being able to send it. It can recommend a payment without being able to authorize one. It can propose a change to production infrastructure without possessing the credentials required to deploy it. It can calculate a path for a robotic arm without being permitted to move the machine.

In each case, the model may have the same underlying capability. What changes is the surrounding authority.

That is why the question facing AI governance is no longer only:

What can the model do?

It is also:

What has the system permitted the model to do in the real world?

The missing layer between intelligence and consequence

Much of the discussion around AI safety begins with the model. We compare reasoning performance, autonomy, planning horizons, context windows, benchmark scores, and the ability to use tools.

These measures matter, but they do not tell us whether a model can produce a real-world consequence.

A highly capable model isolated from external systems has a limited operational footprint. Connect the same model to identity credentials, customer records, communication channels, payment systems, cloud infrastructure, operational technology, or physical machinery, and its position changes.

The model has not merely become more intelligent. It has entered a larger field of authority.

SHIRO & Co. calls this field the Authority Surface: the set of systems, permissions, and operational resources through which an AI system can create effects beyond its own computational environment.

The Authority Surface can be examined across six dimensions:

  1. Access — What systems, data, tools, or machines can the AI reach?
  2. Permission — What actions is it authorized to take within them?
  3. Persistence — Can it maintain access or effects over time?
  4. Propagation — Can one action trigger other systems, agents, or environments?
  5. Reversibility — Can the action be stopped, corrected, or rolled back?
  6. Observability — Can people see what happened and reconstruct why?

None of these dimensions is inherently dangerous on its own. The risk emerges from their combination.

Read access to a database is different from permission to modify it. Drafting an email is different from sending it. A reversible configuration change is different from a financial transfer. A visible action performed once is different from a persistent process that can propagate through connected systems.

The Authority Surface therefore does not attempt to classify an AI model as simply safe or unsafe. It examines the operational territory the model has been allowed to occupy.

A decision is not yet an action

Once capability and authority are separated, another boundary becomes visible.

AI systems continuously generate classifications, recommendations, forecasts, and proposed actions. But a recommendation should not automatically acquire permission to act. A decision produced by a model is not the same thing as organizational authorization.

This is where the Supervisory Layer sits: between inference and consequence.

The relationship can be expressed as:

Capability → Authority Surface → Supervisory Layer → ALLOW / HOLD / DENY → Real-World Action

The Supervisory Layer does not ask only whether an output appears plausible. It asks whether the proposed action is permitted under the current conditions and evidence.

  • ALLOW means that the defined conditions are satisfied and the action may proceed within its stated boundary.
  • HOLD means that required evidence, confirmation, or human judgment is still missing.
  • DENY means that the action exceeds its authority or violates a defined boundary.

HOLD may be the most important state.

Many automated systems are designed around a binary choice: proceed or reject. But real organizational decisions often contain uncertainty that cannot responsibly be collapsed into either answer. A missing approval, an unverified customer statement, an unreadable label, stale equipment data, or the presence of a person in a work area may not justify permanent rejection. They justify stopping until the missing condition is resolved.

HOLD gives uncertainty an operational form.

It prevents a model’s confidence from being mistaken for evidence and prevents speed from silently becoming authority.

From enterprise software to physical action

The distinction becomes more consequential as AI moves through widening operational roles:

Chatbot → Tool User → Agent → System Operator → Infrastructure Participant → Physical Actor

At the beginning of this progression, an error may produce an inaccurate sentence. Further along, it may change a customer record, deploy code, allocate money, interrupt a service, redirect a vehicle, or move a machine.

The transition into physical systems does not make the earlier governance model obsolete. It makes it more concrete.

In an industrial environment, an AI system may propose the next handling or sorting action. But the permission to proceed cannot depend only on the model’s proposed decision. It may also depend on whether the item can be identified, whether equipment status is current, whether communications are available, whether the action falls within the system’s authority, and whether a person is present in the affected area.

An ALLOW result in a simulation is not permission to operate real equipment. It means only that the case may advance to the next defined stage of evaluation. Physical execution requires its own controls, certifications, interlocks, responsible operators, and site-specific authorization.

This is the purpose of the emerging Physical AI Execution Layer: not to make AI act more freely, but to make the boundary between a proposed action and physical execution explicit.

Governance should follow the path of consequence

AI governance is often expressed as principles, policies, or model-level constraints. These remain necessary, but they can sit too far from the moment when an inference becomes an action.

The closer an AI system moves toward real-world execution, the more governance must become operational.

It must identify:

  • what information is treated as fact;
  • what remains inference;
  • which permissions exist;
  • which conditions are still unverified;
  • who can review or override a decision;
  • whether the action can be reversed;
  • and whether the history can be reconstructed afterward.

This is not only a technical problem. Authority is distributed across software, credentials, workflow design, organizational policy, human roles, contracts, and physical infrastructure. A model may generate the proposal, but the surrounding institution determines whether that proposal can enter the world.

For that reason, responsibility cannot be assigned to the model alone. The architecture of permission is itself a form of decision-making.

The next question for AI systems

AI capability will continue to expand. Organizations may not be able to control the overall pace of that expansion, but they can still control how capability is connected to authority.

They can reduce unnecessary access. They can separate observation from execution. They can preserve provenance. They can introduce HOLD states. They can require human review at boundaries where consequences become difficult to reverse. They can design systems so that authority does not expand invisibly each time a new tool, credential, API, or machine is connected.

The task is not to prevent capable systems from being useful. It is to prevent growing capability from automatically becoming growing authority.

The decisive question is therefore not whether an AI system is intelligent enough to act.

It is whether the surrounding system has earned the right to let it.


Further reading

Read more

健康なのに、会社に行きたくない。

健康なのに、会社に行きたくない。

Noriko Interviews Kosuke Session 01 / Episode 06 — LIFE / Body 月曜日、午前7時12分。 Norikoはベッドの上で、天井を見ていた。 目覚ましは、もう止めてある。 睡眠時間は6時間42分。 スマートウォッチによれば、 睡眠の質は「良好」。 安静時心拍数も正常。 体温も問題なし。 昨日はお酒も飲んでいない。 熱もない。 頭痛もない。 お腹も痛くない。 それでも、 Norikoは起き上がれなかった。 Noriko: 今日、会社行きたくなかったんですよ。 Kosuke: 今日だけ? Noriko: それ、嫌な聞き方ですね。 Kosuke: (笑) Noriko: いや。 今日だけじゃないです。 たまにあります。 Norikoは少し考えた。 Noriko: でも別に、会社嫌いじゃないんですよ。 Kosuke: うん。 Noriko: 嫌な上司がいるわけでもないし。 いじめられてるわけでもない。

By Kosuke Shirako