Main

BTU researchers’ analytical assessment is that successful AI performance shows how well a chosen objective is achieved, but cannot determine whose interests may legitimately bear the cost. Georgian businesses therefore need separate authorization outside the model: making an agent more capable does not itself expand its mandate.

Key Takeaways

• Better AI performance evaluates execution of a chosen task; passing a test does not give a system authority to make decisions affecting other people’s interests.
• A task, external document or another agent’s request must be assessed against the current mandate and cannot independently enlarge it.
• BTU researchers assess that economic value for Georgian businesses is supported by arrangements in which better analysis accelerates authorized work while keeping responsibility understandable.

AI agents offer an economic opportunity to help people perform work as well as process information. A system can find alternatives, compare terms, prepare a message and perform a defined action through connected software. These capabilities can reduce time spent searching for information and working across poorly coordinated systems. Yet this raises a question that evaluating model quality alone cannot answer: which of the possible actions has the agent actually been entrusted to perform?

The distinction does not require holding back technological development. As a system handles a wider range of tasks, separating its ability to perform from its authority to act becomes more important. A weak program may simply fail to complete a step. A stronger agent may complete it well while still lacking the organization’s permission. A good result and an authorized result mean different things for a business.

Good performance cannot settle competing interests

A successful test relies on a criterion selected in advance. If the objective is response speed, the test shows how quickly the system works under the conditions examined. Speed cannot tell us which customer deserves priority, which information may be used or which obligation the company should accept. Those questions concern the acceptability of the objective and choices among different interests, not simply the quality of execution.

Consider an illustrative university agent that rearranges teaching-room schedules to improve room use. Some groups receive a better time while others lose a previously agreed slot. Better utilization does not establish whose agreement may be changed. This is neither a quantitative model nor an observed Georgian case. It explains why better performance cannot itself confer authority. The institution must first establish which commitments and restrictions remain fixed during optimization.

This is BTU’s principal addition to the familiar distinction between capability and mandate. Evidence of competence and the basis of authority answer different questions even when an agent makes no mistake. The problem is not confined to the probability of error. An entirely accurate action can still exceed the work entrusted to the system. A further improvement in accuracy therefore cannot resolve the authorization question.

Three different questions

The table is BTU’s qualitative comparison, not a ranking of systems. It combines agent-management research materials with public access-control guidance.

Capability concerns how well a system can perform a particular task: whether it can compare offers correctly, for example, and identify important differences. Appropriate tests and examples resembling the intended work should establish this ability. Persuasive writing is insufficient evidence. Even a successful test tells us about the task and conditions actually examined, rather than everything the agent might subsequently attempt.

Technical access describes which application the system can enter and which functions it can invoke. A connected tool may provide reading, editing and sending together. That technical arrangement does not necessarily reflect the organization’s intended decision. An excessively broad connection can make more actions available than the business deliberately delegated.

Authority establishes which action may be performed on a person’s or organization’s behalf and under what conditions. Here, a mandate means the boundaries of that delegated authority. These can include the recipient, purpose of data use, time or type of action. Permission does not arise simply because a step looks useful. An authorized decision establishes it, and the execution environment must uphold its limits.

A task cannot create its own authority

A company may seek better service or more effective resource use. That assignment establishes a direction without changing existing commitments or the means permitted. A mandate can be expanded through an authorized decision consistent with applicable rules and reflected in the execution environment. The agent’s reasoning that an extra step would be useful cannot substitute for that decision. Nor does an email, web page or another agent’s request create permission merely through its contents.

BTU’s agent-management materials distinguish preparing work, executing after human approval and operating within previously defined limits. They treat an agent’s unilateral expansion of its own authority as unacceptable. This is management guidance, not evidence that all Georgian organizations already follow the same arrangement. [6]

Earlier permission is not proof for a different action

Authority is not merely a permanent characteristic of an agent. It describes a relationship between the agent, a particular action and its circumstances. Knowing that the system was approved once, or that its connected account still works, is insufficient. Permission to analyze information does not authorize using it for another purpose. Sending the same output to a different recipient may also fall outside the original mandate. The agent’s skill can remain unchanged while the basis for action needs checking again.

In the university illustration, assessing room use and publishing an approved timetable for students create different consequences. In the first activity, the institution compares alternatives. In the second, people plan their time around the information received. A positive evaluation of the analysis cannot establish permission to publish. Earlier authorization is relevant only if it actually covers the proposed action and its conditions; otherwise, a separate authorization decision is needed.

Two different checks are therefore necessary: whether the prepared result is acceptable and whether this particular use of it is allowed. Passing one is insufficient. An accurate but unapproved change and a permitted but poor-quality change are different problems. Combining them into one general score leaves the company with “works well” while obscuring the basis for a specific action. NIST’s summary of comments released on 29 September 2026 likewise identifies task-scoped and contextual authorization. It summarizes stakeholder feedback rather than introducing a new mandatory standard. [2]

Why authorization belongs outside the model

A model may explain its own boundaries correctly. The organization nevertheless cannot rely solely on the judgment of the same model that proposed the plan. If an application is to change something, authorization also needs checking in that application or an associated control. The agent’s statement that it can perform an action cannot be the executing system’s only basis for permitting it.

OWASP’s 2025 guidance calls for authorization checks in downstream systems rather than reliance on a language model’s judgment alone. [1] For a business, the issue remains relevant even when the proposed plan is accurate and useful: the plan must pass the controls of the environment in which the real change would occur.

Knowing which agent is acting does not mean accepting every request from it. The system may identify an agent correctly while that agent lacks permission to send information to a particular recipient. NIST’s zero-trust architecture similarly treats authentication and authorization as distinct functions. [3] In technical terms, this is access control. In business operations, it protects the boundary of a delegated decision.

The distinction becomes particularly important when a model is upgraded. A better version may complete more tasks or discover a new route to the objective. The organization’s goals and permitted actions do not automatically change. If quality improvements automatically expand permissions, a technology upgrade can quietly redistribute decisions inside the company. BTU’s assessment is that managing capability and mandate separately reduces this hidden change. Guidance published by Australia’s cyber security centre on 11 September 2026 also distinguishes proposed model actions from the software environment that executes them and enforces controls. [5]

The economic significance for Georgia

Geostat’s release of 29 May 2026 reports that 15.3 percent of Georgian enterprises used a web page or website in 2025. [4] This is not a measure of readiness for AI agents. It describes a particular digital channel and reminds us that discussion of technological use cannot be limited to firms with the most developed systems. Businesses organize customer communication and internal work differently.

Purchasing a powerful agent therefore does not answer what a Georgian company should entrust to it. Businesses can use better search, multilingual drafting and comparison of offers while a person continues to decide whether to approve an order or change public terms. Such a division does not waste the technology. It lets a company improve the preparation of complex work without transferring all associated decisions at once.

The distinction becomes especially visible to customers when a conversation creates a commitment. Explaining available times and changing a previously agreed service time are different actions. Tourism, professional services and education share the same mechanism: an improved schedule can affect different people’s interests differently. These sectoral illustrations explain the mechanism; they do not measure how widely agents have been deployed.

Economic value cannot be judged from an agent’s speed alone. If every rapidly prepared result subsequently requires clarification of an uncertain agreement, work has shifted to another stage. Clear authority reduces uncertainty over whether an output is advice, a draft or an approved company action. Partners can also understand which information they may rely on. That clarity matters for trust, coordination and unnecessary checking, although this article does not measure its monetary effect.

Using strong agents requires clear boundaries

BTU researchers assess that separating capability from authority avoids a false choice between doing everything manually and handing everything to software. An agent can possess broad knowledge while holding narrow authority to act. That arrangement enables sophisticated analysis where the final choice must remain with the organization’s responsible party.

The answer to the title therefore extends beyond preventing mistakes. AI capability determines which work could be performed; authority determines which consequences the organization has entrusted to the system in a particular setting. Better performance can support reconsidering a mandate, but cannot substitute for expanding it. Georgia’s productive opportunity is to use agents whose abilities have genuinely been assessed and whose actions remain within authority actually delegated.

Questions and answers

If an agent performs better than a person, should it receive more authority? Better performance supports considering broader use. Permission for additional actions must still be established separately, with appropriate responsibility and conditions.

Is a request from a particular person always enough? The person must hold the relevant authority. An instruction cannot cancel the organization’s existing restrictions unless that person is authorized to make such a change.

Does this mean approving every step manually? No. Predefined actions can run automatically within their authorized boundaries. The important condition is that the agent’s assumptions do not change those boundaries.

What does a successful test establish? It establishes observed performance under tested conditions. Permission, access and performance in another environment remain separate questions.