Main

BTU researchers’ assessment is that an AI agent’s boundaries become stronger when accountability matches decision authority: the human process owner must be able to influence the conditions for which they are answerable. For Georgian businesses, choosing that manager and giving them meaningful authority matters as much as selecting the technology supplier.

Key Takeaways

• BTU’s analytical assessment is that agent-enabled work needs a named human owner whose accountability is supported by relevant decision authority and working resources.
• One process owner does not remove other participants’ responsibilities: leadership, professional specialists and technical teams remain answerable for their own decisions.
• Buying an agent cannot replace a Georgian company’s agreement about who sets the service objective and who resolves competing departmental requirements.

A company assigns an agent to customer inquiries. The service manager wants prompt answers, the sales team wants more orders, and finance wants lower costs. Each objective makes sense, but they do not always coincide. Which should guide the agent when spending longer with a customer protects service quality without increasing sales?

This is an illustration, not an account of a particular Georgian business. It reveals a problem that precedes model accuracy: the organization must make its own choice. Software cannot legitimately resolve an unsettled disagreement about the purpose of a process on the organization’s behalf.

Here, an AI agent means a system that chooses steps and acts in a digital environment to complete an assigned task. Its mandate is the substance and scope of the work entrusted to it. The question is who has authority to determine that mandate and how the company remains answerable for it.

Who owns the whole process outcome

BTU’s agent-management research distinguishes a business-process owner, a technical owner and the person approving a decision. Within this framework, using an agent does not replace human accountability: responsibility connects to a particular role and decision. These materials provide a conceptual management foundation rather than measurements of practice already implemented in Georgia. [1]

A process owner is not necessarily the software’s proprietor or a shareholder. The term identifies the person entrusted with the purpose, outcome and development of an activity. A customer-service owner, for example, should be able to explain what the company means by good service and why a particular use of an agent serves that objective.

Putting a name on a document does not establish this role. If the manager must answer for outcomes but cannot change priorities, involve a necessary specialist or raise a problem with leadership, the accountability is nominal. The duty has been assigned without the means to fulfil it. Employees and customers can both suffer from that arrangement.

Canada’s federal guide to agentic AI, on a page dated 15 September 2026, calls for designated, trained accountable owners, including where agents delegate to other agents. [2] It addresses Canadian public administration, but its organizational question also matters to Georgian businesses: where does responsibility sit when work is divided among several performers?

BTU researchers’ assessment is that the answer should remain understandable at the level of the whole process. Customers should not need to study software components to discover who can consider their issue. Clear ownership matters in ordinary service, not only after a malfunction.

Who allocates decision authority

An agent’s mandate starts with people who already hold authority over the relevant activity. Leadership determines the overall objective, acceptable risk boundaries and allocation of responsibility. The process owner connects those choices to the conditions of the particular service. The chain of authority should therefore not begin with whichever employee first opened an account for the new software.

The US National Institute of Standards and Technology’s playbook for its voluntary 2023 AI Risk Management Framework connects clear roles, staff preparation and executive responsibility for risk decisions. [3] Appointing an implementer is insufficient; leadership remains answerable for its own choices.

This does not require a director to consider every routine action. Excessive centralization would delay work and overload senior management. Leadership should establish which decisions belong to the process owner and which exceed that role’s mandate. Employees retain room to act while the organization retains a common direction.

Agreement matters especially where departments meet. A sales-process owner may organize customer interactions without being entitled to discard another department’s professional requirements. When objectives conflict, an identified person or body must have authority to make the choice and explain why. A shared meeting helps only if authorship of the final decision remains clear.

Leadership also operates within legal and organizational constraints. Naming an agent owner changes neither the law nor existing professional duties. Internal management accountability and the determination of legal liability in a particular case should therefore remain distinct.

One owner does not do everyone’s work

A process needs someone concerned with its overall outcome. That appointment must not absolve other participants. A domain specialist evaluates professional substance. The technical team explains system behavior and limitations. Relevant risk specialists assess matters within their expertise. The process owner connects this knowledge to the business purpose rather than pretending to be an expert in every field.

The OECD AI Principles relate accountability to participants’ roles and the context of use. [4] This helps frame two organizational extremes: everyone discusses only their small component while nobody owns the result; or everything is assigned to one manager while other participants’ actual influence goes unrecognized.

In the first arrangement, decisions disappear between separate functions. In the second, someone is made answerable after a poor outcome despite having been unable to determine the relevant conditions or resources beforehand. BTU’s assessment favors ownership of the integrated process alongside differentiated participant responsibilities. This is a management proposition, not a universal legal formula.

The arrangement also matters to employees. Professional disagreement should not automatically be treated as resistance to automation. If a service specialist sees that the agent’s objective misses the customer’s real need, that information needs a route to the decision-maker. Otherwise an owner exists on paper but cannot learn from the work.

What the company buys from a supplier

Buying a ready-made agent adds an external participant. The supplier may possess important knowledge about the model, application or technical service. Yet a standard product description cannot settle the company’s customer promises, professional objectives or internal priorities. The purchasing organization must connect those choices to its own activity.

A clear division of responsibilities is therefore necessary. Suppliers should explain which part of the service they provide and what information they can make available. Buyers should know who will use that information to decide. A contract can allocate obligations, but cannot by itself resolve an unanswered internal question: why are we using an agent for this work?

The issue becomes clearer when a system delegates to additional agents. More performers should not mean losing the original business objective. The human owner needs to understand the overall service, regardless of how many software participants handle its components. This requires relationships through which the owner can obtain justified answers, rather than knowledge of every technical detail.

Why this matters in Georgia

For Georgia, the issue connects several different activities. In retail, education, tourism or professional services, agent-enabled work may concern a customer request, an employee document or a service condition. These different uses involve choices made in an organization’s name. Allocating responsibility consequently matters beyond the technology sector.

Article 3 of Georgia’s Law on Personal Data Protection distinguishes the controller who determines processing purposes and means, including when processing is carried out through a processor. [5] This is a specific data-protection role, not a universal definition of an agent owner. Technical execution and responsibility for purpose need to be distinguished; the legal allocation in a particular arrangement requires separate assessment.

In a small firm, one person may hold several management roles. That does not automatically require new departments. The individual and colleagues should understand when the person is setting the business objective, when professional advice is needed, and which matters require another authorized participant. A simple structure can work well; unclear responsibility is not the same as simplicity.

Ownership must survive change

Defining the mandate is not only a launch-day decision. An agent’s work may expand, service conditions may change, or the responsible manager may move to another role. Without a corresponding update to responsibility, an earlier agreement may cease to describe actual practice.

Who can change the mandate, and who takes over ownership, therefore matter. Giving a new manager access to the application is insufficient. They need to understand the reasoning behind earlier choices, the intended outcome and the matters over which other people retain authority. That knowledge supports organizational memory without depending on one employee’s permanent presence.

Ultimately, an agent’s boundaries are set by authorized human governance connecting the purpose of the process, its accountable owner and participants’ duties. BTU’s central conclusion is that delegating work to technology requires clarity in the organization’s own decisions. When the company knows who grants the mandate, who can change it and who explains the outcome, agent autonomy has an intelligible human foundation.

Questions and answers

Must the agent owner be a programmer

Not necessarily. The owner needs to understand the work and hold relevant management authority, with a functioning relationship to accountable specialists who provide technical expertise.

Can the supplier be the only responsible party

Suppliers have their own duties, but cannot fully replace the company’s responsibility for its purpose and deployment decision. Legal allocation depends on the relationship and applicable rules.

Does one owner mean concentrating every decision in one person

No. Ownership of the integrated process should coexist with clear responsibilities for specialists and decision approvers. The purpose is coordination, not replacing everyone else.

Does a small business need a new governance department

Not automatically. One person may combine several roles if their authority, required expertise and matters needing other participants remain clear.