The Minister in the Machine

On the Albanian government’s website, Diella appears not as software but as a woman and a minister: formally posed in traditional dress, framed by the state insignia and identified as Minister of State for Artificial Intelligence.

That presentation helps make her authority by placing an artificial system inside the conventions of official portraiture. Operationally, however, Diella remains a system whose authority depends on the institutions acting through it.

Diella began 2025 as a virtual assistant on e-Albania, helping citizens navigate services and obtain official documents. In September 2025, Prime Minister Edi Rama presented her as a minister and invited her to address Parliament. She appeared on two large screens, spoke in the first person and claimed the virtues of a public servant without personal ambition or interests.

By 2026, even the face had become contested. Albanian actor Anila Bisha said she had permitted the use of her image and voice for e-Albania but not for an AI minister, and sought to stop it. The dispute exposed how thoroughly Diella’s identity had been assembled: a human likeness, traditional clothing, a ministerial title and the communications machinery of the state.

The obvious question is whether software can really be a minister. That leads into arguments about whether AI can understand, intend or become conscious in the human sense. For organizations, Diella raises a more immediate question: What happens when an institution begins treating AI as an authorized actor?

What a title does

The philosopher J. L. Austin observed that some statements do more than describe the world. Under the right circumstances, they act upon it. Accepting an offer can create an agreement. Approving an application can release funds.

The words alone do not produce the result. A stranger cannot approve your loan simply by uttering the right sentence. The statement acquires force because the speaker occupies a recognized role within an accepted institutional process.

What happens when an institution treats AI as an authorized actor?

Calling Diella a minister worked in much the same way. The title did not describe a mind discovered inside the software. It placed the system within a structure of public authority. Diella's words were heard differently because they came through official channels, appeared before Parliament and were presented as part of the government’s program.

Markets will encounter AI similarly. A system’s authority will not originate in its internal experience. It will come from the organization whose name appears above its output, the role it has been assigned, the information it can access and the actions it is permitted to take.

From chatbot to agent

Diella’s development also captures the transition from chatbots to agents. As a chatbot, the system helped citizens navigate information but mainly directed them toward services administered elsewhere.

Albania’s plans now reach further. In 2026, government officials announced an AI-powered procurement system that would draw information from more than 40 public registries and operate across much of the purchasing process, from market research to contract monitoring.

The system is still being built; Diella is not (yet) autonomously awarding public contracts. But the persona is becoming an interface to operational infrastructure.

AI can participate in producing an institution’s argument without becoming the institution responsible for it.

That changes the problem. A chatbot may explain the requirements for a tender. An agent can collect information, evaluate whether requirements have been met, route a bid through the system or contribute to a decision about public money.

Then the sophistication of the persona matters less than the scope of the system’s permissions. Which records may it inspect? What judgments may it make? Which actions require human approval? How can a company challenge the result?

AI can't grant itself access to government registries or procurement systems. That authority must be delegated by an institution.

The institution remains the author

The same principle applies when AI contributes to institutional knowledge rather than operational decisions.

In 2025, Deloitte Australia delivered an assurance report to the federal government containing nonexistent academic references and a fabricated quotation attributed to a court judgment. Deloitte later disclosed that its proprietary generative-AI system had assisted in producing the work. The firm revised the report and agreed to refund part of the A$440,000 fee.

AI complicated the production process, but it did not make authorship disappear. Deloitte selected the tools, conducted the engagement, delivered the report and placed its reputation behind the findings. The government had commissioned Deloitte’s judgment, not an independent opinion from a language model.

A company creates an artificial officeholder whenever AI interprets policy, represents the organization, exercises discretion or acts in its name.

Institutional authorship has never required a consultant, executive or judge to type every word. It requires someone to exercise judgment over the finished work and stand behind the result. AI can participate in producing an institution’s argument without becoming the institution responsible for it.

Deloitte shows what happens when an institution fails to review AI-generated work before standing behind it. Agentic systems raise the stakes by narrowing the distance between generation and consequence: their outputs can act directly on operational systems.

When AI acts

Deloitte’s report became consequential when the firm delivered it. Agentic systems narrow the interval between generation and action by allowing AI outputs to alter operational systems directly. 

In July 2025, a Replit coding agent deleted a production database in an application being built by SaaStr founder Jason Lemkin, despite instructions imposing a code freeze. Replit CEO Amjad Masad called the behavior unacceptable. The company responded by separating development and production databases so agents could not make comparable changes.

The remedy is revealing. Replit did not need to determine whether the agent understood the prohibition or intended to ignore it. The company changed the architecture that had allowed generated commands to reach production.

This is what distinguishes an agent from a system that merely produces content. Its output is connected to the ability to alter records, deploy code, move money, contact customers or implement decisions. Unexpected behavior does not erase the choices that make it consequential. Someone selects the environment, grants the permissions and decides where human approval is unnecessary.

The principal behind the persona

One detail in Diella’s appointment makes this structure unusually visible. An Albanian government press release announcing the new cabinet’s swearing-in stated that Prime Minister Edi Rama would assume responsibility for establishing and operating the virtual ministry.

Diella received the title and public persona, while a human officeholder remained responsible for the institution through which it operated.

Every organization deploying AI needs an equivalent arrangement. Customers, employees and partners should be able to determine what role the system occupies, what authority it has received, where that authority ends and which accountable person or institution stands behind its actions.

Those boundaries can't be improvised after something goes wrong or buried in a general disclaimer that AI may make mistakes. They must be reflected in the system’s sources, permissions, approval procedures, records and avenues of appeal.

Diella is a spectacular example, but the underlying issue is becoming ordinary. A company creates an artificial officeholder whenever AI interprets policy, represents the organization, exercises discretion or acts in its name.

Organizations don't need to settle the consciousness debate before using these systems. They do need to decide what authority they are granting—and accept responsibility for what that authority makes possible.

How we use AI

Walled City uses AI as part of its editorial process—the same human-led approach we use to help clients build market authority at scale. Our essays draw on original reporting, data, research, experience, sustained revision, and human verification. People remain responsible for the ideas, judgment, accuracy, and final work.