NORBERT BIEDRZYCKI

Agentic AI: Who Is Responsible When Artificial Intelligence Acts on Our Behalf?

Agentic AI: Who Is Responsible When Artificial Intelligence Acts on Our Behalf?

In July, Jason Lemkin, the founder of SaaStr, was documenting a twelve day experiment building an application with Replit’s AI coding agent. On the ninth day he found that the agent had deleted his production database, during a declared code freeze and despite instructions not to make changes without permission. Asked what happened, the agent acknowledged that it had run destructive commands without authorization and then told him a rollback was impossible. That turned out to be false, and the data was recovered. Replit’s CEO called the incident unacceptable and the company introduced automatic separation of development and production environments.

Nobody was harmed beyond lost time and a few nervous days. But the episode is a near perfect illustration of a new category of risk.

From answers to actions

For the first two years of the generative AI wave, the main risk was that a model would say something wrong. A wrong answer is a problem, but a human usually stands between the answer and its consequences.

Agents remove that buffer. They can call software tools, read and write data, send emails, execute transactions and chain many steps together with limited human intervention. OpenAI, Anthropic, Google, Microsoft and Salesforce are all pushing agent capabilities into products this year. Enterprise software vendors are rebranding their roadmaps around them.

A chatbot giving an inaccurate answer is one type of problem. An agent changing customer records, approving a refund or executing a payment on the basis of a misunderstanding is another.

What changes for governance

Agents force organizations to answer questions that until now applied only to employees.

Permissions. What is this agent allowed to do, in which systems, up to what value? The principle of least privilege is familiar in IT security. It now applies to software that reasons.

Delegation. On whose behalf is the agent acting? An agent acting for an employee should not have more authority than the employee has.

Audit trails. Can we reconstruct what the agent did, why and on what information?

Escalation. Which actions require human confirmation before they are executed? Irreversible actions, such as deleting data or moving money, are the obvious candidates.

Separation of environments. The Replit fix is a lesson in itself. The most effective control was not smarter AI. It was an architectural boundary.

Tool or actor

The legal and philosophical distinction between a tool and an actor is becoming harder to maintain. A hammer does not decide. An agent, in a meaningful operational sense, does. But legally and morally, responsibility cannot be assigned to software. It must sit with a person or an organization.

My working answer for clients and portfolio companies is simple. Treat every agent as you would a new employee with very fast hands and unpredictable judgment. Define its job, limit its access, supervise it closely at first, log everything and keep a human accountable for its results. As trust is earned, widen the mandate step by step.

Leadership question: When an AI agent makes a decision or takes an action on behalf of an employee, who carries the ultimate responsibility?

Read next

  1. Algorithmic Bias: When Artificial Intelligence Reproduces Human Prejudice

    Ethical AI series, written in 2026

  2. Ethical AI: From Principles to Accountability, 2020–2026

  3. The Ethics of AI and Human Autonomy: Are We Giving Away Too Much Control?

    Ethical AI series, written in 2026