# The Accountability Gap: Who Is Responsible When AI Makes Decisions Independently?

Author: Norbert Biedrzycki
Published: 2026-02-15
Language: en
URL: https://norbertbiedrzycki.pl/en/the-accountability-gap-who-is-responsible-when-ai-makes-decisions/
Summary: Consider a realistic scenario for this year. A company uses an AI agent to process supplier invoices. The agent is built by a software vendor on top of a foundation model from another provider,…

Consider a realistic scenario for this year. A company uses an AI agent to process supplier invoices. The agent is built by a software vendor on top of a foundation model from another provider, integrated by a consulting firm and supervised, in principle, by a finance team. One day it pays a fraudulent invoice after being manipulated by instructions hidden in an email attachment.

Who is responsible? The model developer, whose model followed the hidden instructions? The software vendor, who designed the agent? The integrator, who configured its permissions? The finance team, who approved the process? Or the company, whose money is gone?

Each can make a reasonable argument that the failure belongs to someone else. That is the accountability gap.

## How the gap appears

In earlier generations of AI, the technology mostly supported human decisions. A model produced a score, a human made the call. Responsibility followed the human.

Agentic systems change that. They execute multi step workflows, interact with external services and take actions with real consequences, often faster than any human could review. The decision is no longer a single moment that can be attributed to a single person. It is distributed across a chain of components, each built by a different party.

Courts are beginning to grapple with this. In the Mobley v. Workday case in the United States, a court allowed claims to proceed on the theory that an AI vendor screening job applicants for employers could itself be liable as the employers' agent, and later allowed the case to proceed as a collective action. The Air Canada chatbot ruling from 2024 established, at least at a small claims level, that a company cannot disclaim responsibility for what its AI tells customers.

Regulators are moving too, though not always in a straight line. The EU AI Act defines obligations for providers and deployers. In November the Commission proposed, as part of a broader simplification package, to postpone the application of some high risk obligations. Whatever the final timing, the direction remains the same: responsibility follows roles, and deployers cannot outsource theirs.

## Responsibility cannot disappear

The philosopher Hans Jonas argued that technological power requires a new ethics of responsibility, because our actions now have consequences that reach further than our ability to foresee them. His argument was about civilization. It applies equally well to a finance department running an AI agent.

Complexity is not an excuse. If anything, it increases the need for deliberate design of accountability.

## Closing the gap in practice

**One accountable owner per agent.** Every autonomous system needs a named business owner who answers for its outcomes, regardless of who built it.

**Explicit decision boundaries.** A written mandate: what the agent may decide alone, what requires confirmation, what it may never do.

**Auditability by design.** Logs that make it possible to reconstruct what happened and why.

**Contracts that reflect reality.** Supplier agreements that define responsibilities, testing obligations, incident reporting and liability for AI components.

**Escalation and kill switches.** Clear procedures for stopping an agent and rolling back its actions.

None of this is exotic. It is the same discipline organizations apply to delegating authority to people. The difference is that people come with judgment, context and personal accountability. Agents do not, so the organization must supply all three.

> *Leadership question: Can organizations delegate decisions to AI without delegating accountability?*
