Agent Collaboration Raises Accountability Questions

Gartner projects that by 2028, the average global Fortune 500 enterprise will run more than 150,000 AI agents, up from fewer than 15 in 2025. In the same report, only 13% of 360 survey respondents said their organizations had the right governance in place. As companies coordinate agents across workflows, responsibility for an output that causes harm can become difficult to assign. A California law that took effect Jan. 1 bars defendants from using an AI agent’s autonomous operation as a defense against a liability claim.

Ali Alkhafaji, CEO of Apply Digital, says organizations need to assign a named human owner to every system output before deployment, regardless of how many agents helped produce it. His organization starts each client engagement with at least 50% agents working alongside human practitioners in “matrix teams.” He says that if an organization cannot name the person responsible for an output, it is not ready to deploy the system.

Alkhafaji says handoffs between agents need careful design because an early decision can limit what later agents know, allowing errors to compound before a human review. His team uses a “graceful handoff” that passes context along with the task. A useful handoff record captures at least four items: the agent’s objective, the data it could access, its confidence level and the action it took. These records can help a human decide what to do without first reconstructing the process.

He argues that audit trails should record the reasoning chain, not just final outputs. Before deployment, his team defines which inputs, confidence signals, actions and handoffs to log at every agent boundary. Without those records, incident reviews rely on guesswork; failures may instead emerge gradually as eroding trust, compliance gaps discovered months later and decisions that cannot be traced to an owner. Alkhafaji urges IT leaders to map active agent workflows and name an owner for every output, document handoff protocols at each boundary, and check that logs capture inputs, confidence signals and reasoning steps as well as final outputs. He says retrofitting audit infrastructure after an incident costs significantly more than building it at the start.

Gartner projects that by 2028, the average global Fortune 500 enterprise will run more than 150,000 AI agents, up from fewer than 15 in 2025. In the same report, only 13% of 360 survey respondents said their organizations had the right governance in place. As companies coordinate agents across workflows, responsibility for an output that causes harm can become difficult to assign. A California law that took effect Jan. 1 bars defendants from using an AI agent’s autonomous operation as a defense against a liability claim.

Ali Alkhafaji, CEO of Apply Digital, says organizations need to assign a named human owner to every system output before deployment, regardless of how many agents helped produce it. His organization starts each client engagement with at least 50% agents working alongside human practitioners in “matrix teams.” He says that if an organization cannot name the person responsible for an output, it is not ready to deploy the system.

Alkhafaji says handoffs between agents need careful design because an early decision can limit what later agents know, allowing errors to compound before a human review. His team uses a “graceful handoff” that passes context along with the task. A useful handoff record captures at least four items: the agent’s objective, the data it could access, its confidence level and the action it took. These records can help a human decide what to do without first reconstructing the process.

He argues that audit trails should record the reasoning chain, not just final outputs. Before deployment, his team defines which inputs, confidence signals, actions and handoffs to log at every agent boundary. Without those records, incident reviews rely on guesswork; failures may instead emerge gradually as eroding trust, compliance gaps discovered months later and decisions that cannot be traced to an owner. Alkhafaji urges IT leaders to map active agent workflows and name an owner for every output, document handoff protocols at each boundary, and check that logs capture inputs, confidence signals and reasoning steps as well as final outputs. He says retrofitting audit infrastructure after an incident costs significantly more than building it at the start.