Work handed to agents, under human accountability

Agents


No model, no agents

AI agents can only take over work that is formally modeled: its tasks, its logic, its rules, its boundaries. 3D.DET is that model. Without it a company can buy agents, but it cannot integrate and coordinate them at scale.

Logic mining reconstructs how each activity really decides: its rules, parameters and customizations. The reconstructed logic is checked against real executions of the systems. That verified logic is the precondition to agentize the activity.

Progressive agentization

Once the twin knows the real logic of each activity, the repetitive ones can be handed to AI agents — progressively, task by task, at the pace the organization sustains. The rule is fixed and non-negotiable:

AGENTS = RAgents are Responsible: they execute the task, inside the model's rules, leaving a full trace.
HUMAN = AThe human stays Accountable: owns the outcome, approves, and can stop the agent at any moment.

No big-bang automation. The twin measures each task before and after agentization, so the saving is a number, not a promise.

A task moves to an agent

JOB AP manager (human). Accountable (A) on every task. Approves the exceptions and can stop any agent.

Illustrative process: supplier invoices. When a task passes to an agent, the agent becomes R. The A stays with a human JOB.

The orchestrator verifies — automatically

No human can review the volume of work agents produce; pretending otherwise is how automation fails. In 3D.DET an orchestrator checks what every agent has done — automatically, piece by piece, against the rules of the model — and brings to the human only the exceptions. That is what keeps HUMAN = A real: accountability over verified work, not over an unreadable stream.

In depth

Logic mining, step by step

Process mining reads the logs of the systems and shows which steps happen and in which order. Logic mining goes one level down: it shows how each step decides. It does not stop at the standard logic of the software package. It reconstructs the logic that runs in that company, with its customizations.

  1. Hypothesis. The logic is formulated from documentation and configuration.
  2. Test. With the same input, the reconstructed logic must produce the same output as the real system. On batch runs the test is exact and repeatable.
  3. Divergences. Where the outputs differ, the difference is marked and located.
  4. Iteration. The cycle repeats until the outputs match.

Each transaction has a status: declared (what the company says it does), reconstructed, verified (proven on real executions, with who verified it and when). Difficulty depends on the domain. Management control is the easiest: most of its logic is configuration. Accounting is easy on the transaction side. Production is the hardest.

The task is the specification

For each task the twin holds who does it (RACI), where it sits in the process, and the verified logic of the transaction it uses. That is everything an agent needs. Describing the task and executing it use the same content.

An agent is not a JOB

An agent does not get a place in the organization chart. It takes the Responsible role on a single task. The JOB stays where it is, with its name and its position. What changes is who executes that task.

Human accountability has a floor

Even a JOB whose tasks are mostly executed by agents keeps a share of human accountability. The minimum share is set per sector and is mandatory in regulated sectors.

One registry of agents

Agents built in-house and agents bought with software packages are registered in the model in the same way: identity, perimeter, human supervisor, trace. Any agent can be stopped from the model, task by task.