Organise Work: Human-Agent Teams, Workflows and Role Redesign

All courses / AI Leader / Module 2 / Chapter 10

Learning objectives

After the lesson the participant can:

• Apply the five-question allocation rule to the tasks of a workflow.
• Distinguish the workflow owner, the mandate owner, the agent operator and the platform owner.
• Explain how shadow AI is found and legitimised without watching individuals.
• Draft an accountability map with owners, deputies, verification budget and test results for each step.

Key concepts

Task as the unit of allocation

A job is a bundle of tasks, and AI takes over, helps with or leaves alone individual tasks. Asking which jobs AI will replace is a question the evidence does not support; asking which tasks in a workflow go where can be answered with evidence.

The allocation rule

Five questions per task, the lowest answer governing: can a named person check the output more cheaply than producing it; can the action be undone; is there evidence on the organisation's own tasks; does the task teach the next generation; does it decide about a named person or a decision law assigns.

Four owners

The workflow owner answers for the outcome; the mandate owner for each agent's mandate and output; the agent operator for its identity and credentials, with authority to revoke; the platform owner and the provider for platform and model. Conflating them is the usual failure.

Verification budget

The number of agents a conductor can direct is set by the output the conductor and team can verify properly, not by how many agents can be started. The budget, in hours per week and sampled items per agent, is a figure the owner commits to in the map.

Shadow AI

Use of AI that no one in authority decided on. It shows that people have allocated tasks on their own initiative. It is found by asking and by aggregate, service-level measurement, never by covert monitoring, and a time-limited amnesty makes the answers honest.

Interactive exercises

Do them here and get feedback at once. Choose a card and then where it belongs, or drag it.

Teaching plan

Discussion

1. The Nordvik caseworkers used a consumer chatbot because the approved assistant returned stiff wording. What does hidden use in your organisation tell you about the approved tools and about the allocation of tasks?
2. In ACC-HM-2029-01 the triage agent may allocate by claim type, queue length and language, but not by handler speed or handler history. Why does that line matter under Annex III, point 4(b), of the AI Act?
3. Vendors describe agents as colleagues. What goes wrong in accountability when an organisation takes that description literally?

Written exercise

Self-test

Choose an answer to see whether it is right, or open the answer below each question.

1. According to the chapter, what sets the number of agents a conductor can direct?

   1. The number of agent licences the platform owner has purchased
   2. The throughput of the agents measured in items per hour
   3. The amount of output the conductor and the team can verify properly
   4. The span of control that the organisation recommends for the conductor's human team
   Show the answer

   C. The amount of output the conductor and the team can verify properly
   The limit is verification, not how many agents can be started. That is why spans for human-agent teams cannot be read off the agents' throughput and why the map carries a verification budget.

2. What did Nordvik Region decide in AIDR-1010 about the caseworkers' use of a consumer chatbot?

   1. To add a plain-language template to the approved assistant and rule that case text is never entered in an unapproved tool
   2. To prohibit all chatbot use in the unit and log each caseworker's tool use
   3. To approve the consumer chatbot for case text after a data protection review
   4. To ignore the use because it was a show of hands and not a measurement
   Show the answer

   A. To add a plain-language template to the approved assistant and rule that case text is never entered in an unapproved tool
   The decision was neither to prohibit nor to ignore. The reason for the hidden use, stiff wording from the approved assistant, was fixed, and a clear rule was set for case text.

3. Who answers to the customer when a letter drafted by an agent is wrong?

   1. The provider of the model, under the terms of its contract
   2. The workflow owner, whether an agent or a handler wrote it
   3. The platform owner, who runs the organisation's AI platform
   4. The agent operator, who holds the agent's credentials
   Show the answer

   B. The workflow owner, whether an agent or a handler wrote it
   The workflow owner answers for the outcome of the workflow across every step, human and agent. Neither the platform owner nor the provider answers to the customer.