Who is managing the AI workers?

AI is moving from producing information to executing work. Generative AI responds to instructions with text, analysis or code. Agents go further. Depending on design and permissions, they plan sequences of actions, work inside software systems, retrieve information, carry out tasks and react to changing conditions. An agent can process a customer request, update records, produce a report or coordinate activity across several applications.

That opens real opportunities to cut repetitive work and speed up execution. It also changes what people do. When AI starts working independently, people do not stop working. Their job shifts from doing the work to supervising it: monitoring processes, validating results, investigating odd behavior, handling exceptions and deciding when to step in.

The question is not whether agents can do useful work. It is whether the whole human-agent workflow delivers better outcomes with acceptable effort, reliability and risk.

§The promise

Agentic AI is a different model of automation. Instead of producing one response and waiting, an agent works through several activities toward an objective. In procurement it might gather supplier data, compare proposals, flag missing documents, prepare a recommendation and update the system. In customer operations it might classify requests, pull account information, draft replies and route exceptions. In software it might inspect code, propose changes, run tests and open a pull request.

The potential gain is large. But agents work inside workflows full of dependencies, ambiguity, sensitive information, quality requirements and decisions of varying weight. As autonomy grows, organizations must decide which actions an agent may take, what triggers human review, and how errors will be caught and fixed. An agent that completes ten steps well and causes a serious problem on the eleventh can be worth less than a more constrained one.

The measure is not how many actions an agent completes on its own. It is how reliably the whole process reaches its intended result.

§A new division of labor

Delegating execution to software creates six responsibilities that someone has to own. They can sit with business owners, process managers, IT, security or operations, but they cannot sit with nobody.

Without clear ownership, nobody is sure who should intervene or who has the authority to make the final call.

§The supervision burden

Research on human-automation interaction has long shown that automation changes human workload instead of simply removing it. When people move from doing a task to supervising it, they can struggle to keep track of what the system is doing and to judge when to intervene. Parasuraman, Sheridan and Wickens's model of types and levels of automation remains the established foundation for thinking about how automation shifts human authority and oversight. The NIST AI Risk Management Framework adds that roles, responsibilities, monitoring and risk management need defining across the whole AI lifecycle.

Agent autonomy has to be judged together with the supervision and controls that make it safe and effective.

Human supervision does not mean the automation failed. In many workflows it is essential. The question is whether it is designed well and whether its cost is counted.

§Five hidden costs of agentic AI

§The autonomy paradox

More autonomy is often assumed to mean less human work. The relationship is more complicated. At low autonomy, people do most of the work and AI assists. At intermediate levels, agents execute selected activities while people supervise key decisions and exceptions. At high levels, agents complete whole workflows within defined constraints.

Higher autonomy cuts routine effort. It also raises the consequences of unexpected behavior and makes failures harder to spot and understand. A highly autonomous system can run smoothly in normal conditions and then need substantial expert intervention when it fails. That is the autonomy paradox: the less people are directly involved, the more their judgment matters in the exceptional case.

It grows sharper when one person supervises several agents. Someone can oversee a handful of well-designed, reliable processes. If those processes throw frequent, poorly explained exceptions, the load becomes unmanageable. More autonomy is not always the most efficient model. The right level depends on task complexity, system reliability, whether actions can be reversed, the controls available and the cost of failure. Briefing 07, Cognitive Fit, adds a human factor: people differ in how much control they need to supervise well.

§Measuring the real efficiency of agents

Agents should be judged on the total effort and outcomes of the whole workflow. Counting completed actions or automated tasks is not enough. A useful starting point:

Read the result alongside quality, reliability and risk. Six indicators show where supervision is eating the gain.

• Intervention rateHow often people interrupt, correct, restart or redirect the agent.
• Exception rate and resolution costHow many cases escalate, how complex they are and how long they take to resolve.
• First-pass acceptanceShare of agent-completed work that meets the standard without correction.
• Recovery incidentsHow often actions have to be reversed, and what reversal costs across systems.
• Supervision loadHuman hours spent overseeing agents, and how many processes each supervisor can realistically cover.
• Escalation clarityWhether every exception has a defined owner and a defined route.

§An agent that works, and a queue that grows

Take a hypothetical service organization that deploys an agent to handle incoming customer requests. It classifies requests, retrieves account data, drafts and sends replies for routine cases, and escalates the rest. Routine cases now close in minutes, and the dashboard shows a high automation rate.

Look at the whole workflow and the picture changes. Escalated cases now reach a small senior team with less context than before, because no one has handled the early steps. Some automated replies apply an outdated policy and must be corrected with the customer. Permission changes and rule updates need a weekly coordination meeting between operations, IT and security. No one is formally accountable for the agent's decisions, so team leads review escalations twice.

Illustrative example, not an observed result.

The agent creates value. The workflow around it leaks much of it. The fix is not to switch the agent off. It is to pass context with every escalation, tie replies to the current policy source, assign a named owner, and expand autonomy only as the exception rate falls.

§Designing supervision that works

We recommend six design principles for agentic workflows.

§The GrundMind view

GrundMind evaluates agents as part of a human-agent workflow, not as standalone technology. The question is never only what the agent can do, but what it takes for the organization to let it do it safely and productively.

Two dimensions carry this briefing. AI Net Efficiency counts the total human effort in an agent-supported workflow, including set-up, monitoring, exception handling, correction and coordination, against a baseline. Governance & Risk examines whether decision authority, escalation, auditability and accountability are defined well enough for the autonomy the agent has been given. AI Fit Analysis adds whether the task suits agentic execution at all, and whether the people supervising have the information and control they need. AI Value Leakage locates where gains are lost to exception queues, rework and duplicated review.

The outcome is a recommended level of autonomy per workflow, with the controls and ownership it needs and the evidence that would justify expanding it.

§Five questions for the leadership team

1. Who is accountable for each agent we run?If the answer is unclear, the cost is already showing up as delay and duplicate review.
2. Do our agent business cases include supervision?Monitoring, exceptions, correction and coordination are part of the price.
3. Where do exceptions go, and can that team absorb them?Automation that concentrates hard cases on a few specialists creates a new bottleneck.
4. Can our supervisors see what the agents are doing and why?Oversight without visibility is not oversight.
5. Is autonomy set by risk and evidence, or by capability?The right level is the one the workflow can supervise reliably.