Watch out for the AI pass-the-buck manager

Many companies seriously getting started with Agentic AI are currently building roughly the same structure. At the top is one chief agent acting as a kind of managing director. You discuss your questions, assignments, and priorities with this agent. Subsequently, that agent puts various sub-agents to work, each executing their own task. One agent does research, another analyzes figures, yet another writes texts, processes transactions, or handles customer inquiries. For humans, this is particularly attractive because you only have one point of contact while simultaneously gaining the clout of an entire digital team.

That sounds efficient, and it often is. Certainly in customer service, the temptation is great. A single chief agent can receive a customer inquiry, have a sub-agent check the order history, have another agent consult the return policy, and have yet another agent formulate an appropriate response. Yet there is a serious risk in this way of organizing. The more tasks you route through a single chief agent, the larger the black box becomes. You still see what the chief agent reports back to you, but you have hardly any visibility into what happens underneath. You don't know exactly which sub-agent used which information, what assumptions were made, where errors originated, and how different agents coordinated their work with each other. Precisely because you only talk to that one chief agent, many of the natural control moments disappear from the process.

The pitfall of the pass-the-buck manager

This creates a curious parallel with a type of manager many people from organizations will recognize: the pass-the-buck manager. That is the manager who receives an assignment from above and then pushes it down to the team as quickly as possible. Something is forwarded via email, delegated, or dumped in a meeting, after which others are left to figure it out. In customer service, you see this, for example, when a complex customer complaint is simply passed on to another team without anyone truly taking ownership. The manager thereby saves time, but simultaneously becomes increasingly detached from the actual work. Ultimately, he hardly knows what exactly his team does, how decisions are made, and where the risks lie.

With Agentic AI, we can make exactly the same mistake. We could build a system where we only input assignments at the top and receive a result at the end. We outsource everything in between to agents. For example, a customer reports that their order was delivered incorrectly for the third time. The chief agent puts various sub-agents to work, checks the history, looks at the compensation rules, has an answer written, and perhaps automatically initiates a new shipment. At the end, we only see that the case is resolved. That might seem like the ultimate proof of efficiency, but in fact, we risk becoming the digital variant of a box-mover. We add less and less substantive value while simultaneously becoming completely dependent on a system we understand less and less.

A major conceptual error is that we primarily try to optimize the collaboration between human and agent for time savings. Every human interaction is then quickly seen as something that ultimately must disappear. The less consultation, the better. The less control, the smarter the system. The more autonomy, the more mature the application. In customer service, this quickly translates into goals like less human handling, fewer escalations, and the highest possible percentage of fully automated customer contacts. But good management doesn't work that way in the human world either. A good manager doesn't try to eliminate every moment of contact with their team. A good manager actually wants to understand what is going on, wants discussion, wants pushback, wants to be able to evaluate choices, and wants to remain personally involved at crucial moments. Why should we suddenly assume with agents that every form of interaction is inefficient?

Not blind autonomy, but conscious collaboration

The interesting question is therefore not how much work an agent can execute independently, but what form of collaboration between human and agent fits a specific process. This is certainly true within customer service, where a simple question about the status of a package requires something completely different than a complaint from a loyal customer who threatens to leave after multiple mistakes. Various models exist for this, and they do not all have to be as autonomous as possible.

Six models for human-agent collaboration

The simplest form is the agent as on-demand help. In this model, the human does the bulk of the work themselves and the agent is only brought in at specific moments. A customer service representative could, for example, ask the agent to draft a concept response, summarize a customer's history, find a relevant article from the knowledge base, or double-check a complicated complaint. The agent has no independent decision-making authority and only becomes active when you consciously ask for it. This may sound hardly revolutionary, but for processes where expertise, responsibility, and nuance are important, this can actually be an excellent form of collaboration.

A step further is the agent as a work preparer. In this model, the agent does all the preparatory work. It gathers information, makes analyses, drafts documents, and lines up actions, but the moment a real decision has to be made, the agent returns to the human for approval. Within customer service, the agent could, for example, analyze the entire history of a customer, check which deliveries went wrong, pull up the applicable compensation policy, and then propose to issue a refund or resend a product. The employee then only needs to assess whether that is indeed the right solution. The agent thus takes a lot of work off your hands, while the responsibility for important choices explicitly remains with the human. You save a lot of time without the human disappearing from the process.

A third form is the agent as a sparring partner. Here, there are multiple moments in the process where human and agent actively collaborate. Take, for example, a complex customer complaint where multiple contacts have already taken place and the customer now indicates they want to leave. The agent can analyze all previous conversations, transactions, and contact moments and propose several possible solutions. Subsequently, you engage in a dialogue together about what is wise. Is a standard compensation sufficient? Should a personal gesture be made? Is there perhaps a structural problem that needs to be solved first? The agent helps with thinking and formulating, but ultimately the human chooses which solution is offered and how it is communicated to the customer. So, in this model, you don't try to minimize interaction. You actually use the agent to increase the quality of that interaction.

A fourth model is working with fixed decision or control moments. The agent is given a lot of room to work independently here, but there are fixed moments when human and agent come back together. A customer service manager could, for example, discuss with the agent every Friday what types of questions came in, which complaints are increasing, what exceptions the agent made, what refunds were given, and at what moments customers still had to be escalated to an employee. You do not check every individual customer case then, but you consciously build in moments where you maintain a view of the bigger picture. It essentially resembles a regular management meeting, except there is an agent on the other side of the table.

The fifth form is a model in which the agent is allowed to act completely independently as long as it stays within previously agreed boundaries. The human is only involved in exceptions. Within customer service, an agent could, for example, independently approve a return, adjust an order, or provide compensation up to fifty euros. Above that amount, an employee must give approval. You can also set other boundaries, for example, that a customer with three previous complaints must always be assessed by a human, that a legal threat is escalated immediately, or that the agent asks for help as soon as it lacks sufficient certainty about the right solution. In this model, the role of the human shifts from assessing every task to designing and monitoring the rules within which the agent is allowed to act.

The most extreme form is full autonomy. The agent analyzes, decides, and executes independently, while the human only occasionally performs spot checks to see if everything is going well. For highly standardized processes with limited consequences, this can work perfectly fine. Think, for example, of simple questions about opening hours, the status of an order, or resending an invoice. However, it becomes dangerous when full autonomy is automatically seen as the ultimate goal of every Agent-AI application. After all, maximum autonomy is not the same as a mature application.

Conscious design for interaction

Which form of collaboration fits best depends entirely on the process. A simple question about a delivery requires a different setup than an escalation from a dissatisfied customer, a refund with major financial impact, or a complaint that could possibly cause reputational damage. This is precisely why organizations must think much more consciously about the question of where the human should remain in the process. When should an agent escalate? At what moments do we want explicit approval? When do we actually want to spar over content? Which decisions may be taken fully automatically? What information must always remain visible? And at what moments do we consciously want to check again whether the system is still doing what we expect?

These are not annoying limitations surrounding Agentic AI. This is exactly where real design begins. We must therefore abandon the idea that Agentic AI is mainly successful when the human has to talk to it as little as possible. Sometimes, that very interaction is the most important part of the entire system. Certainly in customer service, where a customer contact is not just a task that needs to be handled, but can also be a signal that something is going wrong in a product, process, or service. Anyone who only optimizes for time savings runs the risk of building an organization where the agents become increasingly smart and the humans understand less and less of what is actually happening.

Then you might have an impressive digital team, a chief agent coordinating everything, and dozens of sub-agents working for you day and night, but you yourself are reduced to the one who occasionally throws an assignment over the digital fence and then waits for the result. Your schedule is empty, your customer service keeps running, and you hardly have any visibility into the work anymore. After all, you can also be a soulless manager of robots.