Task automation: learning to redistribute work between humans and AI agents
Since generative artificial intelligence entered professional use, one question keeps coming back: how many jobs will be replaced by AI? It draws attention because it directly affects workers, executives, and the future of many professions. Yet it does not always help us understand what is actually happening inside organizations.
A job is not a single activity. It brings together dozens of tasks that require very different skills, levels of judgment, and responsibilities. Some are repetitive, others demand analysis. Some consist of searching for or transforming information, while others rest on a human relationship, a decision, a negotiation, or a professional responsibility. When artificial intelligence enters this environment, it therefore does not necessarily transform the occupation as a whole. It first changes the way some of its tasks can be carried out.
The arrival of AI agents now adds a further dimension. Artificial intelligence can progressively move beyond producing a text, an analysis, or a recommendation to carrying out a sequence of actions, using various tools, interacting with systems, and pursuing an objective with a degree of autonomy. The company must then move beyond the question “which tasks can we automate?” to ask a far more structuring one: how should we distribute work among humans, AI agents, and automated systems in order to obtain the best result? This question profoundly changes the way work is designed.
Traditional automation mainly sought to remove human work
For several decades, automation mainly consisted of identifying repetitive tasks that could be handed over to a machine or a piece of software. In a factory, a robot could perform a repetitive physical operation. In an office, a computer system could calculate an invoice, transfer data, or automatically produce a report. The objective was generally fairly clear: reduce the human time required to carry out an activity predictable enough to be codified.
Artificial intelligence considerably widens this territory. It can take part in activities that require understanding language, searching for information, summarizing documents, analyzing certain situations, producing content, interpreting data, or proposing actions. With AI agents, several of these capabilities can now be assembled into a more complete work sequence.
The boundary between human work and automated work therefore becomes far less clear-cut. An activity no longer necessarily has to be fully automatable to benefit from AI. Some parts can be handed to an agent while a human retains those that demand more judgment, responsibility, or interaction. The objective thus shifts from automation to the orchestration of work.
An occupation is an assembly of tasks
This distinction changes the way we interpret forecasts announcing that a large proportion of occupations will be affected by artificial intelligence. Being affected by AI does not necessarily mean disappearing. An occupation can be profoundly transformed while the person remains essential to carrying it out.
Take a sales representative. Their work may include researching prospects, preparing a file, consulting the CRM, analyzing the customer's history, drafting an email, preparing a meeting, the conversation with the customer, negotiation, preparing a proposal, and following up on the business opportunity. Several of these activities can already be accelerated by artificial intelligence. The commercial relationship, the understanding of context, trust, and certain decisions nevertheless remain strongly human.
The same reasoning applies to an accountant, an engineer, a manager, an analyst, a technician, a human resources professional, or a cybersecurity specialist. Their occupation may remain necessary while the distribution of time across their various tasks changes considerably.
The strategic question then becomes far more precise: which tasks should remain human, which should be augmented by AI, and which can be handed primarily to automated systems or agents?
The best distribution is not necessarily the one that automates the most
It can be tempting to measure an organization's maturity by its level of automation. A company that automates more would necessarily seem more advanced. Yet this logic risks producing poor decisions.
Some tasks can technically be automated while still being better off remaining human. A delicate conversation with an employee, a strategic negotiation, a decision carrying significant consequences, or a relationship with a major customer may require a level of contextual understanding, responsibility, and trust that goes beyond the mere ability to produce a plausible answer.
Conversely, some tasks that organizations continue to assign to highly qualified professionals hold very little value when performed manually. Searching for information across several systems, re-entering data, reformatting documents, preparing a first summary, or checking repetitive items can consume a considerable share of a specialist's time without genuinely drawing on their expertise.
A good distribution of work therefore does not seek to maximize the use of AI. It seeks to maximize the value produced by the combination of the human and technological capabilities available.
The AI agent introduces a new participant in the process
The arrival of AI agents changes this thinking even further. Traditional software generally waits for a precise command. An agent can receive an objective, determine some of the necessary steps, use various tools, and carry on with its work until a result is obtained or until human intervention becomes necessary.
Imagine a procurement process. An agent could monitor certain inventory levels, analyze consumption history, check supplier lead times, prepare a purchasing recommendation, and produce a request ready for approval. In a well-governed context, it could eventually carry out certain additional actions within predefined limits.
The buyer does not necessarily disappear from the process. Their role can shift toward exceptions, negotiation, supplier evaluation, decisions with a significant financial impact, and improvement of the procurement strategy.
The process then becomes hybrid. Part of the work is human, another part is automated, and some activities are carried out jointly.
The process must therefore be redesigned before the tool is chosen
This is probably one of the most important consequences of the arrival of AI agents. A company that starts by buying a technology risks then looking for places to insert it. It ends up adding artificial intelligence to processes that were designed for a world in which practically every decision and interaction was carried out by humans or by traditional software.
The opposite approach is far more interesting. The organization starts by examining the process itself. It seeks to understand why each step exists, what value it produces, what information is required, where delays appear, which decisions carry genuine risk, and which activities needlessly consume human time.
This work sometimes reveals that certain tasks should simply be eliminated. Others can be simplified. Some can be automated with conventional technologies. Others become excellent candidates for AI. And a few must remain under human control.
The transformation then no longer consists of automating the existing process. It consists of designing a better process while taking into account capabilities that did not exist when it was created.
The human role can become more important precisely because certain tasks disappear
Automation is often presented as a reduction of the human role. In many situations, it can produce the opposite effect on the tasks that remain.
When a professional spends less time searching for, copying, reformatting, or preparing information, they can devote more time to understanding, deciding, communicating, creating, negotiating, or improving.
A manager can spend less time compiling reports and more time supporting their team. A sales representative can spend less time preparing files and more time with their customers. An analyst can reduce the time spent gathering data and go deeper into interpreting it. A cybersecurity specialist can delegate certain repetitive analyses and focus their expertise on anomalies and on decisions that carry more risk.
The real gain is therefore not measured only in hours saved. It can also be measured in the quality of the human work that becomes possible when certain lower-value activities are handled another way.
Responsibility cannot be automated as easily as execution
This redistribution of work nevertheless raises a fundamental question: who remains responsible for the result?
An agent can prepare a financial recommendation, analyze a contract, pre-screen candidates, identify an unusual transaction, or propose an operational decision. Its technical ability to carry out a task does not automatically determine the level of autonomy an organization should grant it.
The level of risk must factor into the design of the process. An easily reversible, low-impact action can allow significant autonomy. A decision likely to have major financial, legal, human, or security consequences will generally require more control.
AI governance must therefore be built directly into the distribution of work. Organizations must determine the permissions granted to agents, the data they can access, the actions they can perform, the situations that require validation, and the mechanisms that make it possible to trace their interventions.
The AI agent becomes a participant in the process, but the organization remains responsible for what it allows the agent to do.
Human in the loop should not become a formality
The expression human in the loop has become common as a way of providing reassurance about the use of artificial intelligence. It can, however, create a false sense of security.
An employee who has to approve hundreds of automatically generated decisions gradually risks validating them without genuinely examining them. If the AI almost always produces an acceptable answer, human attention naturally declines. The control officially exists, but its real value degrades.
Human presence must therefore be designed according to risk and to the value of judgment. In some processes, a human must approve every decision. In others, they may intervene only on exceptions, anomalies, or situations exceeding certain thresholds. Elsewhere, automated controls can complement human supervision.
The relevant question is not simply whether a human is present in the process. It is to determine at which moment their judgment genuinely adds superior value.
Managers will have to learn to lead hybrid processes
This shift will also have significant consequences for management. Managers have traditionally distributed work among people. They will progressively have to learn to distribute certain activities among people, automated systems, and AI agents.
They will have to understand these systems' capabilities and limits well enough to know which responsibilities to entrust to them. They will also have to monitor their performance, their errors, their costs, and their effects on operations.
Capacity management may itself change. When a volume of work increases, the first response will no longer necessarily be to add an employee. It will first be necessary to determine which part of that load genuinely requires more human capacity and which part can be absorbed by the systems.
This does not turn the manager into an IT specialist. It makes understanding AI and automation a new dimension of operations management.
Performance indicators will have to evolve as well
If a company redistributes work without changing the way it measures performance, it risks retaining behaviors inherited from the old model.
An employee who produces ten reports manually and an employee who supervises an agent capable of producing a hundred are no longer doing exactly the same work. Measuring only the number of reports produced quickly becomes inadequate.
Indicators must more closely reflect the value created: quality, turnaround times, error rates, customer satisfaction, problem resolution, process cost, capacity for innovation, or improvement in operational results.
Artificial intelligence thus forces organizations back to a fundamental question of operational excellence: what value are we really trying to produce?
The transformation of occupations must be supported
Redistributing work also means redistributing the skills required.
A professional who previously carried out every step of a process may now have to learn to supervise certain automated activities, interpret the results produced by AI, intervene in exceptions, and continuously improve the process.
This requires training, but also transparency. Employees must understand why their work is changing, which responsibilities remain theirs, and which new skills will allow them to maintain or increase their contribution.
An organization that introduces AI agents without addressing this human dimension risks provoking perfectly predictable resistance. Conversely, a company that involves its teams in designing the new processes can use their knowledge of the actual work to determine far more precisely what should be automated and what must remain human.
The people who carry out the work every day often know about exceptions, constraints, and subtleties that are invisible in official procedures. Their participation therefore becomes a condition for the quality of the transformation.
Conclusion
Artificial intelligence is already transforming occupations, but the phenomenon is more subtle than the simple disappearance or creation of jobs. What changes first is the composition of the work.
Some tasks become automatable. Others can be accelerated by AI. Some still require significant human judgment. And agents now introduce the possibility of entrusting far more complete sequences of activities to systems.
Companies must therefore learn to design processes in which humans and AI agents intervene where their respective capabilities create the most value. This demands precise knowledge of operations, clear governance, control over data and access, suitable indicators, and support for the people whose roles will change.
At Quantum Beyond, this thinking sits at the heart of organizational transformation in the age of AI. Our experts work alongside operational, technology, and leadership teams to understand existing processes, identify improvement opportunities, determine where AI can genuinely create value, and build governance that allows organizations to retain control of their operations.
The objective of an AI transformation should therefore not be to replace as many human tasks as possible. It should be to design an organization in which every resource — human or technological — is used where it brings the most value.
The question that will determine companies' performance will soon no longer be only “how many employees do we have?”, or even “how many AI agents are we using?”. It will be: “have we learned to distribute work intelligently between the two?”