By Mostafa Sayyadi and Luca Collina
This article analyses the 22 seasons of Arsène Wenger at Arsenal as a case study of leadership. The point is that football can be a model of AI, and that Wenger’s achievements imply a deployment strategy.
Wenger’s five lessons about leadership are translated into five related choices in AI: make the purpose recognizable for the organization; care about the quality of work as well as the score; develop people and build the organizational capability; adapt the approach according to the changing evidence; and delegate the authority with limits.
1. Recognize the Organization’s AI Purpose
Wenger’s first lesson is to make a leader’s name synonymous with the brand. For AI leadership, the key point is recognizability: people should see the organizational purpose in the decisions about AI. But this identity belongs to the organization, not to the individual. The philosophy of Wenger provided Arsenal with the recognizable direction; in the business context, the purpose of AI should govern the investment, design, and safeguards regardless of the champion of the project.
For example, a leader of the customer service may conclude that AI should assist employees to resolve routine queries faster and to preserve human discretion in the case of distress, disputed charges, and unusual circumstances. Such a purpose should guide the choice of the tool, access to customer data, the rules of escalation, and the measures to assess the pilot project. If the objective is only to increase the number of AI-assisted requests, the team may optimize the wrong result.
Before approval of the pilot project, the leaders should be able to state the work result in clear terms: what should be faster, more accurate, more consistent, or more accessible, and for whom? Also, they should identify the conditions in which this result would not qualify as success. This gives the organization a basis for deciding whether the pilot deserves to be continued.
This statement should connect the use case with the organizational priorities. “Using the generative AI” is an activity; “reducing the time customers wait for a correct answer without increasing repeat contacts” is an outcome that leaders can measure. This difference helps to avoid the common drift: once the tool is introduced, teams start finding tasks for it. The purpose-driven approach starts with the priority, finds the place where people waste time or where the service is inconsistent, and asks whether AI can improve this part of the work under acceptable conditions.
2. Care about the Quality of the Work as well as the Score
Wenger’s second lesson is that attractive play may matter as much as the record. The warning for AI leaders is not to treat one score, such as speed or volume, as the entire result. Arsenal’s story talks about the changes that Wenger made in training and dietary practices, as well as about the change of positions that allowed Thierry Henry to play centrally. The lesson for the business environment is the redesign of the work process and assessment of its quality, not the number of AI uses or speed of task completion.
The introduction of a generative AI assistant into an unchanged workflow creates a new burden of review: the employees have to check the output, correct the mistakes, and finish the entire process. The project looks successful as the small group absorbs this burden. Before implementation, the leaders need to map the process end-to-end: what does AI prepare, recommend, and perform? Who checks the work of the tool? What information may it use? How does it react to uncertainties and mistakes? Where can the step be eliminated? Which new control is necessary?
The work process design is not a technical detail. It determines whether AI reduces the friction or moves it somewhere else. The leader should ask the project team to describe the entire process of transformation from input to the result, including exceptions and handoffs, and redesign the process together with people who perform it. The adoption begins when the new way of work becomes clear for others, not when the demonstration looks impressive.
Consider the customer service team that tests the AI-generated drafts. The pilot project may demonstrate that the agents draft answers faster. Before adoption, the leaders need to understand whether this time is saved with the review, whether the answer is correct for all customer groups, and whether customers have to contact the organization again to resolve the issue. The team may decide to use AI for routine questions, require the agent to check the answer in the case of disputed charges, and exclude situations of distress until the process becomes clearer. This is not only the tool configuration. This is the new service process with new handoffs, skills, and accountability.
The measures should include the entire service outcome. The time of drafting alone may reward the fast first answer despite the later corrections. The leaders can compare the existing and the new process on such factors as the time of resolution, accuracy, repeat contact, escalation, rework, and customer experience. Also, they may measure the time agents spend on reviewing the output of the AI tool. All these measures need a baseline and a designated owner. Otherwise, teams may confuse activity with improvement and discover only during the wide-scale implementation that the process has been shifted from one step to another.
3. Invest in the Emerging Talent and Build Capability
The third lesson is to invest in young players without the assumption that they will stay at the club. It means developing people at different levels of experience while creating organizational knowledge, not relying on individuals who possess it. In the customer service study, less experienced agents received greater benefit from the assistant than the more experienced colleagues. The researchers found evidence that the tool assisted workers in learning from their stronger performers. The study covered only one company and one work type, so it should be the source of the questions rather than a universal promise.
The leaders should examine who benefits from the tool, who bears the review burden, and what kind of capability is being created. New employees may need examples, coaching, and a clear escalation route. Experienced employees may need the authority to challenge the system, identify exceptions, and contribute knowledge to the redesign of the workflow. If the expertise is used only to correct the AI output, the organization may get short-term efficiency, but at the expense of long-term weakening of the human knowledge base.
A good pilot project should not only demonstrate the average speed. It should measure the quality of work in different tasks and experience levels, the patterns of errors, rework, confidence, and learning. It should find out whether experienced people have more room to perform the more difficult tasks or are asked to supervise the increased number of outputs. The goal is not to protect the existing tasks. The goal is to make deliberate decisions on how people’s judgement and AI’s capability complement each other.
Change becomes easier to manage when employees participate in defining the problem and testing the proposed process. People who perform the work know which cases look routine but are not, which information is lacking, and what customers or colleagues should understand. Their role should not be limited to the test of the tool and its evaluation. Ask them to identify the failure cases, negotiate the rules of review, and test the revised handoffs. Explain what is known, what is unknown, and how the evidence will impact the decision. It will give the leaders a more realistic test and employees the opportunity to influence the work process that may change their responsibilities.
This involvement will help to distinguish various reasons. Some employees may doubt the accuracy of the system; some may fear losing discretion, getting a new target, being measured by a new metric, or correcting the increased number of outputs in the same time. Such treatment of all these reactions as resistance misses the valuable information. The right response is to make the role changes clear, to provide the time to learn, and to show how employees can reject or override the output. The training should involve realistic work and exceptions, not only the demonstration of the best capabilities of the tool.
The development of people also needs the environment for applying the acquired skills. Wenger’s 22 seasons included the move of Arsenal to the Emirates Stadium, which was the result of the club’s leadership and financial situation, not Wenger alone. The analogy is that AI adoption also needs investment beyond the visible tool: reliable data, integration, security review, training time, workflow ownership, and support.
A pilot project will have a better chance of becoming routine if the operational leader owns the outcome and the organization funds the conditions that make the work dependable. Employees should have a way to report the failure cases, someone should be able to fix the process, and the process of pausing or reversing the tool usage should be discussed. If the budget allows experimentation but not maintenance, review, oversight, or learning, the leaders have funded the demonstration, not adoption.
The question is whether the organization is ready to run the new work process without the project team. Who will maintain the data and connections? Who will monitor the performance when the model or the workflow changes? Who will update the training and procedures? These responsibilities should be assigned before expansion, not after the failure reveals the gap.
This is why implementation requires an operational leader with the authority to coordinate the necessary functions: the business owner is responsible for the outcome and process; technology teams maintain integrations and access; risk, legal, and security specialists define appropriate controls; frontline managers support the practice; and employees provide feedback on the behavior of the redesigned process. In the small organization, one person may perform multiple roles, but the responsibilities should be clear anyway. The project team may coordinate the process, but the operational part of the organization will run the service after the project ends.
The leaders should also budget the post-launch period. Models, data, suppliers, and business needs change. Someone should monitor the performance, resolve the incidents, update the guidance, and decide whether the change in the system requires new testing. If these activities do not have the time, funding, and designated owner, the organization relies on informal efforts. It may work in a supervised experiment; it is the weak foundation of the service that depends upon the customers or employees.
4. Keep a Philosophy, but Adapt the Method
The fourth lesson is to have a philosophy without being rigid. For AI, leaders should be firm about the result they pursue and flexible about the methods to achieve it. The pilot can demonstrate that the task, workflow, or tool did not meet expectations. That would be valuable evidence, not an invitation to defend the original plan.
Pilots therefore have to be designed to produce evidence that could alter the decision. The leader has to test routine and challenging cases, evaluate outputs against the existing process, and detect gaps in reliability or mere shifting of effort. The consultant’s experiment found positive and negative consequences of the AI adoption depending on the specific task. Leaders have to identify where AI works, where it does not, and what has to happen in that boundary case.
Before scaling the solution, leaders can define decision criteria: what quality, error rate, rework, and harm to the user has to be achieved, for which tasks, within which period, and who has the authority to expand. If the evidence does not confirm the initial decision, the choice may be to narrow the application domain, redesign the task, strengthen the controls, or even to abandon the idea. This does not mean that the decision maker failed to lead. It means that he or she led correctly.
Thresholds are most valuable when they are agreed upon before the results are known. Otherwise, a team that spent months on pilot testing could continuously redefine criteria of success so that the original decision is saved. Thresholds do not have to be a numerical cut-off criterion. It can combine the minimum quality requirements, the acceptable error rate, the proof of manageability of review work, and the conditions for escalation. The leader has to define who gets the information about the test results and who can order a break. A review cycle has to be established because performance on the first day of use may not be indicative of the performance under other loads, staff, and machine learning models.
5. Delegation of Authority Does not Mean Lack of Responsibility
The fifth lesson is that delegation of authority does not reflect poor leadership skills. In the case of AI adoption, delegation cannot mean unlimited authority of the system or the expectation that the manager checks every output. It means defining tasks that AI may perform independently, which tasks require review, and who has the right to override the decision.
An AI system that generates the internal summary has a different authority profile compared to the system that changes customer records or performs payments. When the potential impact of the decision increases, so too must there be more controls and more rights of human decision-makers. People need to know what the system may do, what decisions require approval, how to stop or reverse the action, and who is accountable for the decision.
One practical approach for setting boundaries of autonomous decision-making is to distinguish three steps: preparation of the information, recommendation of the decision, and execution of action. The system can prepare the draft to review by the employee, propose the response with reasons and supporting information or, in a very narrowly defined case, automatically perform a low-impact action. Every step provides more influence over the result. Leaders have to define which actions require approval, what events trigger the escalation, how employees can stop or reverse the action, and how incidents are logged. The level of control has to correspond to the consequences of the error and the possibility to intervene in time.
This approach helps employees to implement AI-based practices without guessing whether all outputs of AI need to be approved or whether they are personally responsible for each action. It defines a clear oversight process.
In Conclusion
Wenger’s approach allowed him to combine a philosophy with changes in training, player development, recruitment, and other elements of the environment of performance. It may serve as a good interpretive framework for the leadership of the adoption and implementation of AI: not only the tool, but also the system around it.
Luca Collina
