The biggest mistake in AI training happens before the course begins. Managers approve training without understanding the technical work their teams do every day. Without that understanding, they cannot properly review the process or adapt it to AI.
I would start by closing that distance between management and the people doing the work. Otherwise, we are asking training to compensate for a process nobody has taken the trouble to examine.
Some of what we find may be uncomfortable. A step described as a decision may amount to forwarding an answer to someone else. An approval may add a day of waiting without changing the work or catching a problem. I think of these as pseudo-decisions: they have the appearance of control, but it is hard to explain what judgment they contribute.
Put an AI assistant into that process and it may finish its part sooner. The work can still sit in the same queue afterward. What, exactly, should we train the employee to do about that?
This is why I want process review to come first. Follow a piece of work through the operation with someone who actually does it. Look at where it waits and what each handoff requires. For every approval, ask the person responsible to explain what they evaluate and what would make them reject the result. If the answer remains vague, that checkpoint deserves examination before it becomes a requirement in the new AI workflow.
My background at Nuance, Nissan, Santander, and other companies includes DevOps automation and CI/CD, along with documentation and monitoring. That is the operational perspective I bring to AI adoption. A tool becomes part of a sequence of work, and the value of speeding up one step depends on what happens around it. Experience at Nuance.
At Foursys, I also develop courses and workshops for technical professionals and business managers. The distinction between those audiences matters here. The developer may need to implement an integration. The manager may need to decide whether an approval should exist at all. We cannot leave that second responsibility out of the learning plan and expect technical training to resolve it. Professional background.
Gartner’s discussion on skills strategy gives us a way to choose where to begin. The presenters recommend focusing on one to five capabilities that will distinguish the business, then identifying roughly ten to twenty critical skills. I would use that focus to select a process worth improving and examine it with the people who know its details.
That application also connects with my MIT xPRO course, Inovação Tecnológica: da Estratégia à Aplicação, which covers innovation in processes, products, and business models. For me, the useful question here is what should change in the work itself. The course selection follows that decision. Program · My certificate · LinkedIn credentials.
Start with a business priority, identify the capabilities and skills it requires, and define the evidence of progress.
Imagine a support team that uses an assistant to prepare customer responses. Every response then waits for a manager’s approval. Before teaching everyone to write better prompts, I want to understand that queue. Perhaps the manager catches exceptions that employees cannot yet recognize. Perhaps routine cases wait alongside those exceptions because nobody has defined a different route for them.
Those situations need different changes. In the first, the review reveals judgment worth teaching. In the second, the team may need clearer rules and authority to proceed. Keep a human checkpoint when the decision requires it, and make its purpose explicit. Remove unnecessary bureaucracy when the review shows that it adds no useful decision.
The distinction becomes clearer with a difficult case. Suppose an assistant produces a convincing response that depends on a condition the customer does not meet. Ask the employee to find the problem and explain the next step. That exercise helps establish whether the employee can handle the exception, and gives us a concrete reason for whatever review remains.
Good AI-assisted work includes checking evidence, challenging assumptions, and deciding what happens next.
There is evidence for taking this seriously. An experiment involving 758 BCG consultants found that GPT-4 helped on tasks within its tested capabilities. On a separate task outside that boundary, AI users were about 19 percentage points less likely to reach the correct solution. The researchers tested GPT-4 in 2023. Our tools and tasks need their own evaluation; that result does not justify adding the same approval step to every workflow. Dell’Acqua et al., 2026.
My University of Minnesota course, Engineering Practices for Building Quality Software, covered design and architecture through implementation, testing, and deployment. I bring that engineering perspective to the question of control: what do we need to observe before we can trust this particular work? The answer should help us design the check and the training, rather than default to another signature. Syllabus · My certificate · LinkedIn credentials.
The technical team belongs in this discussion throughout. An experienced employee can show a manager why an apparently simple case took time. A manager can clarify an unresolved responsibility. Reviewing the work together also creates an opportunity to teach: compare the assistant’s draft with the final response and explain the correction.
Managers and technical teams review actual work and explain the decisions that matter.
Gartner highlighted social learning in employees’ perceptions of their preparedness. A separate study measured output during an AI rollout involving 5,172 support agents. Brynjolfsson, Li, and Raymond found a 15% average increase in issues resolved per hour, with gains concentrated among less experienced and lower-skilled workers. Performance during tool outages also provided evidence of learning. These are results from one firm’s deployment, not an estimate of what our redesigned process will achieve. Generative AI at Work, 2025.
For our team, I would measure the time a case spends waiting as well as the time spent working on it. Then examine reopened cases and reviewer corrections. That lets us question an apparent improvement: did the work actually reach the customer sooner, and did we preserve its quality? Staffing and case difficulty can change the figures too, so those need attention before attributing a result to AI or training.
Assess learning separately. In a controlled exercise without the assistant, ask employees to explain their decisions. The live workflow shows what they accomplish with support; the exercise helps reveal what they can demonstrate independently. Those observations can inform how much authority they are ready to take on.
I also want the reasoning behind the revised process to remain accessible. In Agent Memory, decisions, lessons, and handoffs persist in Markdown, with context retrieval and session checkpoints to support continuity. For this support team, the useful record might be why a routine case can proceed directly and which condition requires escalation. That explanation will matter when someone proposes adding the approval back. Project documentation · My account of the project.
Gartner’s “skills promise” idea puts a related obligation on hiring managers. It considers the ability and willingness to learn from a minimum viable foundation for the role. Essential skills still need assessment, and someone must provide assignments, feedback, and time to develop. Hiring for potential while leaving people stuck in an unrevised process gives that potential little room to become useful.
Before approving the next AI training plan, sit with the technical team and follow the work. Decide which steps should change, which decisions people really need to make, and what they will need to learn to do that well. That is where I would spend the first part of the training budget.
Sources and further reading
Gartner. Develop a Skills Strategy That Accelerates Organizational Performance. Webinar informing this discussion.
Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889-942.
Dell’Acqua, F., et al. (2026). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Organization Science, 37(2), 403-423.
MIT xPRO. Inovação Tecnológica: da Estratégia à Aplicação. Program · My certificate · LinkedIn record.
University of Minnesota. Engineering Practices for Building Quality Software. Syllabus · My certificate · LinkedIn record.
Mario Mayerle Filho. Professional experience and Agent Memory project.
If your team is planning AI training, contact INOSX to discuss the process you want to improve, including where the work waits for a decision.
For agent-based delegation, explore INOSX AgentOS and register your interest in the closed beta. You can read more about my work and connect with me on LinkedIn.
#AISkills #ProcessImprovement #BusinessStrategy #OrganizationalLearning #INOSX #AgentOS