Working With AI Without Losing Professional Expertise

In the previous post, I argued that graduates need to protect the information entrusted to them before providing it to an AI system. But protecting data is not the only capability that can be weakened when AI becomes embedded in professional work.

There is another risk that may be less visible: What happens when AI helps people perform work they have not yet learned how to do? This question is especially important for students and early-career professionals.

Experienced professionals often approach AI with knowledge developed through years of practice. They have encountered unusual cases, made mistakes, received feedback, recognized patterns, and learned where apparently simple problems become complicated. When they use AI, they can compare its output with an existing mental model of the work. They may recognize that a recommendation violates a professional standard. They may notice that an important variable is missing. They may see that the explanation sounds plausible but does not fit the situation. They may know which questions the system should have asked before offering an answer.

Novices do not yet have that same foundation. They may be able to use AI to produce work that resembles the output of an experienced professional. But resembling expertise is not the same as possessing it. This creates an important tension. AI may help less-experienced workers perform at a higher level while simultaneously reducing their opportunities to develop the expertise needed to perform, evaluate, and supervise that work independently.

In the short term, the result can look like progress. The report is completed. The code runs. The analysis appears polished. The customer receives an answer. The presentation looks professional. The task is finished more quickly than it would have been without assistance. But task completion does not tell us what the person learned through the process.

That distinction has appeared throughout this collection of posts in relation to education. In the workplace, however, the consequences become even more significant. A student who completes an assignment without developing the intended skill may struggle on a later assessment. A professional who lacks the necessary expertise may fail to recognize a problem affecting a customer, patient, client, colleague, system, or organization.

Productivity and capability are not the same thing.

Research has already demonstrated that generative AI can improve workplace performance. A large study of customer-support agents found that AI assistance increased productivity, with the largest gains among less-experienced workers. The system helped newer employees communicate in ways that resembled the organization’s more successful agents. (Brynjolfsson, Li, and Raymond, 2023)

That is a meaningful benefit. AI can make expert knowledge more accessible and help novices perform tasks that would otherwise remain beyond their current abilities. But we should be careful about what we conclude from the improved performance. Did the less-experienced worker become more capable? Or did the combination of the worker and the AI system become more capable? Those are not necessarily the same outcome.

The distinction becomes visible when the tool is removed, when the situation falls outside the AI system’s capabilities, or when the output contains a subtle error. Can the employee continue? Can they reconstruct the reasoning? Can they diagnose the failure? Can they adapt the process? Can they determine whether the recommendation is appropriate? If not, the AI may have extended the person’s performance without developing their underlying expertise.

This does not mean the assistance had no value. We routinely use tools to extend human capability. Calculators, search engines, diagnostic equipment, spellcheckers, and software libraries allow professionals to accomplish more than they could unaided. Professional expertise has never meant performing every task without tools.

The concern is not that AI assists the work. The concern is whether the way it assists removes the experiences through which people learn to understand the work.

A 2026 randomized study examining AI and software skill formation illustrates this risk. Participants learned to complete tasks using an unfamiliar programming library. Those with access to AI performed worse afterward on measures of conceptual understanding, code reading, and debugging. Their evaluation scores were 17 percent lower on average, while the researchers found no statistically significant improvement in overall completion time. (Shen and Tamkin, 2026)

The most interesting finding was not simply that AI use weakened learning. Different forms of AI use produced different outcomes. Participants who relied heavily on AI-generated code developed less understanding. Those who used AI to request explanations, clarify concepts, or support their own reasoning were more likely to preserve learning.

The important variable was not merely whether AI was present. It was whether the person remained cognitively engaged with the task. That finding aligns with the distinction I explored in the previous series between AI as a tutor and AI as an answer machine. In professional settings, however, another layer must be added.

The employee may not be trying to learn. They may simply be trying to finish the work. Workplaces reward productivity. Deadlines are real. Customers are waiting. Organizations may measure how many cases were resolved, how many documents were produced, or how quickly projects moved forward.

Under those conditions, using AI to complete the task as efficiently as possible can be entirely rational. But if every learning opportunity is optimized for immediate production, where will future expertise come from? Professional expertise is often built through work that initially appears inefficient. A junior technician learns by diagnosing problems that an experienced technician could solve more quickly. A new accountant learns by tracing discrepancies that software might identify instantly. A beginning programmer learns by encountering errors, reading documentation, testing possibilities, and debugging unsuccessful attempts. A new instructor learns by designing activities, observing how students respond, and revising the approach. A healthcare professional develops pattern recognition by seeing cases, discussing decisions, and comparing initial interpretations with outcomes. These activities do more than produce work. They help novices build mental models of the field.

The struggle is not valuable simply because it is difficult. Difficulty alone does not guarantee learning. Some tasks are unnecessarily repetitive, poorly supported, or disconnected from the capabilities professionals actually need. But some forms of effort provide information. An error reveals how a system works. An unsuccessful approach exposes a flawed assumption. A difficult decision forces the learner to distinguish between competing priorities. Feedback helps connect an action with its consequences. Repetition allows patterns to become recognizable. When AI removes these experiences too early, it may remove part of the pathway through which professional judgment develops.

Recent qualitative research in software engineering raises this concern directly. Interviews with junior and senior developers suggested that generative AI is beginning to absorb entry-level tasks that previously gave novice developers opportunities to build expertise. The researchers argue that AI may be changing not only which tasks juniors perform, but also the developmental pathway through which juniors eventually become seniors. (Yu and Moon, 2026)

Although that research focused on software development and involved a relatively small interview sample, the underlying question applies across fields: If AI performs the work traditionally assigned to beginners, how will beginners become experts?

This is not an argument for preserving every entry-level task exactly as it currently exists. Some routine work can and should be automated. We should not require people to spend years completing low-value tasks merely because previous generations did.

But when we automate a task, we need to ask what else the task was doing. Was it building familiarity with common cases? Was it teaching the structure of a process? Was it helping beginners recognize abnormal situations? Was it exposing them to the consequences of small decisions? Was it creating opportunities for feedback from experienced colleagues? Was it developing knowledge they would later need to supervise automated systems? If so, removing the task without replacing its developmental function creates a gap.

The organization may become more efficient today while weakening its supply of expertise for tomorrow. This should matter to higher education because colleges may be one of the few places where students still have structured opportunities to develop foundational capabilities before workplace efficiency becomes the dominant expectation.

But we face the same pressure. If AI can complete a task, asking students to perform it without AI can seem outdated or artificial. Students may reasonably ask why they should learn to do something that software can already do faster. The answer cannot simply be, “Because you might not always have the tool.” That is possible, but it is not the strongest reason.

Students need foundational knowledge because using AI well often depends on understanding the work well enough to direct, evaluate, and correct it. A person cannot meaningfully supervise a process they do not understand. They cannot recognize an exception if they have never learned the normal pattern. They cannot diagnose a failure if they have never developed a model of how the system should work. They cannot determine whether an AI recommendation is appropriate if they do not understand the professional standards, concepts, and context that should guide the decision.

The more work we delegate to AI, the more important it becomes to identify the expertise humans must retain. That expertise will vary by profession. A professional may no longer need to perform every calculation manually, but they may need to understand what the result means and recognize when it is unreasonable. A technician may use AI to suggest troubleshooting steps, but they still need to understand the system well enough to test those suggestions safely. A writer may use AI to generate alternatives, but they still need to make decisions about audience, evidence, meaning, and voice. A healthcare professional may use AI to identify possibilities, but they must understand the patient, the limits of the evidence, and the consequences of acting. An instructor may use AI to generate materials, but they still need to recognize whether those materials support learning.

Preparing students for an AI-enabled profession therefore requires distinguishing between performance-supporting tools and learning-supporting uses of those tools. Sometimes the goal is to produce the best possible result efficiently. In those situations, substantial AI assistance may be appropriate. At other times, the goal is to develop a capability the student or early-career professional will later need. In those situations, unrestricted AI assistance may undermine the purpose of the experience.

This is not a permanent division.

A task that should be completed independently while someone is developing foundational knowledge may later become an appropriate task to automate. The question is not whether the person should ever delegate it. The question is when. Have they developed enough understanding to recognize a poor result? Can they explain the reasoning behind the process? Can they perform the most important parts without assistance? Do they understand which parts of the task are routine and which require judgment? Can they recover when the system fails? These questions move us beyond the false choice between banning AI and allowing it to do everything.

Students can work with AI while continuing to build expertise, but the learning experience must be designed intentionally. They might attempt a problem before viewing an AI recommendation. They might use AI to ask for explanations rather than completed work. They might compare their approach with the system’s approach. They might diagnose errors in AI-generated work. They might complete important parts of a process independently before deciding which steps to automate. They might be asked to reproduce or adapt the work when the tool is unavailable. Most importantly, they should understand why certain capabilities are being protected.

If students experience restrictions only as rules designed to control AI use, they may reasonably focus on finding ways around them. If they understand that the goal is to preserve expertise they will later need, the conversation changes.

The question becomes: What do I need to remain capable of doing, even if AI usually helps me do it?

That may be one of the defining questions of professional preparation in an AI world. We do not need graduates who refuse assistance in order to prove their independence. We also do not need graduates whose competence disappears when the assistance does. We need graduates who can use AI to extend their expertise without allowing it to replace the process through which expertise develops.

That balance will become more difficult as AI systems become more capable. It will become harder to tell when a person understands the work and when the system is carrying the understanding for them.

And it leads to another question: Even when a professional retains expertise, how do we determine which situations require meaningful human oversight?

That is where the next post will turn.

Continuing the Conversation

Series 1: AI Is Exposing Existing Problems✓ Completed
Series 2: What We Do About It✓ Completed
Series 3: Cultivating Human Thinking in an AI World✓ Completed
Series 4: Learning Alongside AI✓ Completed
Series 5: Preparing Students for an AI World
Current Post (5 of 8): Working With AI Without Losing Professional Expertise
Next Up: Knowing When Human Oversight Is Necessary