Instrument
It executes a delimited, reversible operation subject to prior human judgement.
Case studies · Professionals
Perform more without stopping know how to do
Assisted productivity and professional competence are not the same. The question is what do you need to continue understanding, checking and executing when the AI response is incomplete or erroneous. This case study proposes analyzing real tasks, not entire professions, and combining immediate profit with a measure of capacity, traceability and response to failure. The case is organized around tasks and concrete responsibilities, not professional labels.
It does not classify professions. The level of control needed depends on the task, the risk, and your actual supervisory ability.
Indicative overview
Professional control
Consulting · law · management
69/100Knowledge work
Speed increases easily; the challenge is to maintain formulation, contrast and responsibility for the recommendation.
◎
Control principleIf you can't verify, you're not supervising
Counterpoint
Look for objections, omitted assumptions and alternatives after your own formulation.
Tutor
Offer examples, questions or feedback that is reduced when competence increases.
Substitute
It assumes full capability and requires deciding what knowledge will remain available.
The fundamental distinction
A good result today does not prove that you will retain the ability tomorrow.
AI can improve speed and quality in specific tasks, and that improvement can be valuable. But the product belongs to the joint person-tool system: it does not allow us to infer, by itself, how much the person has learned, what errors he will detect without help or how he will respond to an exception. To know the available capacity you have to withdraw assistance, vary the case and observe the reasoning.
Performance
Quality, speed, volume or reach achieved by the joint human-AI system. It should be compared to a baseline and to the cost of review, correction and integration.
Seen on the productAbility
What you can understand, transfer and execute when the context changes. It includes recognizing a poorly formulated case and justifying why a solution is appropriate.
It is checked without helpResilience
What you do when the tool crashes, disagrees, changes, or finds a new case. It requires alternative procedures, practice and authority to stop the process.
It is tested with incidentsGuidance evaluator
The unit of analysis is not your profession. It is the specific task.
Evaluate a real operation and define its beginning and end. The result does not measure your competence or certify safety: it organizes four questions that are usually separated—impact, verifiability, importance of skill, and frequency of practice—to decide how much control, contrast, and independent execution should be retained.
Heuristic model, not a validated scale. The score weighs impact, difficulty of verification, need to maintain capacity, and lack of practice. It does not replace professional, legal or security protocols.
Professional capital
Supervising requires expertise that automation itself can leave unpracticed.
Surveillance is not passive nor is it resolved by asking someone to “check.” To detect a plausible but incorrect answer you need domain knowledge, mental models, time, independence and continued contact with real cases. If automation removes precisely those cases, it can weaken the conditions necessary to supervise it.
Formulate the problem
Decide which question deserves an answer, with what limits and for whom.
Make a hypothesis before generating.Domain knowledge
Recognize concepts, restrictions and relationships that do not appear explicit.
Recover fundamentals without consulting.Detect anomalies
Perceive the figure, assumption or pattern that does not fit the case.
Introduces deliberate errors into simulations.Contextual judgment
Integrate values, consequences, uncertainty and tacit knowledge.
Write the final human justification.Respond to the bug
Continue safely when the tool is missing or makes a mistake.
Rehearse contingencies without assistance.Conceptual trajectories
Assisted performance and available capacity
Assisted performance Maintained capacity Ability without practice
The descent of the dashed line represents a loss of practice hypothesis, not an estimate of magnitude or speed.
irregular border
Similar tasks may fall outside the capabilities of AI.
Reliability should not be inferred from the fluency, tone, or level of detail of the response. It is calibrated with representative cases, external criteria and sufficient knowledge to detect when the system has left its competent zone. The boundary also changes with the model, the data, the language, the context, and the exact way of formulating the task.
→
Rule of thumb The more difficult it is to recognize the error, the less sense it makes to call a superficial review "supervision."
Easier to verify →Greater impact →01Test and calibrate 02Use as an instrument 03Stop or escalate 04Double control
New task or new model, but with verifiable results.
Compare with real casesGood observed performance and reversible or verifiable output.
Sample reviewHigh impact and you don't know how to distinguish a correct answer from a plausible one.
Do not delegate the decisionThe AI seems competent, but the cost of error requires an independent path.
Separate analysis and decisionIndividual differences
The recommendation changes with actual expertise, not just age or title.
Years of experience and formal training provide insight, but the relevant capability is specific to each task: formulate, execute, verify, and respond to exceptions. Expertise is not fixed either; it may grow with deliberate practice, become outdated by changing context, or fail to transfer to a nearby domain.
Autonomy
Consolidation professional
It already solves common cases and needs to expand its scope without freezing its development.
- + Chance
- Contrast of alternatives, documentation and acceleration of reversible tasks.
- − Risk
- Convert a still fragile competence into a permanently delegated operation.
- → Protocol
- Alternate assisted cases with complete cases and compare the transfer.
From performance to maintenance
A protocol for working with AI and still being capable without it.
Select a situation and always use the same case when comparing. The sequence preserves checkpoints before, during and after assistance, and adds a contingency test. There is no need to repeat all the work without AI: just isolate the critical operation and check that it is still available.
Practice case
Prepare a professional recommendation
Define the recipient, the decision the output must support and three quality criteria.
Ask for objections, alternatives, and omitted assumptions; not a closed conclusion.
Contrast the relevant statements and write the final recommendation.
Explain the case and decide with a blank sheet of paper and original sources.
Start an important task before asking for help.
Practice the essential procedure with the missing or wrong tool.
Review which operations you no longer execute and whether you still need to keep them.
Functional signs
When should you review your way of working?
They are not clinical symptoms nor do they prove loss of capacity. They are observable changes that justify recovering practice, contrast or delegation limits. Before concluding, check to see if the cause may be time pressure, a poorly designed process, lack of access to sources, or an organizational expectation that rewards speed over understanding.
You don't know how to start a common task without first opening the wizard.
The revision is limited to style because you cannot reconstruct the reasoning.
You accept a recommendation when it sounds professional, even if you can't verify it.
You have stopped performing an operation that you would need during an incident.
Your productivity goes up, but your confidence in working without AI goes down.
You avoid new or ambiguous cases if you can't immediately query the tool.
!
Longitudinal research is still limited. There is evidence of productivity improvements and errors of overdependence on specific tasks; it does not allow us to affirm that using AI causes a general cognitive deterioration in professionals.
Evidence base
What we know and what we don't.
The results depend on the task, the model, the level of experience, the way of use and the comparison criterion. Each card summarizes a specific scope and should be read as a piece, not as a universal conclusion. The evidence on productivity is more abundant than long-term tracking of expertise and transfer.
Productivity · experimentIn professional writing, AI reduced time and raised average quality.Noy & Zhang · Science, 2023 ↗Consulting · experimentImprovements were concentrated within an irregular technological frontier.Dell'Acqua et al. · Organization Science ↗Critical thinking surveyMore trust in AI was associated with lower perceived critical activation.Lee et al. · CHI, 2025 ↗Collaboration · meta-analysisCombining human and AI does not guarantee surpassing the best of both separately.Vaccaro et al. · Nature Human Behavior ↗High impact · experimentRadiologists followed correct and incorrect suggestions of the system.Rezazade Mehrizi et al. · Scientific Reports ↗