dooopSoftware · Organization · 5 min

How to Preserve Technical Judgment When Working with AI

Make technical decisions verifiable when working with agents. Use intention, alternatives, evidence, and limits to guide review and autonomy.

Published on September 6, 2026

CORE THESIS

Accepting a solution does not demonstrate that the team understood the decision.

Judgment appears when choices can be explained, challenged, and verified.

A team preserves technical judgment when it can explain why it accepted a change produced with AI, which alternatives it rejected, and how it verified the result. Asking someone to read the code before approval is insufficient if the person does not know the task’s intention or lacks time to investigate. Judgment needs to appear in observable decisions throughout the work, from problem definition to acceptance of the change.

Identify Decisions That Cannot Remain Implicit

Start with the concrete change. What behavior needs to change? What should remain? What constraint prevents the most obvious solution? These questions help locate where the team needs to understand the work, even when an agent performs much of it.

It is not necessary to require everyone to memorize every detail of each file. It is necessary to know who understands the relevant decisions and how that understanding can be examined. A change in business rules requires mastery of behavior. An integration change requires understanding the contract and possible failures. Reorganizing a component requires clarity about what will be preserved.

The DORA 2025 presentation describes AI as an amplifier of organizational strengths and weaknesses. The application proposed here is to observe whether the review process makes reasoning more visible or merely accelerates acceptance. The report does not prove that a specific ritual will solve this problem.

Request a Justification That Can Be Challenged

"It was the agent’s suggestion" does not explain a technical decision. A useful justification connects the choice to the task and allows another person to disagree based on something verifiable.

Propose a short record with four elements:

Intention: what behavior motivated the change.
Choice: which alternative was adopted and why.
Evidence: which test, reading, or observation supports acceptance.
Limit: what has not yet been verified and who needs to evaluate it.

This record should not become a mandatory essay on trivial details. Adjust its depth according to impact and ambiguity. A small visual adjustment may require a capture and a brief explanation. A change that modifies who can perform an action needs more explicit criteria and review by someone familiar with the rule.

Use the Agent to Explore Alternatives Without Outsourcing the Choice

One way to work is to ask for alternatives before requesting implementation. Another is to request objections to the already proposed solution. The value of these outputs depends on the ability to verify whether the alternatives respect the context.

The responsible person must distinguish a real constraint from a convention that can be reviewed. They also need to recognize when information is missing. If the agent proposes a solution based on an assumption about the user, the next activity may be to confirm that assumption with product rather than continuing programming.

This care does not require every task to become a formal experiment. The Microsoft ExP describes experimentation as part of development, with hypotheses, impact measurement, and iteration. Use this discipline when there is a hypothesis of improvement to test; for a known fix, verification may be reproducing the failure and confirming the expected behavior.

Fictional Example: A Simplification That Removes an Exception

Imagine a team maintaining a course organization system. An agent suggests simplifying the enrollment rule because two conditions seem redundant. The developer compares the proposal with the current code and realizes that one condition exists for classes with prerequisites still under review.

Judgment appears when they stop implementation, confirm the purpose of the exception, and choose how to represent it. They may keep the rule, reorganize it with a clearer name, or discuss its change with product. The agent can help in any path, but the existence of a short solution does not decide which is correct.

The task record should show the considered exception, the decision made, and the evidence used. The example does not demonstrate productivity gain. It illustrates what to observe to know if the person understood the change rather than just accepted a convincing answer.

Observe Understanding During Review

Review can use specific questions without turning the conversation into a memory test:

What case would make this solution fail?
Which premise, if wrong, would change the choice?
How would we know the change preserved previous behavior?
Which agent suggestion was rejected and why?
At what point does a person from another specialty need to participate?

If no one can answer, do not immediately conclude that individual competence is lacking. Context, time, documentation, or task definition may be missing. Investigate the obstacle before using the episode to classify the person.

The DORA on documentation highlights clarity, ease of location, and reliability. A decision record helps when another person can find and use it. An extensive justification hidden in a conversation does not fulfill this function by itself.

Expand Autonomy by Type of Decision

Evidence of understanding in a task does not authorize unrestricted autonomy. Define which decisions the person can make, in which situations they must consult a peer, and which changes require another approval.

Review these limits as new tasks and evidence appear. If the team always depends on the same person to explain a rule, the priority may be sharing context and practicing review together. If justifications become repetitive and no one reads them, reduce the ritual to what truly affects acceptance.

Preserving technical judgment requires making choices examinable. In the next AI-assisted change, request a justification with intention, alternative, evidence, and limit. Use the review conversation to verify understanding and define the next degree of autonomy.

If you want to discuss this decision in the context of your company, talk to dooop.

Sources

NEXT DECISION

Discuss Application in Your Company

Conversation about the software company context

Content by dooop. Registration allows relating this topic to the reader’s journey and tracking interest in the subject.