Trust and quality notes
- Last updated
- August 28, 2026
When a machine can produce an answer in seconds, getting an answer stops being the main achievement. The harder work moves upstream: noticing what matters, deciding what to investigate, and recognizing when the original question no longer fits the evidence.
That is the future Anne-Laure Le Cunff describes in her original statement.[1] Her argument is not that people should compete with AI by writing more elaborate prompts. It is that cheap answers increase the value of human curiosity, judgment, and the willingness to revise a mistaken frame.
For anyone beginning to use AI at work, this is a practical shift. The useful question is no longer only whether the task can be automated. It is also what the team is trying to learn and what evidence would change its mind.
Answers become abundant
Le Cunff starts with a simple change in supply. AI can propose solutions and act on them quickly, so answers are becoming plentiful. When something once scarce becomes common, its value falls relative to what remains difficult.
In her account, the scarce ability is deciding where attention belongs. A system can respond to a question, but a person still has to notice the situation, choose the question, and judge whether the response addresses the real issue. People also have to detect when reality contradicts the frame they began with.
This distinction matters because a polished answer can still be useless. If a team asks how to speed up a process that should not exist, faster execution only hides the bad premise. If it asks how to improve a plan whose assumptions have not been tested, a detailed response can create confidence without learning. These are interpretations of Le Cunff’s argument, not examples she gives in the source. They show why question selection becomes consequential when answer production is cheap.
Her point also places limits on prompt technique. Better instructions can improve an output, but prompting does not decide what deserves attention. Observation and judgment happen before the request is written and after the result appears.
Work becomes a series of experiments
Le Cunff expects more work to take the form of curiosity-driven experiments. As agents reduce the effort required for research and prototyping, it becomes harder to justify spending months perfecting a plan before checking its assumptions.
She assigns distinct responsibilities to people and AI. Humans define the question, decide what evidence could alter their view, and set the experiment’s boundaries. AI compresses the work needed to run it. Humans then interpret the result and choose the next question.[1]
This is more demanding than telling a tool to try something. A real experiment needs a claim that can be tested, boundaries that prevent the exercise from expanding without limit, and evidence that could produce a meaningful update. It also requires the discipline to look at an unwelcome result rather than defend the original idea.
The source describes this as a change in the basic unit of progress. Instead of treating completion of a predetermined task as the main sign of movement, a team can treat an experiment that expands the available possibilities as progress. Completion still matters, but it is no longer enough on its own. The work must improve the team’s understanding.
My interpretation is that this rewards shorter loops between a question and contact with reality. The point is not speed for its own sake. A fast loop is valuable when it exposes assumptions earlier and gives people a chance to redirect effort before a plan hardens into a commitment.
Judgment moves to the center
Automation does not remove human responsibility in Le Cunff’s picture. It concentrates it. A person must decide whether the evidence is relevant, whether the experiment was well designed, and whether the outcome justifies a new direction.
That responsibility cannot be handed off merely because a system generated the research or prototype. The output may be quick and convincing, but interpretation remains tied to context. What counted as meaningful evidence? Which constraint was excluded? Did the result answer the intended question, or only an easier version of it?
These questions are not objections to using AI. They are part of using it well. Le Cunff explicitly argues that AI can compress the execution of an experiment. Her concern is where people place their effort once that compression is available.
For a reader new to AI, a modest practice follows from this argument. Before asking for an answer, write down the decision the answer is meant to support. Name the evidence that would make you reconsider. After receiving the output, ask what remains uncertain and which new question appeared. This practice is my application of her thesis, not a process prescribed in the source.
Organizations must reward learning
Le Cunff also points to an organizational obstacle. Many workplaces reward accurate prediction and efficient execution. Those habits can make revision look like failure, especially after a plan has been presented with confidence.
Her alternative is an environment where people can question assumptions, test other approaches, and stop treating success and failure as the only possible outcomes. In that setting, the relevant measure is whether the work produced learning. As she puts it, “the only real failure is failing to learn.”[1]
That statement does not make every unsuccessful attempt useful. Learning has to be made explicit. A poorly bounded test, an ignored result, or a repeated mistake can still waste effort. The source’s standard is stronger: people should run experiments that change what they understand and use that understanding to choose what comes next.
Leaders therefore have work to do beyond buying tools. They need to make it safe to surface a wrong assumption, stop a plan, or ask a more basic question. They also need to distinguish thoughtful experimentation from undirected activity. This is an interpretation of the conditions required by Le Cunff’s argument.
The advantage is better inquiry
Le Cunff’s conclusion is restrained. The strongest teams will not necessarily be those that automate the largest amount of work. They will be those that ask the best questions.[1]
That changes the purpose of automation. It is not simply a way to clear a queue faster. It can create room for more careful attention, more tests, and more frequent revision. But none of those benefits happen automatically. People must use the saved effort to improve the inquiry rather than produce a larger volume of unexamined answers.
The practical lesson is to begin with curiosity and end with judgment. Use AI to reduce the cost of finding, making, and testing. Keep people responsible for deciding what is worth testing, what the evidence means, and where attention should move next.
If you want to turn one repeated workflow into a well-bounded learning loop, start with Agentic Workers.