How to Disagree With an AI Recommendation

September 12, 2026 · By Michel Valenzuela, Founder of Daryl

The recommendation looks sensible.

It compares the options, explains the trade-offs, and ends with a clear next step.

Yet something feels wrong.

Perhaps it treats a deadline as flexible when it is firm. Perhaps it values short-term revenue more than you do. Perhaps the conclusion depends on a detail you never supplied.

That is a reason to examine the answer.

A useful AI conversation should leave room for disagreement. You should be able to question the reasoning, correct the context, and decide that the recommendation does not fit.

Find the point where the reasoning stops matching your situation

Start with the specific problem.

“I disagree” expresses your reaction. “This assumes I can move the client deadline, but I cannot” gives the assistant something concrete to reconsider.

Read the recommendation and ask yourself where it departs from what you know.

Has it misunderstood a fact? Overlooked a commitment? Introduced an assumption? Given too much weight to a consideration that matters less to you?

You might respond:

“You have treated the higher fee as the main advantage. My priority is protecting time for the product launch. Reassess the options using that priority.”

The aim is to improve the basis of the comparison.

Separate factual errors from differences in priorities

These require different corrections.

If the assistant has the wrong date, provide the right date. If it has misunderstood the available budget, clarify the amount.

A disagreement about priorities needs a different conversation.

Perhaps you are willing to accept slower growth to maintain independence. Perhaps the assistant favours a familiar approach while you want to develop a new skill.

Neither issue can be resolved by adding more detail about the same facts. You need to explain how those facts should influence the choice.

Try:

“The circumstances are correct. The weighting is wrong. Reliability matters more to me here than speed.”

That makes the disagreement easier to address.

Ask which assumptions support the conclusion

A recommendation can sound complete while relying on unstated assumptions.

For example, an assistant might recommend accepting a project because it assumes the scope is fixed, payment is dependable, and the work fits into your available hours.

If any of those assumptions is uncertain, the recommendation needs qualification.

Ask:

“Which assumptions does this conclusion depend on? Identify anything you inferred rather than anything I explicitly told you.”

Check the response against the original information.

An AI may not provide a complete account of how it arrived at an answer. What you need is an assessable explanation: the facts, criteria, and assumptions it presents as support.

That gives you something to verify.

Challenge the recommendation without supplying the answer you want

It is easy to steer a conversation toward confirmation.

You might say, “Surely declining is the smarter choice?” and receive a persuasive argument for declining.

That response may feel satisfying without improving the decision.

A more useful challenge is:

“Make the strongest reasonable case against this recommendation. Use the same facts and priorities.”

Or:

“What evidence would make another option preferable?”

Those questions invite examination without requiring agreement with your preferred conclusion.

You can then assess whether the alternative argument reveals a real weakness.

Bring your experience into the comparison

You may know something that is difficult to express as a neat criterion.

A client has repeatedly expanded the scope. A colleague follows through reliably. A project resembles work that has exhausted you before.

That experience can be relevant, but make it specific where possible.

Instead of saying, “I have a bad feeling about this,” explain what prompted it:

“On the last two projects, the client added substantial work after we agreed on the scope. That makes me less confident in the time estimate.”

Now the concern has a basis the conversation can examine.

You still need to distinguish a recurring pattern from an isolated incident. The point is to bring your judgment into view so it can inform the assessment.

Watch what changes after your correction

After you challenge the answer, inspect the revision.

Did the assistant update the relevant fact? Did it change the comparison in a way that follows from your clarification? Did it preserve other important constraints?

A revised answer can introduce new errors while fixing the old one.

It can also agree too readily.

If the conclusion reverses, ask what changed:

“Explain which part of my correction changed the recommendation and whether any uncertainty remains.”

A useful revision should have a reason you can understand.

Know when the conversation has reached its limit

You may correct the same misunderstanding several times and still receive an unsuitable recommendation.

At that point, another prompt may offer little benefit.

You can extract whatever was useful, return to the underlying evidence, and make the decision through another process. Depending on the situation, that could involve a colleague, a relevant specialist, or a direct test.

The assistant is one resource available to you.

Its fluency does not create an obligation to keep working toward agreement.

Keep the final judgment yours

You can reject a recommendation even when its reasoning is coherent.

It may represent a trade-off you are unwilling to make. It may fit the stated criteria while missing something you have not yet articulated. You may need more time or another perspective.

What matters is understanding your reason well enough to act responsibly.

A productive conversation can help you get there. Agreement is only one possible outcome.

Why disagreement belongs in Daryl

Daryl is a personal AI for Mac, currently in development, designed around local conversation and support for decisions.

I want the experience to make correction and reconsideration straightforward. People should be able to question an assessment, revise the context, and retain control over the choice.

These are principles guiding the product’s development, rather than claims that every behaviour described here is already available.

An assistant earns its usefulness through the quality of the discussion.

Sometimes that discussion ends with a recommendation you accept.

Sometimes it helps you explain precisely why you are choosing another direction.