How Do You Know Whether an AI Recommendation Is Any Good?
· By Michel Valenzuela, Founder of Daryl
A useful AI recommendation should fit your actual options, respect your constraints, and explain what could change the answer.
Fluent writing is not enough. Neither is confidence.
Before acting, check whether the assistant understood the decision you need to make and whether its reasoning holds up against what you know.
You can do that without a technical background.
Did it understand the decision?
Start by asking the assistant to restate the question.
Suppose you are deciding whether to accept a project. You need to know whether you can deliver it alongside existing commitments.
An answer about how attractive the opportunity sounds may miss the decision entirely.
A better starting point is:
“Before recommending anything, state the decision I am making, the options available, and the constraints I gave you.”
Read that summary carefully. If the assistant has misunderstood the problem, correct it before continuing.
Did it respect your firm constraints?
A firm constraint determines which options are viable.
If you cannot move the deadline, an option that requires more time does not solve the problem. If a project requires skills you do not have, enthusiasm alone does not make it deliverable.
Separate requirements from preferences.
| Type | Example | How it should affect the recommendation |
|---|---|---|
| Firm constraint | Delivery must happen by Friday | Options must fit that deadline |
| Preference | I would rather work with the existing supplier | Consider it alongside other tradeoffs |
| Unknown | I do not know whether the supplier has capacity | Identify what needs to be checked |
| Assumption | The client might accept a smaller scope | Keep it conditional until confirmed |
This gives you a practical way to inspect the answer.
An assistant should not quietly treat a preference as mandatory or an assumption as a fact.
What evidence supports the recommendation?
Ask which information the answer relies on.
Some inputs come directly from you. Others may come from documents or external sources. Some may be the assistant’s inference.
Those distinctions should be visible.
For example:
“Option A is cheaper” needs a basis. Does the quoted price include the same work? Are there additional charges? Is the figure current?
“Option B is easier to deliver” also needs a basis. Does it require fewer people, a shorter approval process, or less unfamiliar work?
Ask:
“Separate the facts supporting this recommendation from the assumptions.”
That makes the answer easier to assess.
Does it explain the tradeoff?
Most worthwhile decisions involve giving something up.
An option may be faster but less flexible. Another may provide more control while demanding more of your time.
A useful recommendation should explain that cost.
Suppose the assistant recommends keeping a project’s original scope. The tradeoff might be disappointing the client’s new request while protecting the agreed delivery date.
You need to see both sides to decide whether the recommendation fits your priorities.
Ask:
“What do I gain by choosing this, and what do I give up?”
A list of benefits alone is incomplete.
What would change the answer?
This is one of the most useful questions you can ask.
A recommendation may depend heavily on one uncertain detail.
If a supplier can deliver by Thursday, one option makes sense. If delivery takes another week, a different option may be preferable.
Ask the assistant:
“Which missing fact could change your recommendation?”
Then:
“What is the simplest way to check it?”
You may discover that the next useful step is a short conversation, a written confirmation, or a small test.
That is often more valuable than another page of analysis.
Is it overlooking a better alternative?
An assistant works from the options and context it receives. If you give it two choices, it may focus on choosing between them.
There may be another approach: reduce the scope, postpone one part, run a limited trial, or decline both options.
Ask:
“Are these the right options to compare? Is there another realistic approach?”
The word “realistic” matters. An alternative is useful only if it fits your circumstances.
You are looking for an overlooked possibility that you could actually pursue.
Does a percentage make the recommendation more reliable?
A number needs an explanation.
An assistant saying “85% confident” does not establish an 85% chance that the decision will succeed.
You need to know what the number represents and how it was produced. It might express a comparison between options, a model score, or an unsupported estimate.
If that explanation is missing, assess the recommendation through its evidence and assumptions.
Precision in presentation should not substitute for a clear basis.
What does a practical review look like?
Consider a small agency deciding whether to accept an additional deliverable.
The client wants it within the existing deadline. The team has limited capacity.
You can review the recommendation with five questions:
- Has it recognized that the deadline is fixed?
- Has it accounted for existing commitments?
- Has it distinguished confirmed capacity from assumed capacity?
- Has it explained what accepting the work would displace?
- Has it identified what to confirm before replying?
If those questions expose gaps, revise the analysis.
The goal is an answer you can inspect and use, including when you disagree with it.
How does this relate to Daryl?
Decision assistance is central to what we are building with Daryl.
Daryl is being developed as a personal AI for Mac, with local core conversations, voice and typing, and help comparing possible next steps.
The intended value is making a decision easier to examine: what matters, what is uncertain, and what needs attention before you proceed.
Daryl is still in development. Its behavior and launch capabilities need to be evaluated through testing.
A recommendation should support your judgment. You remain responsible for deciding whether it fits the situation.
Frequently asked questions
Should I ask the same question several times?
Repeating it may produce different answers without resolving the uncertainty. First identify which fact or assumption needs examination.
Can I ask the AI to challenge its recommendation?
Yes. Ask for the strongest practical reason against it and the circumstances under which another option would be better.
Is agreement with my preference a good sign?
Only if the reasoning supports it. Check whether the assistant has examined the costs and constraints rather than simply reinforcing your initial view.
Can a local model still make mistakes?
Yes. Processing location does not guarantee answer quality.
Make the answer inspectable
Before acting, check the decision, constraints, evidence, tradeoff, and uncertainty.
If you can see those clearly, you have something useful to work with.
That is the kind of assistance we want Daryl to provide: a clearer basis for your next step.
Meet Daryl and get the launch announcement