keepcalculator

06 Numbers and Powers

Why a Planning Agent Must Ask Before It Picks

Card No. 06·79179 · Updated · Standard formula, shown below

A planning agent must ask before it picks because choosing an action without confirming the user’s current constraints, preferences, or intent can commit resources, violate a hard limit, or produce a result the user did not want. Asking first converts an assumption into a verified fact, which is the only safe basis for a plan that will actually be executed.

Why Asking Matters: The Core Problem

A planning agent (whether a software scheduler, an AI assistant, or a human project manager) operates under incomplete information. The user’s real constraints—deadlines, resource limits, personal preferences, or hidden dependencies—are rarely fully specified in the initial request. If the agent picks an action without asking, it risks:

  • Committing to a path that violates a hard constraint (e.g., a task that must be done by noon, but the plan schedules it at 3 PM).
  • Using a resource that is actually unavailable (e.g., a meeting room already booked, a person on leave).
  • Choosing an option that is technically valid but contradicts the user’s unspoken preference (e.g., scheduling a workout at 6 AM when the user is not a morning person).

Asking is not about being polite; it is about reducing the probability of a costly wrong choice. In decision theory, the expected cost of a wrong pick is the probability of the pick being wrong multiplied by the cost of that error. Asking eliminates most of that probability, making it the rational move when the cost of asking is low and the cost of a wrong pick is high.

How a Planning Agent Decides When to Ask

Not every decision requires a question. An agent should ask when the following conditions hold:

  1. Uncertainty is high: The agent has multiple plausible options and no data to distinguish them.
  2. Irreversibility is high: Once the action is taken, it cannot be easily undone (e.g., sending an email, purchasing a non-refundable ticket).
  3. Impact is significant: The action affects other tasks, people, or resources in a way that is hard to reverse.
  4. User preference is unknown: The agent has no history or stated rule that indicates the user’s choice.

Conversely, if the agent has a clear rule (e.g., “always book the cheapest flight”), or if the choice is trivial and reversible (e.g., which color to use for a label), it can act without asking. The key is to balance the cost of asking (user time, friction) against the expected cost of a wrong pick.

Common Mistakes When an Agent Doesn’t Ask

Real-world examples illustrate why skipping the question leads to failure:

  • Calendar conflict: An agent schedules a meeting at 2 PM because it sees the user is free, but the user has a standing doctor’s appointment every Tuesday at 2 PM that is not on the calendar. The agent should have asked about recurring blocks.
  • Resource overbooking: A project management agent assigns a developer to two tasks simultaneously because it only checks task deadlines, not the developer’s capacity. A quick question about workload would have prevented the bottleneck.
  • Preference mismatch: A travel agent books a hotel near the airport because it assumes convenience, but the user prefers a downtown location for sightseeing. The agent’s assumption was wrong, and the user must cancel and rebook.

These cases share a pattern: the agent had a default assumption, but the user’s actual situation differed. Asking would have surfaced the difference in seconds.

How to Choose the Right Moment to Ask

An effective planning agent uses a few heuristics to decide when to interrupt the user:

  • Ask early for high-stakes decisions: If the choice affects the whole plan (e.g., destination, budget, timeline), ask before generating any detailed schedule.
  • Batch questions: Instead of asking one question at a time, group related questions into a single prompt (e.g., “Do you prefer morning or evening workouts? Any days you can’t train?”). This reduces user friction.
  • Use defaults but confirm: Offer a sensible default and ask for confirmation only if the default is risky. For example, “I’ll book the 9 AM flight unless you prefer a later one.”
  • Learn from past answers: If the user has answered similar questions before, reuse that knowledge and only ask if the context has changed.

Costs of Asking Too Much or Too Little

Asking has a cost: it interrupts the user, takes time, and can become annoying if overdone. Asking too little has a cost: wrong plans, wasted resources, and user frustration. The optimal strategy is to ask only when the expected value of information exceeds the cost of interruption. In practice, this means:

  • For low-impact, reversible choices (e.g., which font to use in a report), don’t ask—just pick a reasonable default.
  • For high-impact, irreversible choices (e.g., whether to cancel a subscription, or which medical procedure to schedule), always ask, even if you have a default.
  • For medium-impact choices, ask if you have any doubt, but phrase the question so the user can answer quickly (e.g., “Is X okay?”).

Safety and Accuracy in Planning

When the plan involves safety-critical actions (e.g., medication reminders, equipment shutdown, financial transactions), asking is not optional—it is mandatory. An agent that picks without asking can cause harm. For example, an automated reminder system that assumes the user has taken a dose could lead to a double dose if the user actually skipped it. In such cases, the agent must confirm the user’s state before acting. Similarly, any plan that involves legal or financial commitments must be confirmed in writing or through an explicit yes/no prompt. The general rule: if a wrong pick could cause injury, loss, or legal trouble, the agent must ask.

Frequently Asked Questions

Why can’t the agent just use my calendar to know my preferences?

Calendars capture explicit events but not implicit constraints like “I don’t want meetings before 10 AM” or “I prefer to work from home on Fridays.” Those preferences are often unstated, so the agent must ask to learn them.

Does asking slow down the planning process?

It can, but the slowdown is usually small compared to the time wasted fixing a wrong plan. A well-designed agent asks only a few targeted questions and then proceeds quickly. The net time saved is often positive.

What if the user doesn’t know the answer?

Then the agent should offer a sensible default and explain the trade-offs. For example, “If you’re not sure, I’ll assume you prefer the cheapest option—you can change that later.”

Can an agent ask too many questions?

Yes, and that’s a common failure. Users abandon agents that ask trivial questions. The key is to prioritize questions by impact and batch them into one or two prompts.

How does this apply to AI assistants like Siri or Alexa?

Those assistants are planning agents too. When you ask them to “book a table for two,” they must ask about time, location, and cuisine because those are unknown. If they just picked a random restaurant, you’d be unhappy. Asking is what makes the plan useful.

Related

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →