LLM agents are increasingly tasked with creating plans for humans, such as driver scheduling or nurse assignment. Human requests often omit crucial facts, and supporting information may be outdated, forcing agents to ask clarifying questions. Existing benchmarks either assume fully specified problems or rely on uncalibrated LLM‑simulated users, which does not adequately test this ability.\n\nWe introduce Ask2Decide, a suite of 43 planning tasks that specifically evaluates whether agents request missing information that matters for the decision. An operator holds all facts absent from the agent’s request, and optimization solvers identify which facts would change the optimal plan. We also inject similarly worded facts that leave the optimal plan unchanged to test agents’ ability to distinguish useful from irrelevant details.\n\nWe collected 79 interaction sessions from 19 participants. After an incorrect plan, human replies omitted the correcting fact 43% of the time, far higher than the 7% observed with an uncalibrated simulated user. Using these data we calibrated a simulated operator whose reply pattern more closely matches real humans.\n\nEvaluating four frontier LLMs, we found that when information is missing their initial plans are infeasible on 58%–98% of tasks, compared with only 5%–28% when all facts are provided upfront. Agents discover facts that change the optimal plan only slightly more often than facts that leave it unchanged (39%–52% vs. 32%–41%).\n\nThese results show that acquiring decision‑relevant information through interaction remains a major bottleneck for LLM planning.\n\nReview: The work offers a systematic benchmark that highlights current models’ shortcomings in information gathering, pointing to clear directions for future improvement.