What is the core difference between these two kinds of "ask"?
The difference is the result when nobody responds. Per the changelog, a tool your organization's approval policy marks ask makes the scheduled run show a prompt and wait, while an MCP tool its server marks as needing a person asks only on the in-app card and the Scheduled Task goes on without those calls. One stalls; the other does one step less.
Why would Anthropic have MCP tools skipped in Scheduled Tasks?
The changelog gives no reason. A reasonable inference is that a tool marked as needing a person is designed to require someone present, and in an unattended scheduled setting it is preferable to skip that call and continue rather than stall the whole task. This is inference, not an official statement of motive.
How do I know which tool belongs to which kind?
The first is set by your organization's administrator in the approval policy; the second is marked by the MCP Server itself. The changelog does not give a way to see the server's marking in the interface, so the practical approach is to ask your administrator about the policy and to observe whether a Scheduled Task's output is missing what that tool would provide.
How can I reduce the risk now?
Do not make a critical step depend on a tool that may need a person; compare recent scheduled output for gaps; and split any step that might be skipped into a manual process. Update the desktop app to a version that includes v3.6556.1, since that release records this behavior.
A Scheduled Task runs with nobody present, so how "needs confirmation" is handled is the core of its reliability. The official changelog actually records two different treatments, and their outcomes are opposite.
The October 8 update (v2.31226.0) says that when a scheduled task reaches a tool the organization's approval policy marks "ask," it used to be skipped; now the run shows the approval prompt and waits, and a newer run of the same task retires an earlier run's unanswered prompt. The outcome here is stop and wait for a person.
The October 22 update (v3.6556.1) says MCP tools their server marks as needing a person now ask only on the in-app card, so their notifications have no Allow button, and Scheduled Tasks go on without those calls. The outcome here is one step fewer, everything else runs.
In interactive use both look like "Claude asks me." In a scheduled task, the first leaves the whole run hanging, which you will notice soon; the second lets the task finish as usual, just missing a piece, which is harder to spot. The changelog does not say whether a skipped call is recorded or notified, and the official scheduled-tasks page likewise does not cover approval behavior during runs, listing only an "approval mode" field at setup. The judgment that this is easy to miss is my inference.
Do not hand a critical step to a tool that may need a person's confirmation. If a task must use one, run it manually while you are present, or split that step out so a gap in the result becomes visible.
List which connectors and MCP tools your scheduled tasks use, and check which your organization marks ask and which their servers mark as needing a person. For the first, expect a run that may sit waiting for approval; for the second, expect a result that may be missing a piece without an error. Either way, the most reliable check is comparing the task's actual output against what those tools would normally have produced.