Ask the org
One question, answered by every role you hold, with the plan shown to you before anything runs.
Overview
Ask the org takes one question and puts it to every role agent you hold. Each role answers from its own records only: Customer Success from health scores and tickets, the Sales Engineer from technical validation and open product gaps, and so on. A final pass reads those answers and writes one synthesis.
The role you are asking from is left out of the plan on purpose. You already have that view, so paying a second time for it adds nothing.
You see the plan before you spend
"Preview who gets asked" runs the planner and stops there. It shows exactly which roles would answer, the instruction each one would get, and roughly how many turns the run costs. Nothing has run at that point and nothing has been charged.
This is on the page rather than hidden in the API because the plan decides whose perspective ends up in the answer, and that is the one thing you are best placed to judge. If the plan leaves out the role that actually knows, you want to find that out before the run, not after.
If the preview shows the run would cost more than your remaining budget for the period, it says so with the figure, and the run button goes away. You can narrow the question or raise the cap.
Every claim traces to a role that read it
The answer carries a chip per role saying whether that role answered. The pass that writes the synthesis has no tools and no access to the underlying records: it can only work from what the roles reported. That is what makes the attribution meaningful rather than decorative, and it is why a synthesis can never quietly introduce a fact no role actually found.
A role that fails is shown, not hidden
One role failing does not fail the run. An answer built from three roles out of four is worth having, so you get it, with the missing role marked and its reported reason shown beneath. What you never get is a gap filled in silently.
If every role fails, there is no answer at all. A synthesis over zero readings would be the model's assumptions wearing your org's name, which is worse than nothing because it looks like something.
When it refuses, and why that is the feature
Three refusals happen before any spend. The run costs more than your remaining budget. No other role you hold can add anything, which happens when you hold one role and are asking from it. Or a role faces more tools and topics than one agent routes well, in which case it refuses and names the role rather than answering worse and saying nothing about it.
In all three cases nothing runs and nothing is charged, and the message says which case it was.
Nothing here writes anything
Every role agent in a run is read-only, enforced before any of them start rather than by convention. A run can tell you a renewal is at risk; it cannot update the opportunity, send the email, or create the task. Those go through the normal proposal path where a human approves them.
How it is shown
A question box with the role you are asking as, then two buttons: preview, and ask. The plan appears as one card per role with the instruction that role would receive. The answer appears below it with its attribution chips, and any refusal or missing role appears as its own flagged note. Changing the role you are asking as clears the plan, since the plan depends on it.