The Assistant refused to do it
A refusal is one of four things, and they need different responses. The Assistant lacks the authority; the channel's autonomy rule says a person decides; the evidence a rule requires is missing; or it could not form a plan from what you said. Only the last is about your wording. The first cannot be lifted at all — a person has to act instead.
What you are seeing#
You asked for something specific and got a sentence declining it, usually naming the limit. No proposal appeared, or one appeared and could not be confirmed. The status line finished normally, because a refusal is a completed turn rather than a failure.
A refusal is worth reading rather than rephrasing around. Three of the four are rules working correctly, and rewording the request to get past a rule either fails again or, worse, produces something the rule existed to prevent.
The four refusals#
| Kind | It sounds like | Who can lift it |
|---|---|---|
| Authority | It says it is not permitted to do this | Nobody. A person with the right does it on the screen instead |
| Autonomy | It says a person has to approve this first | The workspace, by changing the channel's mode — or you, by approving it in Needs You |
| Evidence | It says it has no source for the thing you asked it to use | You, by supplying the source; the rule itself stays |
| Plan | It says it is not sure what you want, or asks a question back | You, by being specific about the record and the change |
The Assistant's rights are deliberately narrower than yours. It does not act on pricing, and it cannot clear a do-not-contact entry — those two are the clearest examples of authority refusals, and both exist because an agent that can quietly re-enable contact with somebody who asked not to be contacted is a liability rather than a convenience. Separating the agent's authority from the person's is the reasoning.
What Connect completed, and what it did not#
- What Connect completed
- The reading. A refusal usually comes after the Assistant has looked — it found the record, read the rule, and then declined. The answer often contains that finding, which is useful.
- What Connect completed
- The audit. A refusal that mattered is recorded, so a later question about why something did not happen has an answer.
- What Connect did not complete
- The action itself. Nothing was written, nothing partial exists, and no compensating step is needed.
- What Connect did not complete
- Any surrounding steps that depended on it. If you asked for three things and one was refused, the other two either ran or did not — the step list says which, and it is worth opening.
What you can do#
Read which of the four it is before rewording anything.
Result Three of them will refuse the reworded version too, and you will have spent two more model calls learning that.
For an evidence refusal, supply the evidence.
Result Add the source to Knowledge, or point at the record that carries it. The refusal was about the missing source, not about the request.
For an autonomy refusal, open Needs You.
Result The held item is there with the message it answers. Approving it is the same decision, taken in the place designed for it.
For a plan refusal, name the record and the change in one sentence.
Result Most plan refusals come from an instruction with two possible targets. Pinning the record first removes the ambiguity for every request in that tab.
What an administrator can do, and when to escalate#
An administrator can change the autonomy mode for a channel, or narrow it to one contact so that a single sensitive account keeps a person in the loop while everything else moves. They can add the Knowledge source an evidence refusal is asking for. They cannot widen the Assistant's authority: that boundary is part of the product, not a setting.
Escalate when a refusal names a limit that does not appear to apply — a suppression on a recipient who never asked to be suppressed, or an authority refusal for something well inside the Assistant's ordinary work. Report the exact sentence it gave you and the record it concerned; those two identify the rule that fired without needing anyone to reconstruct the conversation.
Questions#
Can I be given permission to override a refusal?
Not for an authority refusal, because the limit belongs to the Assistant rather than to your account. You may well be able to do the thing yourself on the relevant screen — that is the point of keeping the two sets of rights apart.
Why did it refuse something it did last week?
Usually the channel's autonomy mode changed, or the recipient reached a suppression list in between. Both are visible: the mode on the channel, the suppression on the contact, with the reason it was added.
Is a refusal counted against my AI budget?
The model call is metered like any other turn, because the model ran. What is not charged is the action, since no action took place.