Trust and letting an agent send by itself

Minimum role: Account admin
Applies per: organisation
Last updated: 12 October 2026
aishaa version: 80c056408

Menu path: Account > Agents > (open agent) > Trust

In short: Every reply an AI agent writes starts as a proposal to a colleague. aishaa keeps track, per agent, organisation and topic, of how often colleagues send those proposals unchanged. When there are enough of them, aishaa proposes letting the agent send by itself on that topic. You decide: only when an Account admin clicks Let it send by itself may the agent reply on that one topic without a colleague, and even then only if every guardrail allows it. With the Emergency stop you turn everything back into proposals in one click.

Where to find this

Go to Account > Agents, open an agent and click the Trust tab. The tab appears on an agent that has already been saved.

Who: you use Let it send by itself, Take back and the Emergency stop from the Account admin role. An Organisation admin sees the tab with all the figures, but the buttons and the switch are disabled for them. The tab shows the organisations of the account you have access to; trust and permission apply per agent, per organisation and per topic.

The Trust tab of an agent with, per organisation, the Emergency stop switch and a table with the columns Topic, Decided, Unchanged, Edited, Rejected and Sent by itself, with the text aishaa proposes and the Let it send by itself button on an earned topic, menu Account > Agents.

Every reply starts as a proposal

When an agent writes a reply to a customer message, it arrives as a proposal in the ticket’s editor. A colleague reads it, adjusts it if needed and sends it, or clears it with Clear AI answer. That is always how it goes, until someone switches on Let it send by itself for an agent on a topic.

The explanation on the tab puts it like this: Every reply starts as a proposal to a colleague. Where this agent’s proposals go out unchanged often enough, aishaa proposes letting it send there by itself; you decide.

How you review and send a proposal is described in Replying to a ticket.

How trust is measured

aishaa counts, per agent, per organisation and per topic, the proposals a colleague decided on. A topic is the query pattern from the taxonomy that aishaa gave to the customer message; in the table the support area is shown below it. See Taxonomy: support areas and query patterns.

Column	Meaning
Topic	The topic, with the support area below it.
Decided	Proposals a colleague decided on: sent or rejected.
Unchanged	The share of those that a colleague sent as the agent wrote it, apart from whitespace or a typo.
Edited	Proposals that were sent after a colleague adjusted the text slightly or substantially.
Rejected	Proposals that a colleague cleared without using them.
Sent by itself	Replies that the agent sent by itself on this topic. These never count towards trust.

A topic is earned when all three conditions are met:

Condition	Value
Period	The last 30 days.
Number decided	At least 50 proposals a colleague decided on.
Unchanged	At least 95% of those sent unchanged.

The summary above the table also mentions these thresholds: Every reply starts as a proposal. Earned on a topic: at least 50 proposals a colleague decided on in the last 30 days, 95% of them sent unchanged.

What does not count:

Replies that the agent sent by itself. Trust is never built on itself.
Proposals that were set aside without anyone deciding on them, for example because the customer sent a new message in the meantime or a colleague sent a different reply.
Replies that a colleague wrote with aishaa on this ticket. That is writing help for a colleague, not work by the agent.
Messages without a topic. If a message gets no topic from the taxonomy, no permission can apply to it.

A change to the agent starts over

Trust belongs to the agent as it is now. That is why aishaa only counts proposals from the last change to who the agent is. Starting over happens on:

a change to the agent itself: the General instructions, the knowledge, the memory, the skills, the tools and what the agent may do with them, or the conditions per tool;
a change to a custom skill that the agent has;
a change to the Additional instructions on the Organisation information page of that organisation.

A new name, a different description or the Enabled switch change nothing about what the agent does and leave the trust as it is.

If someone had already given permission before the agent changed, that permission lapses. The agent proposes again on that topic until it has built up the trust again and someone switches on Let it send by itself again. So nothing ever quietly stays on after a change.

Switching on Let it send by itself

On a topic where the agent has earned the trust, the last column shows aishaa proposes: with the number and the percentage, for example aishaa proposes: 60 times, 97% unchanged, and the Let it send by itself button. On a topic that is not that far yet, it shows how many decisions are still needed, for example 12 more to go.

Go to Account > Agents, open the agent and click the Trust tab.
Under the right organisation, find the topic with aishaa proposes.
Click Let it send by itself. You see The agent now sends by itself on this topic. The table then shows the chip Sends by itself.

The button only appears on an earned topic. Giving permission for a topic that has not been earned yet is not possible. If the Emergency stop of that organisation is on, the button is disabled.

A permission applies only to this agent, on this topic, in this organisation. Other topics and other agents keep proposing.

The guardrails

Even with permission, a reply only goes out by itself if every guardrail allows it. If one guardrail stops it, the reply stays an ordinary proposal for a colleague. aishaa checks in this order:

Guardrail	When the reply stays a proposal
Emergency stop	The organisation’s emergency stop is on.
AI paused	The AI is paused for the organisation.
Marketplace	The conversation runs via a marketplace. Marketplaces do not allow automatic messages to buyers. See Messages from marketplaces.
Permission	Nobody gave this agent permission on the topic of this message, or the permission has lapsed because the agent changed.
Still earned	The agent no longer meets the thresholds on this topic, for example because colleagues have been adjusting more often lately. As soon as it meets them again, the permission applies again.
Sample	One in 10 replies always goes past a colleague, even on a topic with permission. Which reply that is, is fixed per proposal.
Other workflow	Another workflow has already written to the customer about this message. Without a colleague, nobody would notice a duplicate message.
Placeholders	The reply contains a placeholder: a value that only a colleague can fill in.
Agent’s doubt	The agent that wrote the reply indicated itself that a colleague should look at it.
Safety check	A separate, independent check reads the customer’s message, the conversation, the proposal and what the agent looked up, and does not let the reply through.

If the reply does go out by itself, aishaa sends it under the organisation’s name, with the organisation’s Signature for automatic messages (under Organisation > General, see General organisation settings). If the message cannot be sent in that conversation at that moment, or the conversation has changed in the meantime (a new customer message, a colleague who already picked up the proposal), it stays a proposal and a colleague sends it.

Sending by itself in a workflow

Who replies to a customer message is determined by the organisation’s workflow with the Write the reply step. That step has the Send by itself choice with two options:

Option	Meaning
Where the agent may (earned and allowed per topic)	Default. The reply goes out by itself where someone gave the agent permission and every guardrail allows it; otherwise it stays a proposal.
Never: always a proposal	Every reply from this step stays a proposal, even on topics with permission.

If a reply goes out by itself, the step leaves the workflow via the Sent by itself exit. If it stayed a proposal, the step’s outcome shows under Why it stayed a proposal which guardrail stopped it, for example sample for a colleague or not allowed on this topic.

In the workflow editor, Try it on last week shows what a draft of a workflow would have done with the current permissions: per message, whether a reply Would go out by itself, stayed a Proposal to a colleague or needed a person. Nothing is sent, and the safety check does not run in this case. See Answering customer messages with workflows and Create, test and publish a workflow.

A follow-up that a colleague releases

There is one other way a reply can go out without a colleague: a colleague snoozes a ticket with Let aishaa pick this up when the snooze ends and ticks aishaa may send the reply immediately themselves. That colleague’s choice replaces the permission per topic and the sample for that one ticket. The other guardrails still apply: the emergency stop, the marketplace, the placeholders, the agent’s doubt and the safety check. Without the tick, the reply becomes a proposal. See Snoozing and wake-up calls.

Taking back and the emergency stop

Taking back one topic: in the row with Sends by itself, click Take back. You see The agent proposes again on this topic. From that moment on, every reply from this agent on that topic is a proposal again. If you want it again later, someone has to switch on Let it send by itself again.

Everything back at once: above the table of each organisation is the Emergency stop switch (On: every reply of this organisation is a proposal again.).

Switch the Emergency stop on. You see Emergency stop on: every reply is a proposal.
From that moment on, no agent in that organisation sends by itself any more, on any topic. Follow-ups that a colleague had released also become proposals.
The permissions are kept. If you switch the emergency stop off again (Emergency stop off.), they apply again as they were.

The emergency stop applies per organisation and to all agents at once. You find it on the Trust tab of every agent; switching it on or off on one agent therefore works for all agents of that organisation.

Seeing what was sent by itself

On the Trust tab, the Sent by itself column shows per topic how many replies the agent sent by itself.
In the ticket, a reply sent by itself appears in the conversation with the organisation’s name as the sender instead of a colleague’s name, and with the Signature for automatic messages.
In the workflow that replied, you see per run whether the Write the reply step went via Sent by itself, or why it stayed a proposal. See Answering customer messages with workflows.
Every permission, every taking back and every time the emergency stop goes on or off appears with name and time in the Changes overview. The automatic lapsing of a permission after a change to the agent appears there too. See Changes and reverting.

A workflow step that acts by itself from now on

Apart from replies to customer messages, a workflow has steps that send or carry out something, such as sending a message, forwarding or an action in a connected system. Such a step waits for a colleague by default. If a colleague has passed on the proposal of the same step unchanged 10 times in a row, aishaa offers Do this itself from now on in the ticket, with the question From now on this step goes out at once, without a colleague. Sure? Confirming is possible from the Organisation admin role, because it changes the workflow. After confirming, that step carries it out by itself from then on; you turn that back in the workflow. See Answering customer messages with workflows.

Good to know

aishaa only sends a reply without a colleague if a person has allowed it: with Let it send by itself for that agent on that topic, or with the aishaa may send the reply immediately tick on a single snooze. Without that permission, every reply is a proposal.
A disabled agent writes nothing, so it sends nothing either.
Trust is measured per organisation. The same agent can have earned a topic in one organisation and not in another.
A change to an agent’s instructions via Tell aishaa or via Back to the standard also counts as a change: trust starts over and earlier permissions lapse.
Actions from a connected system that an agent may carry out while writing a reply are set per tool on the Skills tab. That is separate from sending by itself. See Skills.