How AI changes are approved, recorded and undone
In AgentticCRM, AI arrives by two paths, and neither reaches further than the person using it. The Assistant prepares each change and waits for your approval. A connected ChatGPT or Claude carries out the work with the role of the person it represents, and each of its changes lands in Agent activity, where a person reviews it and, if needed, undoes it.
- The Assistant changes nothing until you approve
- A connected agent works with the role of whoever asks
- Every change is logged with who asked for it
The AgentticCRM Activity log inside Settings, with fictitious data: the filters by person and by action, the search, and the table with the When, Who, Action and Detail columns. Ten entries from one morning, newest first: a project credential opened, a proposal the client accepted in their portal, a role change, an Assistant action a person approved, a confidential send created, a ticket moved to in progress, sign-ins, a failed attempt and a recovery point the System verified.
The Assistant proposes the change and you approve it
When you ask the Assistant for something that changes data, such as creating a task or moving an opportunity to the next stage, it answers with a proposal: the steps it would take, the data for each one and what it will not do. Before you decide, you can remove steps and correct fields in the proposal itself.
Nothing happens until you approve it. At that moment AgentticCRM checks your role and permissions again, so a proposal prepared yesterday is no good if your role changed today. And if nobody attends to it, it expires on its own.
A connected ChatGPT or Claude works with the role of whoever asks
If your team already uses ChatGPT or Claude, each person can connect it to AgentticCRM from ChatGPT or Claude itself. The connection brings no permissions of its own: what counts is the permissions of the person it represents, and with no identified person behind it there is no call, not even to look something up.
Within that role, a connected agent carries out what it is asked, and its work is reviewed afterwards. Every change is kept in Agent activity, where anyone with permission reviews it and puts it back the way it was. If someone has touched that record since, undo is not offered, so their work is never overwritten.
What is recorded for every change
Two screens keep the history: the Activity log, with what happens across your organization, and Agent activity, with what the connections did.
-
Who asked for it
Every action is recorded under the person it was done for and, when it came through a connection, with the agent's name in the detail.
-
Exactly what changed
The data that was written is kept, and the change also shows on the client, invoice or ticket it touched.
-
Who approved it
For the Assistant's proposals, it records who approved or discarded them, and when.
-
What was undone
Every undo leaves its own entry, so you see the change and its correction too.
-
A log nobody can edit
Nobody can edit or delete an entry in the Activity log, and restoring a recovery point does not change it either.
-
Who can read it
Only the people with permission to view activity, which you grant through their role or their individual permissions.
Questions about AI and control over changes
Can the AI see or do something I cannot?
No. The Assistant and connected agents work with the role of the person using them, and that role is checked again on every action. If you change someone's role today, what the AI can do for that person changes today.
Who approves what a connected ChatGPT or Claude does?
For what that person's role already allows, nobody beforehand: the agent carries it out and its work is reviewed afterwards in Agent activity. It only waits for approval when the action cannot be undone, when AgentticCRM marks it as always needing approval, or when your organization chose to approve it before that agent takes it.
What happens to an Assistant proposal nobody approves?
It expires on its own and is never carried out. If you need that change later, you ask again and it prepares a new proposal with the data as it stands then.
Can everything a connected agent does be undone?
What it carries out without waiting for approval, yes, because an action with no way back always waits for a person. Undo stops being offered once someone has changed that record since, so their work is not erased, and the screen tells you why.
How do you disconnect a ChatGPT or Claude?
Each person disconnects their own from My agents, in their profile, where they also see when they last used it. The keys your organization registered for its own agents are revoked in Settings, under External agent keys.
Can the AI get into AgentticVault or confidential sends?
No. Neither the Assistant nor a connected ChatGPT or Claude has access to AgentticVault, and they cannot open or create a confidential send. Both are encrypted in the person's own browser.
More about security in AgentticCRM
- How we protect your dataThe security overview: separate data, permissions, accounts, AI and recovery.
- Separate data and permissionsHow your organization is kept apart from the others and what each person on your team sees.
- Backups, recovery and deletionRecovery points, how you go back to an earlier state and how deletion really happens.
- Activity and auditInside the Activity log: what it records, how you filter it and who reads it.
- AssistantThe Assistant's screen, what you can ask it and how your ChatGPT or Claude connects.
- AI CRMWhat AI does for your team day to day, with your data and your permissions.
See a change approved and undone, live
In a demo we ask the Assistant for a real change, review the proposal before approving it, and open together the record of what it did and how to undo it.
Prefer to explore on your own? Work out your current costs, see how the Assistant works or use the CRM evaluation guide.
Create your organization online and try it for 14 days, no card required.