In CRMown, the AI approval workflow runs through a queue called AI Receipts. When the AI proposes an action (moving a deal to a new stage, say), it waits there until you approve or reject it. You choose which kinds of action can skip the queue, you can undo an AI action after it runs, and an admin can switch AI off for the whole organisation.
Why AI in a CRM needs a checkpoint
There are two easy ways to get AI in a CRM wrong.
The first is a chat box that answers questions but can’t actually do anything. Nice demo. Not much help on a Tuesday.
The second is the opposite. The AI acts with nobody checking, and one confident mistake updates a hundred customer records before anyone notices. Picture a salon where the AI decides a batch of regulars have “gone cold” and moves them all to the wrong stage the week before the holidays.
CRMown sits in between. The AI can do real work, but by default that work is proposed, not performed. You see what it wants to change before anything changes. That’s what people usually mean by human in the loop AI, and it’s the reason the AI approval workflow exists.
How does the AI approval workflow work?
The queue lives at AI Receipts (/dashboard/ai-receipts), in the AI Intelligence group of the sidebar. Here’s what’s in it:
- Every action the AI proposes lands in the queue before it touches your data.
- You approve or reject each one. A badge in the sidebar shows how many are waiting, so the queue can’t quietly pile up.
- You can leave feedback on each decision, and that feedback feeds accuracy tracking.
- Action stats show your acceptance rate over time.
- Per-contact action history shows every AI action that touched a given record.
Can you undo AI changes?
Yes. AI undo reverses an AI action that has already run. Combined with the per-contact history, you can see exactly what the AI did to a record and take it back.
That matters most for the actions you’ve allowed to run without review, which brings us to rules.
Approval rules: what can run on its own
Nobody wants to click approve on every suggestion forever. This is where you tune your AI approval workflow: approval rules decide which action types run automatically and which always wait for a person. They’re in Settings, AI Usage, next to token usage, AI memory and per-persona settings.
Here’s a rollout that works for most small teams:
- First week: everything needs approval. Watch what the AI suggests. Reject freely and leave feedback.
- Check the acceptance rate by action type. If you’re approving nearly every suggestion of one kind, that kind is a candidate to run automatically.
- Let low-risk, reversible actions go first. Anything a customer would see, or anything hard to reverse, stays on manual for longer.
- Look again every month. Rules are settings, not promises.
A five-person consultancy might be comfortable auto-running a few kinds of action after a couple of weeks. A business with a lot of customer data and a new team might keep everything manual for a quarter. Both are reasonable.
Where do the AI’s suggestions come from?
Several parts of the product feed the queue or follow the same suggest-then-apply pattern:
- Pipeline, scan for advances. The AI reviews your deals and proposes stage moves. Each move goes to the queue instead of happening directly.
- AI-suggested field values. On pipelines, the AI can suggest a value for a custom field, and you confirm or clear it.
- Meetings. After a call, the extracted action items are applied to the CRM in one click. Your click.
- Tasks. Plan my week proposes a week of work, then applies it if you accept.
AI permissions for your team, with personas
In AI Agents you can set up AI Personas. They’re assigned per user or per team, come with five presets, have their own scoped knowledge, and carry their own approval rules.
So the assistant your operations lead uses can work with different permissions from the one a new hire uses. Same system, different limits.
The off switch
Some businesses want AI off completely while they evaluate it. Settings, AI Usage has a master switch for the organisation. Turn it off and the whole AI Intelligence group disappears from the sidebar.
How to test AI safely before it touches real data
If you’d like to try approval rules without risking production data, Sandbox (/settings/sandbox) makes an isolated copy of your organisation for training and testing. Sandboxes are isolated by default and intercept email, so test runs don’t send real messages. When you’re happy, you can promote changes back to production and review them in a diff view first.
FAQ
Does the AI change my data without asking?
Can I undo something the AI did?
How do I know if the AI is actually useful for my business?
Can different team members have different AI permissions?
Can I turn the AI approval workflow off, or AI entirely?
Where can I see everything people asked the AI?
Keep reading
- How to replace your tool stack with all-in-one business software
- What a CRM for service businesses needs on day one
Try the approval queue on your own data, with everything set to manual approval to start.
Capabilities described from CRMown’s feature inventory, generated from the live codebase on 30 August 2026 (Tools, AI Intelligence, AI Receipts). Plan availability varies; check pricing for current details.