Safe CRM automation with AI: what should always ask first
An AI can now click through your CRM for you. That is only useful if it stops before the actions you cannot take back. This guide sorts common CRM jobs by how much damage a mistake does, and gives a checklist you can use on any tool, including ours.
Sort jobs by what a mistake costs
| Tier | Examples | Sensible rule |
|---|---|---|
| Read only | Look up a contact, list stalled deals, check which workflows are on | Safe to automate. Nothing changes |
| Reversible edits | Add a tag, change a lead source, create a draft | Let the AI do it, but show what changed |
| Edits to many records | Bulk stage changes, bulk tagging | Ask first, show the exact list, cap how many at once |
| Messages to people | Texts, emails, voicemail drops | Always ask. A sent message cannot be unsent |
| Publishing automations | Turning a workflow on | Always ask. One switch can message every contact that matches |
| Deleting or merging | Deleting contacts, merging duplicates | Always ask. Often permanent |
| Money | Charges, refunds, invoices, subscription changes | Do not let an AI do these |
A checklist before you trust any AI CRM tool
- Does it show the exact change before it makes it? "I will update 14 contacts" is not enough. You want the names and the new values.
- Can the AI approve its own actions? It should not. The approval must be your click, not the model's decision.
- Does it build in draft first? A workflow should be created switched off, tested, and only then switched on by you.
- Can you test on a fake contact? Make a test contact with your own email and phone number and run the automation against it before any real lead.
- Does it check the result afterwards? Reading the CRM after the change, rather than trusting the AI's report, is how you catch "done" when it was not.
- What happens if the page changes mid-task? A good tool stops and asks again. A bad one clicks whatever is now under the cursor.
- Does it treat text in the CRM as untrusted? Notes, form answers and contact names can contain instructions. The tool must not follow them.
- Are there limits? Cap the number of records per run and the spend per task.
- Can you stop it at once? There should be a stop button that works mid-task.
- Is there a record? You should be able to see afterwards what was changed and when.
Where kmand stands
kmand is a Chrome extension in private testing. kmand requests approval before recognised sensitive actions, including sending, deleting and publishing. It refuses steps it recognises as payments, such as paying or checking out, and lets you stop a task at any time. It cannot see everything a website does behind a button, so check each approval. No tool should be trusted to catch every risky click, including ours. On a practice app it builds a lead follow-up system in draft, tests it, and waits for approval before switching it on. The details, including what has not been tested and where it failed, are on how we test. Do not take our word for it. Use the checklist above on us too.