An AI employee in CRAIM runs in one of three modes. Manual: it acts only when a person asks. Copilot: it drafts every customer message and waits for approval. Autopilot: it sends on its own, except for the actions you marked as always needing a person. New employees start in copilot, and this is what each mode changes.
The mode is a setting on the employee, and it decides one thing: what happens between the moment the employee has prepared an action and the moment that action reaches a customer or the CRM. Everything else, the knowledge it reads, the channels it works in, the tools it may call, stays the same in all three.
A new employee is created in copilot. That default is deliberate. The first week of an employee's work is the week you learn what its instruction is missing, and copilot is the only mode where you see every draft before a customer does.
What each mode changes
The three modes differ at two points: whether the employee wakes up on its own, and whether what it prepared goes out without a person.
| Mode | Wakes up on an inbound message | Customer messages | Actions marked "always approve" |
|---|---|---|---|
| Manual | No. It acts only when a person asks it to in the conversation; the blocked wake-up is recorded | Wait for approval | Wait for approval |
| Copilot | Yes | Wait for approval | Wait for approval |
| Autopilot | Yes | Sent by the employee | Wait for approval |
Manual is for an employee you want to consult, not run: a colleague you call into a thread with a question. It does not answer the customer at night, because it does not wake up at night.
Copilot is the working default. The employee reads the incoming message, writes the reply, and puts it in the approvals queue as a card with the draft, the customer context and the facts it used. A person presses approve, edits, or asks for another version. The customer sees nothing until then.
Autopilot removes the wait for messages. What it does not remove is the per-action rule: an action whose approval policy is "always" waits for a person in autopilot too. That is how a discount or a change of terms stays with the team while routine replies go out in seconds.
What to read during the copilot week
Every draft in the queue is a test of the instruction and the materials, and the corrections sort into three kinds.
- A wrong fact. The price is old, the delivery time is wrong, the condition does not exist. The fix is in the materials, not in the instruction: update the document or the catalogue row, and the next draft is right.
- A right fact in the wrong place. The employee answered a question it should have handed over, or promised a callback nobody will make. The fix is a line in the instruction: what it never promises, when it calls a person.
- The wrong tone. Too formal, too chatty, a different name for the product. The fix is the tone document and, if that is not enough, an example of a good reply pasted into the instruction.
If the same correction comes up three times, it is not a draft problem. Stop approving and fix the source. Every draft you approve without noticing a repeated mistake is a draft autopilot would have sent.
When to switch to autopilot
There is no counter that turns green. The honest signal is the approvals queue: when you open it and press approve without editing for a few days in a row, the drafts have stopped surprising you. That is the moment, and the checklist for that week lists what to look at before pressing the switch.
Switch one employee, not all of them. Keep the actions that cost money on "always approve". And keep reading the history: autopilot removes the wait, not the record. Every message the employee sends is in the conversation with its run, so a week later you can still see what it said and why.
When to go back
Back to copilot is one click, and there is no shame in it. The usual reasons are a change in the materials the employee has not been retrained on yet, a new channel with a different kind of customer, or a person joining who wants to read the drafts before trusting them. The mode is the employee's, so one employee can be in autopilot for a Telegram bot while another, on a new WhatsApp number, is in copilot.
Common questions
- What is the default mode for a new AI employee?
- Copilot. The employee wakes up on incoming messages and writes replies, and every customer message waits for a person's approval before it is sent.
- Does autopilot send everything without a person?
- Customer messages, yes. Actions whose approval policy is set to always still wait for a person in autopilot, and read-only tools never wait in any mode.
- What does manual mode do?
- The employee does not wake up on events. It acts only when a person asks it to in the conversation, and a blocked wake-up is recorded so you can see what it would have reacted to.
- Can different employees run in different modes?
- Yes. The mode is a setting on each employee, so a proven employee can run in autopilot while a new one on another channel stays in copilot.
- Is there a signal that says an employee is ready for autopilot?
- Not a numeric one. When you approve drafts for several days without editing them, the instruction and the materials are complete enough. Switch one employee, keep the expensive actions on always-approve, and keep reading the history.