List situations that need a person.
Start with decisions that an automated agent is not authorised to make. Common examples include refunds outside policy, unapproved discounts, legal or safety complaints, threats, repeated service failure, requests involving sensitive personal information and any promise with an unclear cost or deadline.
Add language signals as well as topics. A customer may not write “I am angry”; they may use insults, repeat the same complaint, threaten to publish the exchange or demand a manager. Treat uncertainty as a trigger too. An agent that cannot verify an answer should hand it over instead of filling the gap with a plausible guess.
Define three actions, not one vague instruction.
“Escalate when needed” leaves too much room for interpretation. Give every trigger one of three explicit actions:
- Continue: answer from an approved source and keep the conversation moving.
- Acknowledge and pause: confirm that the message was received, make no new promise, and alert the owner.
- Pause immediately: send nothing further until a named person takes control.
Write the exact acknowledgement the agent may use. It should identify the next step without pretending that a decision has already been made. The agent should also be clear that it is an AI assistant; a difficult conversation is the worst time to create confusion about who the customer is speaking to.
Make the handoff useful.
An alert without context transfers the workload, not the problem. The handoff should include the customer, channel, original request, short timeline, facts already confirmed, promises already made, detected trigger and the exact decision needed. Include a link to the full conversation for detail, but do not force the owner to read it before understanding the issue.
Represent relevant messages and attachments in the handoff when they affect the decision. The summary must distinguish the customer’s claim from a fact the business has verified. If a customer contacts the business through both WhatsApp and email, do not assume the two conversations belong to the same person without confirmation.
Protect the moments most likely to cause harm.
Consider an additional review before messages involving money, promises, anger or a customer’s first impression are sent. Price and link validation must be designed, configured and tested separately; do not assume a written instruction creates a deterministic guardrail. If uncertainty remains, the defined rule should pause the reply for a person rather than allow an improvised answer.
Kai currently gives the owner Telegram controls to pause the agent or take over one conversation. Document what happens to queued messages when the pause is used and who can return the conversation to the agent. Without a resume rule, a human reply and an automated reply can collide. Kai identifies itself as AI when a customer asks.
Test the boundaries with past cases.
Create tests for obvious triggers and near misses: a normal refund, a refund outside policy, mild frustration, direct abuse, a known price, an unknown bundle and a question containing both a routine request and a threat. Verify the message sent, the alert, the handoff summary and the paused state. Keep the same test set for later changes so a fix in one case does not quietly break another.
Pair these rules with a clear brand voice guide and price-list controls. You can also review Kai’s current capabilities, operating profile and available roles.