Yes, for service messages triggered by facts in your own systems: a delayed delivery, a failed payment, an outage, a document about to expire. The agent uses approved templates, respects contact preferences, and sends one message per customer per event. Large batches, sensitive subjects such as arrears or fraud, and anything resembling marketing need a person's approval.
Why is proactive contact riskier than replying?
Because mistakes multiply. A wrong reply reaches one customer who asked a question. A wrong trigger, such as a misread status code or a stale list, can send the same wrong message to thousands of customers at once, none of whom expected to hear from you. The damage is done before anyone notices.
So the unit of approval changes. For replies, you approve the agent's capability and sample its work. For proactive messages, a person approves each batch, or at least each new trigger, before it runs.
What conditions decide it?
- A trigger from a reliable record: a carrier delay, a payment returned by the receiving bank, an incident in the status log. Not a guess.
- A lawful reason to contact and contact preferences that allow it on that channel.
- An approved template for each trigger, with only the customer's own details filled in.
- Caps: one message per event per customer, and a limit on batch size before approval is needed.
- Quiet hours, so non-urgent messages do not arrive at 3 am.
What may the agent send, and what stays with a person?
| Message | The agent | A person |
|---|---|---|
| Delivery delay for one order | Sends from the template | None |
| Failed or returned payment, one customer | Sends what happened and what to do | None |
| Outage notice to all affected customers | Prepares the batch and the message | Approves before sending |
| Document or card expiry reminder | Sends from the template | None |
| Missed payment or arrears contact | Never on its own | Collections team decides |
| Message to a flagged vulnerable customer | Holds | Decides channel and wording |
| Offers or promotions | Never as a service message | Marketing, under its own rules |
Outage notices are the clearest case for batch approval; the process is in how to send proactive service outage notices with an AI agent.
What does this look like at a payments app?
Take a payments app. One morning, 300 customer transfers come back from receiving banks, most because of an invalid account number. The operations team works through the rejected payment messages between banks, repairing what it can.
For the customers, the agent prepares a message: the transfer was returned, the money is back in their balance, and they should check the recipient's details before trying again. The message names the amount and recipient from each customer's own record. Because this is a batch of 300, the operations lead reviews the template and a sample of filled messages, then approves. The agent sends them and logs each one. Customers whose transfer was repaired and resent receive a different template, approved in the same way.
Should proactive messages be reviewed before sending?
For established triggers with approved templates, the template is the review; individual messages need not be checked. New triggers, unusual volumes and any message with free text need a person first. The trade-off between speed and review is covered in whether an AI agent should send replies to customers without review.
Some outbound messages are never proactive service. A final response to a complaint is a formal document with its own approval, covered in whether an AI agent can send the final response to a complaint.
Which controls make it safe?
- Triggers defined and approved by a named owner, each tied to a reliable source.
- Approved templates, with only record fields filled in.
- Batch size limits above which a person approves.
- Contact preferences and quiet hours checked at the moment of sending.
- A kill switch to stop a batch mid-run, and a filed record of every message sent.
The wider approach is in the guide to AI customer support agents.
Questions people ask
Can customers reply to proactive messages?
They should be able to, and the reply should reach the agent with the context of the message that prompted it.
Do proactive messages need AI disclosure?
Where the customer can interact with the agent through the message, they should know it is an AI. Templates can say so briefly.
What if a batch goes out with an error?
Stop the batch, send a correction to those affected using an approved template, and treat it as an incident to find the cause.