Short answer: They can take the repetitive, reversible work: order status, a stored answer, a draft a person sends. They should not, on their own, make a promise, issue a refund, or handle a complaint where the customer is angry or at risk. Rate the use before you connect it to the inbox.
I'm Vinay Pasricha. I run GoodSpace AI in Noida, and I wrote AI for Business Leaders. Customer service is where founders want an agent first, because the inbox is visible and the team is tired. It is also where a wrong sentence leaves the building with your name on it.
What the agent can hold
Most service work is not one thing. Some of it is a lookup. Some of it is a judgement. Some of it is an apology that has to come from a person. The function playbooks in the book separate those, and the separation is the whole answer.
An agent can tell a customer where an order is, if that fact is in a system a person also trusts. It can draft a reply from answers you have already approved. It can file the conversation so the next person does not start from zero. Those are useful, and they are enough for a first month.
An agent should not decide that a customer is wrong, offer a remedy you have not pre-approved, or keep writing after the customer has asked for a person. The handoff has to be real: a named human, a queue someone watches, and a rule for how fast they answer. A hidden "ask for a human" link that nobody reads is not a handoff.
Rate it before you turn it on
- Who is affected: a customer, not an internal note. The stakes start higher than a back-office draft.
- Worst case: a wrong promise, a leaked detail, a person left without help. Write the worst case in a sentence.
- Undo: can you recall the message, refund the error, or only apologise? If you can only apologise, a person sends it.
- The ladder: low-risk lookups can sit lower. Anything that commits the company sits higher, with a person deciding.
The Risk Ladder in the book runs from a low tier to a critical tier so the controls match the stakes. Use the same habit here. Do not pick the control because a vendor's demo was fluent.
A fluent reply is not a correct reply. Read twenty drafts against the last twenty real replies your team sent. If the agent invents a policy you do not have, it is not ready for customers, however polite it sounds. Fix the source of the policy, in writing, and run the twenty again.
A first month that stays safe
Let the agent draft. Let a person send. Count how often the person sent the draft unchanged, how often they rewrote it, and how often they threw it away. That count is your evidence. Widen what the agent may send alone only for the replies that were unchanged and that you would happily show a customer again.
Do not measure the month by how many tickets the agent "closed". A closed ticket can be a customer who gave up. Measure replies that were right, and the ones a person had to save.
Use the free Risk Ladder before the agent is allowed to act, and Automate, Assist, Escalate to sort the tasks around it. The eight steps are in AI strategy for business. Collapse or Collapse is the longer argument: one person owns one result and runs agents to deliver it.
One thing to do this week
Take last week's customer messages. Mark each one lookup, draft, or person. Only the lookup pile is a candidate for an agent to answer alone, and only after you have checked that the fact it would quote is actually in your system.
Where this comes from
This answer draws on The Risk Ladder and Function-by-function playbooks in AI for Business Leaders.
Your company will have its own details. To ask a follow-up in your own words, try Ask Vinay's books.
Vinay Pasricha is the founder of GoodSpace AI and the author of six books, including AI for Business Leaders.
Go deeper in AI for Business Leaders
This answer is the short version. The full argument, and the method behind it, is in the book.
