What AI should and should not do in customer support

Most teams are no longer deciding whether to use AI in support. They are deciding where it belongs. Our answer is narrow on purpose: AI is good at the first draft and the tedious rewrite, and it should stay away from the decision to send.

Let it draft, not decide

A reply written from a long thread takes a minute of reading and a minute of typing. A draft removes the typing and leaves the reading, which is the part that needs judgement. The agent still checks the facts, adjusts the tone and decides whether this customer needs a paragraph or a phone call. A draft that goes out unread is not a time-saver. It is a risk you have not found yet.

Hold it to the facts in the ticket

The failure that costs the most is the confident invention: a refund policy that does not exist, a deadline nobody promised, an order detail that belongs to a different customer. A support assistant should work only from the conversation in front of it. When something needed is missing, the right behaviour is to ask the customer for it, or to leave a visible gap for the agent to fill, not to guess.

Keep the customer’s language

Customers write in the language they think in. A draft should answer in that language rather than defaulting to English, and a rewrite of the agent’s own words should stay in the language they wrote it in. If you serve customers in more than one language, check this first, because it is where weak tools show.

Teach tone, not facts

Past replies are a good guide to how your team sounds: how it greets, how it apologises, how long its answers run. They are a poor source of facts. Another customer’s order number or outcome should never leak into this customer’s reply, so past replies should shape style only.

Turn solved tickets into answers

Every solved ticket holds an answer someone worked hard to find. AI can turn that conversation into a first draft of a help article, but it needs two safeguards: strip anything that identifies the customer, and say plainly when the conversation never reached a confirmed fix. A person reviews and publishes it.

Know where the text goes, and who can use it

Using AI on tickets means ticket text goes to an AI provider. That is a decision for your team, not a default. Ask who processes it, whether your content is used to train public models, who on your team can use it, and whether you can switch it off without losing anything else. Ask for a record of use too: who used it, for what, and how much.

A short checklist

• An agent reads and approves every reply before it is sent.
• The assistant works from the ticket only, and leaves gaps visible.
• It answers in the customer’s language.
• Past replies guide tone, never facts.
• You can see who used it, and switch it off person by person.
• Ticket text is not used to train public models.

This is how we built the AI assistant in nabudesk, available on Pro and Enterprise. It drafts, refines and turns solved tickets into article drafts, and it leaves the decision to send with the agent. See how it works.