AI for customer support: draft replies, deflect, keep a human in the loop

A customer emails "where is my order, it has been five days." A support AI pulls the order from Shopify, sees it shipped Tuesday with a tracking number, drafts a reply with the link and an apology for the delay, and posts it in Slack for an agent to send. The agent reads it, tweaks one line, and hits send. Forty seconds, not four minutes. That is the shape of AI support that works in 2026: the AI does the lookup and the draft, a person owns the send.

What to automate, and what to keep human

The fastest way to get this wrong is to point AI at the whole queue and hope. Sort tickets by how repetitive and how reversible they are, and only the top band is safe to lean on heavily.

• Lean on AI: order status, password resets, "how do I" questions, refund-policy lookups, the same five questions you answer every day.
• AI drafts, human sends: anything customer-facing that is a judgement call, an apology, a goodwill credit, a frustrated account.
• Keep human: cancellations from a key account, complaints heading for a refund dispute, anything legal or safety-related.

The dividing line is not how hard the question is. It is whether a wrong answer is recoverable before it reaches the customer. Repetitive and reversible goes to AI; sensitive and irreversible stays with a person.

Drafting in your voice, not a template

Generic support copy is worse than a slower human reply, because customers can smell it. The fix is grounding the AI in your real material: your help docs, your past replies, your tone. An AI that reads how your team actually answers before it drafts produces a reply that sounds like you, with the specific order number and the specific fix, not "we apologise for any inconvenience caused." You review and send; you do not start from a blank box.

Deflection without a wall of bot

Deflection has a bad name because most of it is a chatbot blocking the path to a human. Done well, it is quieter. The AI answers the genuinely repetitive questions instantly, and the moment a ticket needs judgement, it hands off to a person with the context already gathered. The measure of good deflection is not how many tickets the bot closed. It is how many customers got the right answer without waiting, and how cleanly the rest reached a human.

Escalation: hand off with the work already done

The point of escalation is not to dump a ticket back on an agent cold. A useful support AI escalates with the lookup finished: the order pulled, the account history summarised, the likely fix drafted. The agent opens a ticket that already has the facts and a proposed reply, and spends their time on the decision, not the digging. Escalate the judgement, not the busywork.

Why full autonomy on day one is a mistake

The tempting move is to let the AI send on its own and watch the queue drain. Resist it. The cost of one wrong auto-sent reply, a refund promised that should not have been, a curt message to your biggest customer, is far higher than the time saved on the easy tickets. Start with the AI drafting and a human approving every send. Once you have weeks of evidence that the drafts are right, you can auto-send the narrowest, safest category, order-status replies with a tracking link, and keep approval on everything else. Earn autonomy ticket by ticket; do not assume it.

Running it through Echo

Echo handles support from Slack. It connects to your help desk, Shopify, Stripe, and inbox, reads the ticket and the order, and drafts a reply in your voice into the thread. It asks before anything reaches a customer, sending, refunding, or editing a record all wait for your yes, so nothing goes out unreviewed. It does not train on your data, and because it does not charge per seat, your whole support team can use it without the bill climbing. The first $50 of work is free, which is enough to test it against a real morning of tickets.