Customer support with AI

Support is where AI looks most obvious and where a miss hurts most. A bad-tone draft or an invented lead time goes out over your signature. The visitor does not know “it was the model”. They know you lied to them.

That is why the first design is not a 24/7 bot. It is an internal copilot that writes 70% of the email or ticket, with sources, and a human who sends. The public bot, if it arrives, is earned with a month of high accept rate and a cited FAQ.

This is the order we hold in SMEs that live on the phone and email, not a 200-agent contact centre.

The draft flow (the one you pay back quickly)

A ticket or mail arrives. The system proposes a reply from your templates and FAQ. It marks what it does not know. The agent edits and sends. You store whether it was accepted as-is, rewritten or discarded.

In two weeks you see the real house tone and the questions missing from the FAQ. That feeds the site and the internal copilot. The bot is not needed yet.

Rules that prevent the mess

• Never a lead time, price or availability that is not in the source.
• Never an automatic “sorry for the delay” when there is no delay: it sounds like a cheap template.
• Anger, legal or health cases leave the automatic queue immediately.
• The name on the signature is a person. The model does not sign.
• A kill switch per channel (mail, chat, WhatsApp).

When (if ever) the public bot

When the FAQ already answers the twenty questions and the form does not drop leads. When the internal team has a month of high accept rate. When the bot cites and escalates. When there is a real person behind it in real hours — not “the bot covers nights” as an excuse not to staff.

If volume is fifteen emails a day, the bot is not the lever. The draft and the FAQ are. Do not buy theatre for a volume that fits a well-run inbox.

Support metrics

First response time, resolution time, draft accept rate, reopen rate, and a weekly reading of three ugly tickets. CSAT if you already measure it; do not invent it for the pilot.

If time drops but more tickets reopen, the model is closing badly. That shows up on the sheet, not in the demo.

Training and programme

Two champion agents, real cases, review at day 14. House tone is taught with your best emails, not a generic “be empathetic” prompt. Kit Digital can touch web presence and customer management; the deliverable is an operable queue, not the widget.

House tone: how it is actually taught

Pick ten of your emails you are proud of and five you would not send again. Those fifteen are the tone manual. The prompt says: imitate the ten, avoid the five, do not invent lead times, cite the FAQ. Nothing about “be warm and professional” — everyone writes that and it sounds like no one.

The champion marks which sentence in the draft is not us. At day 14 those marks become two more rules. Tone is not downloaded. It is cut.

If you do not have ten good emails, the problem is not AI. Support has no voice. Write three templates by hand before you automate.

WhatsApp, without making the owner’s phone the system

If the channel is WhatsApp, the risk is not the model: it is that the thread lives on a phone. First, messages must land in the same queue as email (a company number, a seat). Then drafts. A bot inside WhatsApp is the last step, not the first.

An owner answering at 11 p.m. from their mobile is not “closeness”. It is maturity 1 with a notification. The pilot cannot rest on that.

Once a week, read out loud a ticket the draft would have sent badly. Not to humiliate: to adjust the rule. That fifteen-minute ritual prevents the “it is going fine” that precedes a public mess. The champion picks the ticket, not leadership.

Do not measure the support pilot by “messages generated”. Measure it by tickets closed well and by reopens. A prolific generator that reopens 30% is slower than the old inbox. The four-number sheet, applied here, stops you falling in love with the model’s prose.

Frequently asked questions

Can we connect the bot to the ERP to “look up the order”?

Not in month one. Mis-identifying the customer and leaking an order status is an incident. FAQ and drafts first.

What about the phone?

Post-call transcript and summary can help. An automatic agent that “handles” the customer in shaky Spanish, not at the start.

Do we unify WhatsApp, mail and web now?

Unify the queue (one place everything lands). Unify the model later. Mixing channels and model on the same day is chaos.

Next: a site that captures and internal copilots. Tell us channel and volume.