牙科集團 AI 接待員 0.6 秒回覆，首次回覆中位時間仍達 4 小時 12 分鐘

A dental group I worked with had an AI receptionist that answered in 0.6 seconds. Their median first response time, measured across all inbound enquiries, was 4 hours 12 minutes.

Both numbers were correct. The AI was fast. The business was slow, because the AI was only watching one door out of five.

The 0.6-second reply that nobody received

Here's the setup. They had a website chat widget with an AI agent behind it, and it worked well — answers in under a second, correct prices, clean escalation. Then we pulled the channel data for one week:

Fifty-eight percent of enquiries never touched the AI at all. They landed in a shared inbox that two people checked when they had time. The customer experience was: ask a question, hear nothing, wait until tomorrow.

The AI wasn't underperforming. It was solving 42% of the problem at world-class speed while the other 58% rotted in a queue nobody had instrumented.

First response time is a routing problem before it's a model problem. If you tune prompts for a month and never check which channels your agent actually covers, you will improve a number that isn't the bottleneck.

Layer 1 — Define the clock before you optimise it

Most teams can't state what FRT means in their business. Pick one definition and apply it everywhere:

• The clock starts at the customer's first message on any monitored channel — including at 11pm on a Sunday.
• The clock stops at the first substantive reply, not the auto-acknowledgement.
• Reply time is measured per enquiry, not per message. A thread with 12 back-and-forths is one enquiry.
• Off-hours enquiries count. Excluding nights and weekends is how companies convince themselves they're at 45 minutes when the customer experiences 14 hours.

Write those four rules down. If your team argues about them, you've already found your first problem: nobody knows what's being measured, so nobody is accountable for it.

Layer 2 — Map every door, then count it

List every place a human can start talking to you, and next to each one write who or what replies first. Do it on paper, in one sitting. Include the channels you think nobody uses.

The uncomfortable result is usually the same: the AI is on the channel you built it for, and your customers are on the channel you forgot. Instagram DMs and Facebook Page messages are the two most commonly orphaned — they were set up years ago for marketing, and now they're an unmonitored support queue.

Once the list exists, coverage becomes a menu:

1. Agent on the channel natively — best, and possible on WhatsApp and IG if you're willing to do the API work or use a platform that abstracts it.
2. Channel forwards into the AI's channel — a one-line auto-reply pointing to the AI conversation, plus a real answer to the same question pre-loaded.
3. Human-in-the-loop with an AI draft — the AI writes the reply, the human approves it in one tap. Slower than option 1, still 10x faster than typing.
4. Close the channel. If you can't cover it, remove it from your site and profiles. An unmonitored inbox is worse than no inbox: the customer assumes you're ignoring them.

Layer 3 — Never let the customer sit in silence

The single highest-leverage change is decoupling acknowledgement from resolution.

A customer who receives "Got it — checking this now, you'll hear back within 15 minutes" and then waits 20 minutes is calmer than a customer who waits 8 minutes in silence. The second one opens a competitor's site.

This is a copy-paste acknowledgement block. Adapt the wording, keep the structure — time promise, what happens next, one option to skip the wait.

Thanks for reaching out — this is on my screen now.

Here's what happens next:
• I'm checking your question against our current records.
• You'll have a real answer within [15 minutes / 1 hour / by 9am tomorrow].
• If it's urgent, call [number] and mention this message.

Nothing else needed from you right now.

Three things make it work: a specific time promise (not "shortly"), a named next step, and an escape hatch. Send it within 60 seconds, every time, on every channel — that's what the AI is for. It costs nothing and it converts waiting into patience.

The rule I use on deployments: no enquiry stays unanswered past 60 seconds, ever. The first message can be an acknowledgement. It can even be "I don't know yet." It cannot be silence.

Layer 4 — The after-hours lever nobody pulls

Here's the number that changes CFO conversations. In most service businesses, 35–45% of enquiries arrive outside working hours. For the dental group it was 39%.

So if your business hours are 9-to-5 with a human, you are structurally guaranteeing that nearly two out of five enquiries get their first response the next morning — a response time of 12 to 16 hours, no matter how good your daytime team is.

You cannot fix that with more staff at 3pm. You fix it with three moves:

1. Turn the AI on overnight. Same agent, same facts file, no human required. The overnight median drops from hours to seconds.
2. Set one honest overnight promise. "You'll hear back by 9am with the full answer" is fine — customers accept it, because it's specific and it's true.
3. Book, don't just answer. An agent that takes a booking at 11pm is worth more than one that politely says someone will call tomorrow.

The dental group turned the agent on for nights and weekends only. Nothing else changed — no new prompts, no new model. Median first response time went from 4 h 12 m to 22 minutes in the first week, and the overnight median went from 15 h 30 m to under a minute.

Layer 5 — Report it weekly, or it decays

FRT regresses. A new channel gets added, someone changes a routing rule, a prompt edit breaks the acknowledgement, and three weeks later you're back to silence on Instagram.

One weekly report, five lines, is enough:

Week of: ____
Total enquiries: ____
Enquiries answered under 60s: ____%   (target: 100%)
Median first response: ____          (target: under 2 min in hours, under 9am next day off hours)
Slowest channel: ____ at ____        (fix this one first)
Enquiries with no first reply: ____  (target: 0)

That last line is the one that matters. Every enquiry with no first reply is a customer who thinks you don't want their money.

The honest version of "our AI is fast"

A fast agent behind a door nobody watches is a demo. A mediocre agent that greets every enquiry within a minute on every channel is a business.

Test it today, and take ten minutes: open your site and every social profile you own, send yourself the same question from each one, and time how long until a real reply lands. Do it at 9pm, not at 2pm. Whatever number comes back is your actual first response time — and it's almost never the one in the dashboard.

The routing map, the acknowledgement templates, and the weekly report format are part of a kit I run on client deployments. Fill-in versions: Agent Audit Pack