How Our AI Email Autoresponder Actually Works

In a previous post, we talked about how we automated our email inbox — classifying, labelling, and triaging every message so our office manager only sees what actually needs attention. That system saved us about 25 minutes a day.

But sorting emails is only half the problem. The other half is responding to them. And a surprising number of those responses are the same thing, worded slightly differently, sent fifty times a month.

"Do you do Saturday appointments?" Yes. "What time do you open?" 8am. "Can I get a quote for a crown?" Here's what to expect, please call to book. "Please find attached our updated price list." Thanks, forwarded to the practice manager.

So we added auto-response capabilities to our email AI. Not canned templates — actual contextual replies that read the email, understand what's being asked, and respond appropriately. Here's exactly how it works, what it responds to, what it never touches, and the safeguards we built in to make sure it doesn't go rogue.

The five-step process

Every email that arrives goes through the same pipeline:

1. Classify the email — the AI reads the message and determines what type it is. Appointment request, general enquiry, supplier correspondence, spam, personal, insurance, complaint, clinical question. This classification drives everything that happens next.
2. Decide if auto-response is appropriate — not every email should be auto-responded to. The AI applies a set of rules about which categories are eligible for automatic replies and which must always go to a human. More on this below.
3. Pull relevant context — for eligible emails, the AI gathers the information it needs to respond. Practice hours, available appointment slots, service descriptions, pricing guidelines, policies. It's not guessing — it's pulling from a knowledge base we maintain.
4. Generate the response — this is where it gets interesting. The AI doesn't slot values into a template. It actually reads what the person wrote and writes a response that addresses their specific question. If someone asks about Saturday hours and also mentions they're a new patient, the response covers both.
5. Send or queue — depending on how confident the AI is in its classification and response, it either sends immediately or queues the draft for a human to review before sending. The confidence threshold is the most important safeguard in the system.

What it auto-responds to

We've tuned this over months of real-world use. The AI handles three main categories automatically:

Appointment requests

Someone emails asking to book, asking about availability, or enquiring about a specific service. The AI responds with something like:

"Hi Sarah, thanks for your enquiry! We have availability this Thursday afternoon and Friday morning. Would either of those work for you? You can book online at [link], call us on [number], or just reply to this email with your preferred time and we'll get you sorted."

It checks actual availability before suggesting times. It doesn't promise slots that don't exist. And if the request is for something complex — a full treatment plan, a specific clinician, a procedure we don't commonly do — it queues for a human instead.

General enquiries

Questions about hours, location, parking, services offered, what to bring to a first appointment, whether we accept a particular health fund. The AI answers the specific question asked, not a generic dump of practice information.

"Hi Mark, yes we're open Saturdays from 8am to 1pm. Would you like to book a Saturday appointment? Here's our online booking link, or call us on [number] and we'll find a time that works."

Simple, direct, helpful. The same response a receptionist would write, just sent in seconds instead of hours.

Supplier correspondence

Updated price lists, delivery confirmations, account statements, order acknowledgements. The AI sends a brief acknowledgement and forwards the email to the practice manager.

"Thanks David, received. I've passed this on to our practice manager for review."

Nothing fancy. Just a professional acknowledgement so the supplier knows their email didn't vanish into the void. It's a small thing, but suppliers appreciate knowing their correspondence landed.

What it NEVER auto-responds to

This is the more important list. Getting auto-responses right is good. Getting them wrong is catastrophic. So we drew hard lines:

• Patient complaints — always flagged for a human response. A complaint deserves genuine attention, not an AI-generated platitude. The office manager or practice owner responds personally, every time.
• Clinical questions — if a patient describes symptoms, asks about a diagnosis, or enquires about a treatment outcome, it goes straight to a clinician. The AI is not a dentist and it absolutely should not pretend to be one.
• Legal and insurance correspondence — anything from a solicitor, an insurer, a regulatory body, or the ADA gets flagged for the practice manager. Zero tolerance for AI involvement in anything with legal implications.
• Anything ambiguous — if the AI isn't sure what category an email falls into, it doesn't guess. It queues the email for human review and moves on. False confidence is worse than no response at all.

We'd rather have a human respond to an email that could have been automated than have the AI respond to something it shouldn't. That's a deliberate design choice, not a limitation.

The safeguards

Giving an AI the ability to send emails on your behalf is a big deal. We took it seriously. Here's what keeps things safe:

Confidence threshold

Every classification and response gets a confidence score. If the AI is less than 80% confident in either its classification or its drafted response, the email gets queued for human review instead of sent. In practice, this means about 15% of auto-response candidates end up in the review queue. We'd rather over-queue than over-send.

Daily summary

Every morning, the office manager gets a summary of everything the AI sent the previous day. Every auto-response, the original email, and the reply. She scans through it in about two minutes. If anything looks off — wrong tone, missing information, a classification error — she flags it and we adjust the system. This has caught exactly three mistakes in six months, all minor.

Category overrides

We can flag any email type or sender as "always queue for review." If we take on a new supplier and want to personally handle the first few exchanges, we add them to the override list. If we run a promotion and expect unusual enquiries, we can temporarily queue all appointment-related emails for human review. The system is flexible, not rigid.

Tone matching

The AI doesn't write like a corporate helpdesk. It writes like our practice actually communicates — friendly, professional, concise. We spent time calibrating this, giving it examples of real responses our team has sent, so the output feels like it came from a person who works here. Because the worst auto-response isn't a wrong one — it's one that sounds so obviously robotic that the patient feels dismissed.

Real examples

Here are some actual exchanges, anonymised:

Incoming: "Hi, I'm looking to get my teeth cleaned. Do you have any appointments available next week?"

AI response: "Hi [name], thanks for getting in touch! We have availability next Tuesday at 10am and Thursday at 2:30pm for a clean and check-up. Would either of those suit? You can also book online at [link] or give us a call on [number]. Looking forward to seeing you!"

Incoming: "Please find attached our updated price list effective 1 March."

AI response: "Thanks [name], received. I've forwarded this to our practice manager for review."

Incoming: "I had a filling done last week and it's still quite sensitive. Is this normal?"

AI response: None. Queued for clinician review with a "clinical question" flag. The dentist responded personally within the hour.

That last example is the one we're most proud of. The AI recognised a clinical question, knew it shouldn't touch it, and routed it to the right person. That's the system working exactly as designed.

The time savings add up

Combined with the inbox triage system, our email workflow has gone from 30+ minutes a day to about 5 minutes reviewing flagged items and the daily auto-response summary. The AI handles roughly 60% of incoming emails completely autonomously — classified, responded to, filed, done.

For the remaining 40%, it still does the classification and prioritisation. The human just writes the response. Even those emails are faster to deal with because the context is already pulled up and the priority is already assigned.

As we covered in our post on what AI can actually do for small businesses, the value here isn't replacing people. It's giving your people back the time they're currently spending on repetitive tasks so they can focus on work that actually needs a human brain.

Could this work for your business?

If your team spends more than 15 minutes a day responding to routine emails — the same questions, the same acknowledgements, the same replies worded slightly differently each time — then yes, this would probably save you a significant chunk of time.

The system is built to be adapted. Different businesses have different categories, different rules about what's safe to auto-respond to, and different tones. A law firm's auto-responses sound very different from a dental practice's. But the underlying logic — classify, decide, draft, safeguard, send — is the same.

If you're curious about what this would look like for your business, get in touch. We'll walk through your current email workflow, identify what could be automated, and talk through the safeguards that would make sense for your situation. No obligation — just a practical conversation about what's possible.