Customers do not keep your hours. They message at 9:40 PM from the couch, when they finally sit down and remember the leak, the appointment, the noise the car started making. They send the same question to you and to the competitor whose name came up next to yours — and in the morning, the one who answered first usually has the job. That is not rudeness or impatience. It is just how a person with a problem behaves when two businesses look equally good and one of them is awake.
You cannot be awake for all of it, and you should not try. What you can do is make sure a night message gets an immediate, honest, useful answer to the routine parts — hours, prices you publish anyway, where you are, how booking works — and that everything else is captured properly and waiting for a human in the morning, with the customer’s name, number, and problem already written down. That is the whole job of an after-hours AI assistant. It is a narrower job than the sales pitches suggest, and that narrowness is exactly why it works.
Start with the boundary, not the tool
Before you pick any software, write down two lists. Do this on paper first; it takes ten minutes and it decides everything else.
The assistant may answer, on its own:
- Hours, location, parking, directions — the facts already on your Google profile
- Services you offer, in the words your customers use for them
- Published starting prices and what a quote depends on — never an invented exact price
- How booking, deposits, and cancellations work
- “Yes, we handle that” / “No, we don’t — here’s who does” for the clear cases
The assistant must never handle alone:
- Emergencies — anything involving safety, flooding, gas, a locked-out child, a sick animal. It gives the emergency number or instruction and stops.
- Complaints. A frustrated customer at night gets acknowledged, given a name and a time a human will call, and nothing argued with.
- Exact quotes, diagnoses, or promises about dates. It can say what a job typically involves; a human commits.
- Anything it doesn’t actually know. The correct night answer to an unusual question is “I don’t want to guess at that — here’s what happens next,” and then it takes the message properly.
If a tool you’re considering can’t be held to a list like this — if it wants to improvise answers to sound impressive — it is the wrong tool, whatever it costs.
| The assistant may answer alone | The assistant must hand to a human |
|---|---|
| Hours, location, parking, directions | Emergencies — give the emergency number or instruction, and stop |
| Services, and published starting prices | Exact quotes, diagnoses, or promises about dates |
| How booking, deposits, and cancellations work | Complaints — acknowledge, and name the time a human will call |
| Clear yes-or-no questions about what you handle | Anything it does not actually know — “I don’t want to guess” is the right answer |
The worked example: one night conversation, question by question
Here is the shape of a good exchange, for a fictional heating-and-air shop. Read it for what the assistant does not do, as much as for what it does.
Customer (9:42 PM): Do you service heat pumps in Gray?
Assistant: We do — heat pump service and repair across Gray and Washington County. Are you looking for a repair, or regular maintenance?
Customer: Repair. It’s blowing cold air.
Assistant: I’m sorry — that’s a cold night for it. I can’t diagnose it from here, and I don’t want to guess. What I can do is book you first thing: a technician can be scheduled in the morning, and dispatch calls by 8 AM to confirm the window. Can I take your name, the best number, and your address?
Customer: [gives details] Is it going to be expensive?
Assistant: Honest answer: it depends on the cause, and anyone who quotes it exactly tonight is guessing. Our service call is [published fee], which goes toward the repair if you go ahead, and the technician confirms the full price before any work starts. You’re booked for the morning call — if the heat fails completely tonight, call [number] and it reaches the on-call line.
Notice the moves: answer the factual question immediately and in plain words; show you heard the actual problem; refuse the two things a machine should not do (diagnose, quote); convert the rest into a concrete morning commitment with a time on it; and leave an emergency door open. A customer who gets that at 9:42 PM mostly stops shopping. Not because a robot charmed them — because their problem is now scheduled, and the competitor’s silence is not.
Set it up in this order
- Feed it only what’s already public. Hours, services, area served, published prices, booking rules — the same facts as your profile and your homepage. If those facts disagree with each other, fix that first; an assistant trained on a messy record just repeats the mess faster, which is the same consistency problem covered in how to show up when people ask ChatGPT for recommendations.
- Write the hand-off before the greetings. The morning summary is the product: name, number, what they need, what they were told, and anything unusual flagged. If that summary lands in the inbox or phone you actually check first, the system works. If it lands in a dashboard nobody opens, it doesn’t.
- Say what it is. The first line identifies it as the shop’s after-hours assistant and says when a human reads everything. Customers forgive a machine that is honest about being one; they do not forgive discovering they were fooled, and they tell that story in reviews.
- Test it like a difficult customer. Before it goes live, message it yourself at night for a week: ask the five questions your real customers ask most, then ask two it should not answer — an exact quote and a complaint — and confirm it hands both to a human instead of improvising. Fix the wording where it sounds stiff; a sentence you rewrite is cheaper than a customer you lose.
- Put it where the button already points. Article #6’s one-button rule says the homepage asks for one action. After hours, that same button is what reaches the assistant — the customer shouldn’t have to find a second, different way to reach you at night.
The morning habit that makes it pay
The assistant books the work; a human still has to close the loop. Fifteen minutes, first thing:
Copy this — the morning check:
- ☐ Read every overnight conversation summary, not just the count.
- ☐ Call the booked ones by the time the assistant promised — that promise is now yours.
- ☐ Scan for anything the assistant got wrong, sounded wrong on, or shouldn’t have answered. Correct the source fact or the boundary list the same day.
- ☐ Note the questions it couldn’t answer. The third time the same question appears, it probably belongs on the “may answer” list — with a human-written answer.
That last line is how the system gets better without getting riskier: it grows by your reviewed answers, not by the software’s confidence.
And one firm rule on reviews: never let the assistant ask for one. A review request belongs after real work, from a person, in the pattern covered in how to get more Google reviews without begging customers. An automated ask at the end of a night chat cheapens the whole thing.
The mistake that makes this fail
Turning it on everywhere, for everything, on day one — website, Facebook, text, reviews, quotes, scheduling — and never reading what it said. The businesses that get burned by these tools are almost never burned by the technology’s limits; they’re burned by scope. An assistant that answered one wrong price question to one customer has done more damage than a month of 7 AM replies ever did. Start with one channel (the one your customers actually message at night), the boundary lists above, and the morning check. Broaden only after a month of clean summaries.
The other version of the same mistake: using it to avoid customers instead of to reach them faster. If the night answer becomes a wall — no phone number, no morning commitment, no human name attached — customers can tell the difference between a business that answers while it sleeps and one that hides while it sleeps.
How long this takes / what it costs to DIY
An evening to write the two boundary lists and gather the public facts. Most website and Facebook messaging tools now include a basic assistant or connect to one; expect roughly $0 to test on one channel and, for most small shops, tens of dollars a month rather than hundreds — the honest variable is your setup time, not the subscription. Budget a second evening for the week of your own night tests, and fifteen minutes each morning after. If a quote for this comes back in the thousands with a long contract, you are being sold a bigger system than this article describes.
When to hire this out
Set it up yourself once if your setup is simple — one channel, a short service list, published prices. You’ll learn exactly what your customers ask at night, and that knowledge is valuable even if you hand the system over later. Hire help when your rules have real branches (multiple locations, emergency rotation, deposits that depend on the job), when the messages live in three places that all need to feed one morning summary, when a wrong answer carries real legal or safety weight, or when you’ve read the summaries for a month and know precisely what you want changed but not how to change it. And if the honest problem is that nobody reads the morning summaries at all, that’s a workflow fix with a name — it’s exactly the kind of thing we walk through on a first call.
AI FOR APPALACHIA
Book a Free Consultation
A consultation is a working session, not a sales pitch — and it’s free. On the call, we look at the messages your business actually gets after hours, draw the two boundary lists with you, and tell you straight whether a simple setup covers it or your rules need a custom build. No pitch, and what you do with it is up to you.
We talk first. If there’s work to do together, we’ll talk numbers on the call, once we understand what you actually need.
Book a Free ConsultationMore reading: All articles from Southern Marketing