What "AI-powered" should actually mean for a small shop
The pitch problem
Every small business owner I talk to has been pitched "AI" by now, and most of them are quietly suspicious. They're right to be. The standard pitch is autonomy: a bot answers your phone, books your jobs, replies to your customers, sends your invoices. You stop doing the boring parts.
The trouble isn't that it doesn't work. It's that it works about ninety percent of the time, and the other ten percent quietly costs you a customer you never hear from again. A bot that quotes the wrong price on a job, or promises a Tuesday you don't have, doesn't announce the mistake. You find out when someone doesn't come back.
For a shop with four employees, that math is much worse than it is for a company with a support department to absorb the damage. Small businesses don't need AI that acts. They need AI that prepares.
The model I actually run
I ship mobile apps solo using a fleet of AI agents — I wrote about how that works in more detail. The short version: agents do the volume, I do the judgment, and nothing reaches production because an agent said it was finished. It ships because I checked.
Four rules make that safe, and all four transfer directly to a business that has nothing to do with software.
Route work by cost. Not every task needs the expensive model. The cheap one handles the routine eighty percent — sorting, drafting, summarizing, the tenth near-identical reply. The expensive one gets the exceptions, where being wrong is costly. Most vendors bill you for the expensive model on every task because it's simpler for them.
Draft, don't send. This is the whole thing. The system writes the reply, fills the quote, drafts the follow-up — and a human presses send. You keep the ten seconds of review and lose the ten minutes of composition. Nothing customer-facing goes out unread by a person.
Corrections become permanent rules. When it gets something wrong, you don't just fix that one instance. The correction gets written down as a rule the system loads every time after. Tell it once that you don't work on diesels, or that the Callander job site is an hour out and gets priced accordingly, and it stops making that mistake. The competence accumulates in the setup, not in the model.
Sandbox anything that touches money. Reading your calendar is safe. Writing to it is not. Drafting an invoice is safe. Sending one is not. The boundary is drawn at actions you'd want to undo.
A worked example
I've built sites for a barbershop here in North Bay and an auto shop, so this is the shape of business I have in mind: five or six people, phones ringing during work you can't step away from, one person doing quoting and scheduling and invoicing in the evenings.
To be clear about what follows — this is a proposal, not a case study. I'm describing what I'd build, not reporting numbers I've already delivered.
The actual problem is not "we need AI." It's that the phone rings while everyone's hands are busy, and the person who calls a second shop gets served first. The evenings disappear into quoting and chasing invoices. That's the pain worth spending money on.
What I'd build:
Inbound calls that aren't answered go to a transcription. Within a minute, the owner gets a message on their phone: who called, what they want, how urgent it sounds, and a drafted reply ready to send — or a drafted quote if there's enough detail to price it. One tap sends it. If the caller is an existing customer, the draft already knows their history.
Emailed and web-form inquiries get the same treatment, sorted by whether they're a real job, a price-shopper, or a supplier. Overnight, the routine ones already have drafts waiting.
At week's end, unpaid invoices get drafted follow-ups, in the tone the owner actually uses, ready to review as a batch.
The arithmetic — illustrative, and worth checking against real numbers before anyone spends money on it. A shop missing five calls a week that converts two more of them, at a $400 average ticket, is $800 a week. Recovering an hour a night from quoting and invoice chasing is worth whatever an hour of the owner's evening is worth, which most owners undervalue badly.
What it costs
Running costs are lower than people expect, because the cheap model does most of the work. For a few hundred inquiries a month, the model usage runs roughly $15–40. Hosting is essentially free at this scale. There's no per-seat SaaS subscription, because this isn't a product — it's plumbing between tools the business already pays for.
The real cost is the setup: learning how the business actually talks to its customers, writing the rules down, and wiring it into the phone, email, and calendar that already exist. That's one engagement, not a subscription that escalates every year.
The point of building it this way is that the business owns it afterward. No vendor holding the data hostage, no renewal negotiation, no platform deciding to triple the price once you depend on it.
What I would not automate
Anything that commits you. Booking, pricing, promising a date. Draft it, don't send it.
Anything with a regulatory surface. Health information, financial advice, employment decisions. Different rules, different risk, out of scope for a system like this.
The relationships. The reason people call a small shop instead of a chain is that they get a person. Automating that away is competing on the one axis where you're already losing. Automate the paperwork so there's more time for the part that isn't automatable.
The judgment calls. The system should surface the odd inquiry, not decide it. If it can't tell whether something is routine, that's the signal it goes to a human — and the failure mode should always be "ask," never "guess confidently."
Where this breaks
It breaks on vague setup, the same way my agent fleet breaks on vague specs. If nobody can articulate how the business prices a job, the system can't draft a quote — it'll invent one that sounds plausible. Most of the real work in a project like this is writing down decisions the owner has been making by instinct for fifteen years.
It also breaks if the owner stops reading the drafts. The whole safety model rests on a human gate, and a gate nobody stands at is just a slower way to send the wrong thing. Any setup where the review step feels like friction will get skipped within a month, so the review has to be genuinely fast — one screen, one tap — or it isn't sustainable.
Agents for the paperwork, a human on everything that commits you. "AI-powered" should mean your evenings back, not your judgment outsourced.
I'm Greg Roy — I run Roy Systems Consulting out of North Bay, building practical AI and automation into the tools businesses already use. If this sounds like your shop, get in touch.