The hygienist is in operatory two, running late. The coordinator is at the front desk with a vendor PDF still open on her screen — "AI receptionist that also handles recall, reviews, and insurance." The demo was slick. The monthly price is small enough that you could put it on a card and forget it. She's looking at you, because you're the owner, and someone at the study club said they "have AI now."
You already know how this goes. Last year it was the review widget. The year before it was a chatbot that sat on the website and answered "what are your hours" while new-patient forms leaked into a shared inbox nobody owned. The chairs still sit empty on Thursday afternoons. The insurance pile still eats the first ninety minutes of her day.
If that tool does not make you money, save you real hours, or make the patient's experience better, you don't need it. Those are the only three doors. I don't care how pretty the dashboard is. I don't care that the other dentist "has AI." I care whether it moves one of those three numbers in your shop. I call them Effectiveness, Efficiency, and Quality. Every tool and every "we should try this" idea has to pass at least one. If it can't, skip it. Without that filter you will spend a year collecting logins.
Effectiveness: does it make money?
Effectiveness is not "we feel more modern." It is cash, booked work, close rate, or capacity you can actually sell. If you cannot name the number, you do not have an Effectiveness play. You have a vibe.
In an HVAC company, Effectiveness looks like a missed after-hours call becoming a booked diagnostic instead of a voicemail that sits until morning. The estimate that used to take two days to type goes out the same afternoon, while the homeowner is still looking at the broken unit. Follow-up on unclosed quotes happens on a clock, not when the owner remembers. None of that is a chatbot personality. It is more jobs from the demand you already paid to generate.
In a dental office, Effectiveness is recall that actually lands, new-patient forms that don't leak, and the hygiene chair that doesn't sit empty because nobody confirmed. If a tool helps you fill a hole in next week's schedule, that's money. If it writes a blog nobody reads, that's a hobby. In a law firm, Effectiveness is consult-to-retain. Intake that captures the facts so the attorney can say yes or no faster. A follow-up that goes out the night after the consult, not three days later when the prospect has already hired someone else.
Agencies win Effectiveness on proposal speed and follow-up on the pipeline they already have. Multi-location retail wins it on the local page, the review reply, and the shift note that stops a stockout from becoming a lost basket. Same test in every shop. Does revenue move, in a window you can see?
Hypothetical (not a client, not a promise): say an HVAC shop loses four after-hours calls a week, and even one of those would have been a $280 diagnostic. That's a napkin, not a case study. If a missed-call workflow captures part of that, it pays. If a "content AI" writes posts that don't book a truck, it doesn't. Run your numbers.
Efficiency: does it give hours back?
Efficiency is time. Not "the AI is fast." Hours your people get back, fewer repeats, a shorter cycle on work you already do. If you cannot name the hours, the person, and the task, you don't have an Efficiency play. "We'll save time" is a wish.
The office manager in a dental practice who spends the first ninety minutes of the day on insurance follow-up and no-show calls is the person I want to talk to. If a helper drafts those calls from yesterday's schedule and she edits and sends, that's Efficiency. If she still has to hunt the same four screens for the same four facts, you bought a faster way to hunt. HVAC dispatch is the same story in a different costume. If the board is a group text, a whiteboard, and a tech calling in from the driveway, AI will not magically "optimize routes" until you write down how a job actually gets assigned.
Law firms burn hours on time entry, intake notes, and the letter that is 80% the same as last month's letter. The associate or paralegal still owns the file. The helper does the first pass. If your "AI" requires a forty-minute prompt to spit out something you'd rewrite anyway, it failed Efficiency. You added a step. Agencies burn hours in research-to-first-draft and in status updates clients didn't ask for in that format. Multi-location retail burns hours in weekly reporting that is copy-paste between stores.
A real Efficiency sentence sounds like this: "The coordinator spends Tuesday morning on recall. We cut that in half without dropping the confirmation rate." That is a project. "We'll save time with AI" is a slide. Write the sentence before you buy the seat.
Quality: is the customer's experience better?
Quality is the one owners skip because it doesn't show up as a line item. It is fewer errors, more consistent service, less "depends who was working today." If a tool makes you faster at producing slop, you failed Quality. Faster mess is still mess.
In dental, the recall text sounds like your office, not a spa in another state. The new-patient packet doesn't ask for the same insurance card twice. The review reply is on-brand and a human still sends it. You never fabricate a Google review. Not for a ranking, not for a demo, not "just this once." Quality includes not lying. In HVAC, the estimate is complete the first time. The homeowner isn't calling to ask what "misc. materials" means. The tech's notes from the last visit are in the next tech's pocket. A voice agent that cuts the customer off or books the wrong window is a Quality loss, even if it "saves" a minute at the front desk.
In a law firm, intake is complete. The client isn't repeating their story to three people. Engagement letters don't promise things the firm doesn't do. Anything that goes to a client or the other side still leaves a lawyer's desk. AI that sends a message because it "sounded confident" is how you get a bar complaint, not a productivity win. Agencies: the deck doesn't invent a metric. The ad copy doesn't claim a result you can't stand behind. Multi-location retail: the complaint gets the same fair treatment at store three that it got at store one. Quality is consistency the customer can feel.
How you actually use the filter
Take the tool sitting in your inbox. Write three lines on a legal pad. If you cannot fill one line with a straight face, don't buy it. If you can fill one, keep going. If you can fill two, that's a real candidate. Three is rare. I don't require three. I require one, honestly.
Paste this into whatever chat tool you already pay for, then answer it out loud before you enter a credit card:
You are helping me decide whether to buy an AI tool for my shop.
Shop type: [dental / HVAC / law / agency / multi-location retail]
Tool name and what the vendor claims:
The job this is supposed to touch:
Answer in three lines only:
1. Effectiveness — what money number moves, and in what window?
2. Efficiency — whose hours, which task, how often?
3. Quality — what does the customer notice, and what could go wrong if the model is wrong?
If you cannot fill a line with a specific number or person, write "FAIL" on that line.
Do not recommend the tool. Do not suggest alternatives. Score the three doors.
Then ask the ugly questions the prompt will not ask for you. Who owns this on Tuesday? If the answer is "we'll all use it," nobody will. Pick a person: the office manager, the dispatcher, the intake paralegal, the store lead. What does "done" look like in two weeks? Not "we adopted AI." A number: unclosed quotes touched within 24 hours. Hygiene openings filled. Consult follow-ups sent the same day. Complaint tickets with a same-day human reply.
What happens when it's wrong? Because it will be. A human sends the customer-facing message. A human checks the estimate before it goes. A human does not let a model invent a review, a medical claim, or a legal conclusion. What are you willing to stop doing? A new tool on top of the old mess is a new mess. If the filter says "this step shouldn't exist," delete the step. Don't automate it.
What usually fails, and what usually passes
Conference-booth tools that promise a full "AI receptionist that also does billing, reviews, and marketing" fail because that is a catalog, not a filter. Seats for the whole office with no workflow attached fail because ChatGPT on eight phones is a calculator farm. A $20 login is not a plan. Anything that requires you to become a prompt hobbyist after 8 p.m. fails Efficiency. If the only person who can run it is you, you bought a second job.
Anything that wants to email patients, clients, or homeowners without a human in the loop fails Quality and liability. Drafts are fine. Send is a person. Anything whose only proof is someone else's dashboard screenshot fails honesty. Hypothetical math on your shop is more honest than a vendor case study you can't verify.
What usually passes is one hated loop. High frequency. Dollar impact you can sketch on a napkin. Messy enough that a first draft would help. Visible to you as the owner so you'll actually protect the change. Dental recall and confirm. HVAC missed-call to booked diagnostic. Law intake to first follow-up. Agency proposal first draft from a call recording. Retail review replies and shift notes. Those are not magic. They are boring, daily, and they hit the filter.
Do this on a legal pad before you buy anything
List the five things that ate last week. For each one, mark money, time, or customer experience. Circle the one that shows up every day and that you, the owner, can see. That's the candidate. Everything else is a later pile.
The owner who gets this right is not the one with the most AI. It's the one who can say: this makes us money, this gives us time, or this makes the customer feel the difference. Everything else is garnish.
FAQ
How do I choose AI tools for a small business without wasting a year? Run every purchase through Effectiveness, Efficiency, and Quality. If you cannot name a money number, a person whose hours come back, or a customer-facing change, don't buy it. One honest door is enough. A dashboard is not a door.
Do I need different tools for dental, HVAC, law, an agency, or retail? You need different jobs, not a different philosophy. Recall, missed-call capture, intake follow-up, proposal drafts, and review replies all pass the same three doors. The vertical changes the example. It does not change the test.
What if a tool fails two doors but the vendor swears it "saves time"? If you cannot name whose hours, which task, and how often, it failed Efficiency too. "Saves time" without a person and a task is marketing. Skip it.
If you want a room where we run this on your vertical — dental, HVAC, law, agencies, or multi-location retail — that's the free 15-minute Mini Margin Check. You leave with the Three Outcomes filter and one process mapped, not a stack of logins. Bring the tool you almost bought. We'll put it through the three doors. If it survives, you'll know why. If it dies, you saved a year of a subscription.