The managing partner has a note on his phone with twelve "AI ideas." Intake. A voice thing for after hours. Something that "summarizes case law." A chatbot on the contact form. The vendor demo from last month. His cousin's dentist has one on the website. He is standing in the hallway between the reception desk and the conference room, trying to pick where to start, and every option looks equally shiny.
Meanwhile the intake coordinator is staring at a shared inbox. A consult from Tuesday still has no follow-up. The good ones have already hired someone. The twelve-item note is a vision exercise. The quiet consults are the job.
Start where the pain shows up every day, not where the demo looks cool. The first AI workflow in a small business is not a strategy offsite. It is a scoring problem. Frequency. Dollar impact. Messiness. Whether you, the owner, can see it without a report. One workflow, done, beats four half-started pilots. I'll walk you through the score, then through dental, HVAC, law, agencies, and multi-location retail so you can steal the pattern, not a slogan.
The score, not the vibe
Take a legal pad. List the jobs that ate last week. Not "marketing." Actual jobs. Confirming hygiene. Chasing unclosed estimates. First-pass intake. Writing the proposal after the call. Compiling store notes on Sunday night.
For each job, mark four things. Frequency: does this happen every day, or a few times a year? Daily wins. Annual "we should automate onboarding" loses for round one. Dollar impact: if this job improves, does money show up — more booked work, fewer leaks, capacity you can sell — or is it just tidier? Effectiveness first. Messiness: is the job ugly enough that a first draft would help a human, or is it already a clean checklist? Clean checklists want a rule, not a model. Ugly, repetitive writing and routing want a helper. Owner-visible: will you notice in two weeks if it's working? If the only proof lives in a staff member's head, you will forget to protect the change.
Circle the job that is high frequency, real dollars or real hours, messy, and visible. That's first. Everything else goes on a later list. The later list is not failure. It is how you avoid starting four things and finishing none. Efficiency (time) and Quality (customer experience) still count. A daily mess that burns the coordinator and makes patients repeat themselves can beat a slightly bigger dollar item you only touch at month-end. The filter is still: money, time, or the customer. The score is how you pick which of the passing ideas goes first.
Hypothetical (not a client, not a promise): if unclosed HVAC quotes sitting four days is the pile, even a faster first draft plus a same-day follow-up is a revenue play. Say four quotes a week sit, average ticket $4,200, and a live same-day nudge closes even one extra a month. That's a napkin. If "rewrite the about page" is the pile, that's a rainy Saturday. Run your unclosed pile. I will not invent a shop that "recovered $X with AI."
Use this to score the list before you buy anything:
Score these jobs for a first AI workflow. Shop: [type].
Jobs from last week:
[paste 5–8 real jobs, not slogans]
For each job, score 1–5 on:
- Frequency (daily = 5)
- Dollar impact (booked work / leaks / sellable capacity)
- Messiness (ugly writing or routing a first draft would help)
- Owner-visible in two weeks (a number I can see)
Return a table and circle ONE first workflow. Put the rest in a parking lot.
Do not recommend tools. Do not invent jobs I didn't list.
Dental: what first usually is
Not the website chatbot. Not "AI that does perio charts." Those demos win conferences. Usually first: recall and confirm. It happens every day. Empty chairs are money you already paid to acquire. The messages are repetitive and easy to get wrong in tone. You can see openings on next week's schedule with your own eyes.
Runner-up: insurance follow-up notes and no-show calls, if that's what eats the first ninety minutes. Still daily. Hours back (Efficiency). Patient isn't calling twice for the same answer (Quality). Human still sends. What I would park: a custom "AI receptionist that also files claims." That's many jobs glued together. You will drown. Also park anything that wants to post a Google review for a patient. You never fabricate reviews. Draft a request at the right time, after a good visit, in your voice. A person sends it.
Map the real path. Phone rings or list prints. Who looks at the schedule. What they say. Where they get stuck. Last week's actual no-show list, not a sample. Then a helper for the coordinator only. She beats it up for a week before anyone else touches it.
HVAC: what first usually is
Not route optimization as a product. Not a brand-new "AI that sells the replacement." Usually first: missed call to booked diagnostic, or first-pass estimate from the tech's notes. Both are daily in season. Both leak money. Both are messy writing plus routing. You can see the board: calls that died, quotes that didn't go out.
If your after-hours voicemail is a black hole, that's Effectiveness. Demand you already paid for, sitting in a recording. If estimates take two days because someone has to type in the evening, that's Effectiveness and Efficiency. The homeowner cooled off. The tech is doing office work in the driveway. Quality: the estimate is complete the first time. No "misc. materials" that triggers a confused call. A human checks the number before it goes.
Park the thing that tries to change dispatch, billing, reviews, and sales in one install. That's too many people and too many handoffs, and then the shop says "AI doesn't work." Start easy: one estimator or one dispatcher, one loop, real jobs from last Tuesday.
Law, agencies, and multi-location retail
Law: not "summarize all case law." Not a bot on the contact form that pretends to give legal advice. Usually first: intake to first follow-up. Frequency is every new inquiry. Dollar impact is consult-to-retain. Messiness is the story arriving by voicemail, web form, and a napkin from the receptionist. Owner-visible is the pipeline of consults that went quiet. The helper drafts the recap and the follow-up from the notes. A human — intake coordinator, paralegal, or attorney — sends it. Same day. Nothing customer-facing leaves without a person. Nothing that sounds like legal advice gets to a prospect because a model was confident. Time entry and first-draft letters are good second workflows. They win Efficiency. They are not first if matters are going cold.
Agencies: not "we'll generate 40 ads a day." Faster creative on a muddy offer is still a muddy offer. Usually first: call or notes to first-draft proposal, or follow-up on the open pipeline. Daily or near-daily if you're selling. Dollars on close speed. Messy because every draft starts from scratch even when the offer isn't new. Visible: proposals out, stalled deals touched. Give the helper the offer sheet, the "claims we never make" list, and one good proposal. The account lead edits. The account lead sends. Quality is not inventing a metric in the deck. Park the full content machine and the custom tool you were going to build for clients. Your first workflow is your shop, not a product.
Multi-location retail: not a fine-tuned model. You don't need that. Usually first: complaint routing, review replies (human sends, never fake), or shift notes the next manager can use. These happen every day across stores. Dollars show up as retained customers and fewer "I asked store two and they said the opposite." Hours show up as Sunday-night reporting that stops being copy-paste. Start in one location. The store lead is the first user. Copy the loop after it works. Rolling out seats to every manager on day one is how you get twelve calculators and the same messy close.
How you run the first two weeks
Write the job as it really happens. Cut duplicate steps before you add a helper. A 15-step estimate path that is actually 8 steps should become 8 on the whiteboard first. Automating a bloated path makes a faster blob.
Feed real examples. Last week's list, last Tuesday's voicemails, the three texts you were proud of, the one you regretted. One person uses it. They correct it like a new hire. You save the corrections. That file is how it stops repeating the same mistake. Human sends anything a customer, patient, or client will see. Pick one number. Two weeks. If it didn't move, you learned. You don't add workflow two to cover the embarrassment. You fix workflow one or you kill it.
Refuse anything that needs five departments. Anything whose demo required a perfect script. Anything you can't score on money, time, or customer experience in one sentence. Anything that sends without a human. Anything that fabricates social proof. Anything whose owner is "we'll all share it." Also refuse "we'll start with a custom build." Most shops need a workflow and an off-the-shelf tool with context, not a product. Custom is later, and only when the shop's way of working is the advantage.
FAQ
What should the first AI workflow in a small business be? The daily, messy, dollar-touching job you can see without a report. In most shops that's recall, missed-call capture, intake follow-up, proposal drafts, or complaint routing — not a website chatbot.
Can I start two workflows at once if both score high? No. One workflow, done, beats two half-started pilots. Put the second on a later list. The later list is how you finish.
How do I know it's working after two weeks? You picked one number in week one: openings filled, quotes out, follow-ups sent, complaints closed. If that number didn't move, fix the loop or kill it. Don't add a second workflow to hide the miss.
If you want this done with someone who will make you circle one job and put the rest in a parking lot, that's the free 15-minute Mini Margin Check. We score frequency, dollars, mess, and visibility on your vertical using the Three Outcomes filter. You leave with one workflow chosen and mapped, not a stack of ideas. Bring last week's pain. Bring the person who actually does the job if you can. Leave the vendor PDFs in the car.