The agency lead wants a portal. An "AI OS" with the shop's logo, a client login, and a roadmap. His own Friday status notes still start from a blank page. The PM tool they already pay for has the latest deck. The client could have read an email. Someone is quoting a custom build anyway, because custom looks like a product and a $20 tool plus a one-page SOP does not.
If you own a dental practice, an HVAC company, a law firm, an agency, or a handful of retail locations, someone will eventually try to sell you that build. A portal. An agent. An app with your logo. Most of the time you should say no.
Most shops need a workflow, not a product. A $20 tool plus a written way of working will beat a six-figure science project that nobody opens on Tuesday. I filter it the same way every time. Effectiveness: does this make you money? Efficiency: does this save hours you will actually use? Quality: is the customer better off, or just more automated? If the custom build cannot name which of those it is buying, it is a wishlist item. Custom AI software for a small business is rarely the first honest card on the page.
Buy and configure first
The default is: use the software you already pay for, or a cheap tool that already exists, and write the steps.
HVAC: your dispatch board, your estimate software, a $20 chat tool, and a script for first-draft estimates. That is a system. A custom "AI dispatcher" is a product someone wants to bill you to maintain. Dental: your PMS already has recall lists, appointment status, and a texting add-on you half-use. Configure the list. Write the recare script. Have a model draft the text. A human sends. You do not need a custom patient-communication platform. Law: your practice system already stores matters. Intake is a form plus a voicemail plus a human. A custom "AI associate" that files documents is how you buy a malpractice story. Buy a transcription tool. Buy a better intake form. Write the checklist.
Agency: you already have a PM tool, email, and a drive. The Friday status note is a prompt plus a folder plus a human. You do not need a custom client portal that "uses AI" to tell the client what they could have read in an email. Multi-location retail: POS, inventory, a shared sheet, store-level texts. The Monday out-of-stock report is a workflow. A custom "retail brain" is a vendor with a pitch deck. When the step is generic — transcribe this, draft that, remind them, tag the email — buy the tool. Build only where the shop's way of working is the actual advantage, and only after a cheap version has proven the loop.
The $20 tool plus an SOP is a real strategy
I mean that literally. ChatGPT or Claude at the paid seat. A one-page SOP. Three examples of good work. The decision rule: human sends.
Hypothetical (not a client, not a promise): an HVAC office manager spends 45 minutes turning a tech's job notes into a customer-facing estimate email. You write a one-page SOP: paste the notes, attach the price list rules, never offer a discount, always include the permit line if it's a change-out. The model drafts in four minutes. She edits for five. You just bought back half an hour a job without a developer. If she runs 8 of those a week, that's hours on a napkin, not a case study. You still have to use the time — on more estimates sent, or on going home. I will not pretend I already ran this in your shop.
Dental, same shape: review replies sit for three days because nobody wants to write them. SOP: pull the review, classify (praise / specific complaint / legal-adjacent), draft a reply in our voice, never invent a clinical claim, office manager hits send. That's Quality plus Efficiency. It is not software you own. It is a way of working. Law: consult recap emails. SOP: bullets from the meeting, no legal advice in the recap, list the documents we still need, paralegal sends. Custom software does not make that safer. A rule does. If the SOP is ugly and people still use it, you have a loop. Then you can talk about wiring it into the PMS or the estimate tool. If the SOP is ignored, a custom app will be ignored with a nicer login screen.
Wishlist hygiene: most "AI projects" are ordinary software
This is the conversation I have in the first hour. The owner has a list. Half of it is not AI.
Typical dental list: online forms that don't dump into the PMS, a waitlist that actually texts the next patient, a way to see who is overdue that isn't three spreadsheets, automatic reminders, a dashboard of production. That's software and process. Some of it you already bought and didn't turn on. Typical HVAC list: missed-call capture, a map of who's closest, estimates that leave the truck the same day, a parts list that matches what you actually stock, follow-up on bids older than a week. Again: telephony, dispatch, templates, a CRM field called "next action." AI might draft the follow-up. It will not invent a definition of a lead.
Typical law list: intake that doesn't fall into a void, time entry that isn't a Sunday night archaeology dig, document naming, a reminder when a statute date is near. Checklists, fields, calendar rules. An agent on top of a messy matter file will be a confident mess. Typical agency list: a place where the latest deck lives, status the client can see, a QA pass before anything ships. Project management. Maybe a model on the first draft. Not a custom model. Typical multi-location retail list: which store has the SKU, who called out, what sold on the promo, a clean way to push a local page update. Inventory and a content calendar. Not a model.
Sort the wishlist into four buckets before you spend. Plain software or a setting you already own. Turn it on. Stop calling it AI. One-person helper: a model plus a SOP plus a human. Do this next. Later: multi-person, multi-system, after the one-person version works. Never: anything that auto-files, auto-diagnoses, auto-emails a patient, or auto-publishes a legal letter. Naming the buckets saves you from funding a science fair. It also saves me from quoting a build I shouldn't sell.
When custom is actually justified (rare)
I'm not allergic to building. I'm allergic to building first.
Custom starts to make sense when all of this is true. You already run the loop by hand with a cheap tool and a SOP, and the bottleneck is now the clicking, not the thinking. The shop's way of working is the advantage — HVAC who quote a certain way that a generic estimator keeps flattening, a dental group with three locations and a recall rule-set no vendor will encode, an agency with a QA gate that is the product. The data already lives in one place you can actually connect, or you are willing to clean it. Custom on top of four spreadsheets and an owner's head is still a mess. A human checkpoint is designed in. No unsupervised run that can email, file, charge, or diagnose. You can name the scoreboard: hours back, slower hiring, or a revenue/quality number. Not "we'll see." You have someone who will own it after month one. If that's not you and it's not me on a retainer, the app will rot.
Even then, start with the smallest spine. A button that pulls today's list and dumps a draft into a doc the office manager already uses is a build. A platform with users, permissions, a mobile app, and a roadmap is a company you don't want to run. A four-location dental group that has proven recare drafts in chat for six months, where the pain is now "we paste from three PMS exports," might justify a small integration that parks drafts in a queue for human send. A "patient engagement OS" with a vendor CSM would not. An HVAC company with a first-draft estimate SOP that works might justify reading job photos from a folder you already use and spitting a draft into the estimate tool the tech still confirms. An autonomous agent that books the install and orders parts would not. Law firms: custom that touches filings, advice, or anything a court will read is a no unless you have a process a partner will sign in blood and a human still files. I'd rather you buy better intake and a checklist.
The Margin Scan should say "no new tool, just fix the process" when that is the truth. It should say "buy and configure" when that is the truth, and recommend a small build only after the cheap version works. An embedded operating relationship in the tools you already use is often more honest than a custom quote designed to look impressive. Efficiency includes not paying for software you will not maintain. Effectiveness includes not tying up cash in a portal while missed calls still hit voicemail. Quality includes not letting a half-tested agent become the customer experience.
Write the project in one sentence. Then run this test:
Custom-build smell test. Answer yes/no. Do not design an app.
Project in one sentence:
1. Can a $20 seat and a one-page SOP get us 70% of this in two weeks?
2. Does this already exist in the PMS, CRM, dispatch tool, or POS we pay for?
3. Are we asking for AI because the last software install failed?
4. Who is the one person that will run this every Tuesday?
5. What is the human checkpoint (send / file / price / diagnose)?
6. Which door does it hit — money, time, customer experience — and how will we notice in 30 days?
If (1) or (2) is yes, do not build. If (4), (5), or (6) is blank, do not build.
Bucket it: plain software / one-person helper / later / never.
If you cannot answer those, do not build. Come get a map.
FAQ
When should a small business build custom AI software? After a cheap loop already works, the bottleneck is clicking not thinking, the shop's way of working is the advantage, and a human checkpoint is designed in. That's rare. Most first cards are "no new tool" or "buy X, write this SOP."
Is a $20 tool plus an SOP actually enough? Often yes. First-draft estimates, recare texts, consult recaps, review replies. If people ignore the SOP, they will ignore the custom app too.
Will you quote a build on a free call? No. The assessment sorts your wishlist into plain software, one-person helper, later, and never. You leave with a ranked page, not a custom quote you didn't ask for.
That's the paid assessment: $1,500 for a 45–60 minute owner interview and seven-business-day analysis. I sort your wishlist into those four buckets. You leave with a ranked page, not a custom quote. No free discovery. If you only want the filter and one SOP in a room, the free 15-minute Mini Margin Check is the cheaper door. The retainer exists after we know we are not about to fund a product nobody needed. Most "AI transformation" in a shop is a cleaned-up job and a model in the dull step. That's not a product company. That's operations. Treat it like operations and you will skip a very expensive year.