A 20-person plumbing company called me last month with a problem I hear constantly. One vendor had quoted them $40,000 for a custom AI system. Another was selling a $149-a-month tool that promised to do "basically the same thing." Both salespeople sounded confident. Both had slide decks. The owner had no way to tell which one was right for him, and he was smart enough to know that guessing wrong in either direction would cost him real money.
So let us settle it. Not with a slogan, but with a way to think that works every time.
You are asking the wrong question
"Build vs. buy" sounds like buying a truck versus building one in your garage. That picture is wrong, and it leads smart owners to wrong answers.
The real question is simpler: are you buying software, or are you buying configuration?
Real software keeps working when the vendor looks away. You pay the subscription, it does the job, and nobody on your team thinks about it. Configuration looks like software on demo day, but the moment something changes, a phone number format, a new service you offer, a price update, somebody has to go in and fix it by hand. If that somebody is you, on a Sunday night, you did not buy software. You bought a part-time job.
Here is the rule of thumb I give every owner: if it breaks and nobody notices for a week, you bought configuration. If it breaks and the vendor's monitoring catches it before you do, you bought software. Ask that question about anything anyone tries to sell you, and half the confusion clears up immediately.
This matters because the AI market right now is full of configuration dressed up as software. A chatbot bolted onto your website with no connection to your schedule, your CRM, or your team is configuration. It will impress you on Tuesday and embarrass you by Friday. Knowing the difference is the whole decision.
The 5-question scorecard
Forget the slide decks. Run any workflow through these five questions. Your answers will tell you whether to build, buy, or walk away.
1. How unique is the process? If every competitor does it the same way, buy. Appointment reminders, review requests, basic phone answering: these are commodities. There is no prize for having a unique appointment reminder. But if the process is yours, the way your shop quotes custom work, the way your dispatchers route techs, the way your team qualifies a buyer, that uniqueness is exactly what off-the-shelf software will flatten. Buy the commodity. Build the fingerprint.
2. How sensitive is the data? Customer lists, health information, financial details, anything you would not want sitting on a stranger's server: the more sensitive the data, the stronger the case for building, or at least for a setup where your data stays inside systems you control. This is not paranoia. It is the same reason you lock the office. Off-the-shelf tools are fine for marketing copy. They are a different conversation when they touch your customer database.
3. How many systems does it have to touch? One or two systems: buy, all day. A tool that lives inside your CRM and sends texts is a solved problem. Five or more systems that all need to agree with each other: buying gets expensive and fragile fast, because now you are paying for five subscriptions and praying their integrations do not break on the same Tuesday. This is the zone where a custom build starts winning on total cost, usually around the second year.
4. Is it a competitive edge or table stakes? Table stakes, the things customers expect every business to have, should be bought as cheaply as possible. Nobody picks a plumber because of their reminder texts. Competitive edge, the thing that makes you faster or better than the shop across town: that is worth building, because your competitor cannot buy it off the same shelf. Be honest about which is which. Most owners overestimate how many of their processes are truly an edge. It is usually one or two.
5. Who fixes it at 9 PM? This is the question nobody asks and everybody should. If something breaks on a Friday night, who gets the call? If the answer is "the vendor's support team," buy with confidence. If the answer is "nobody, it waits until Monday," that is fine too, as long as you said it out loud. If the answer is "me, I guess," do not build unless you are also buying someone to watch it. Unmonitored custom software is a liability with a login page.
Score it simply. If three or more answers point toward unique, sensitive, multi-system, edge, or needs-a-watcher, you are in build territory. If three or more point the other way, buy. If it is 2-2 with one shrug, keep reading, because the hybrid answer is probably yours.
What "build" actually means (for the technically curious)
You are not training AI models. Let us be clear about that, because vendors love to make "custom AI" sound like a research lab. Nobody at a 20-person company is training a model, and nobody needs to.
When we talk about building, we mean buying architecture. The AI models themselves are rented, interchangeable, and getting cheaper. What you own is everything around them: the API connections between your systems, the logic that decides what happens when, the approval gates where a human says yes, the logging that tells you what the system did and why. That plumbing is yours. It lives in your accounts. If a model provider raises prices or shuts down, you point the plumbing at a different provider and keep going. That is what model-agnostic AI means in practice, and it is the single most important property a custom build can have.
What "buy" means, by contrast: per-seat pricing, their roadmap, their price increases, their deprecation schedule. That is a fine deal for commodities. It is a bad deal for the thing that makes you different, because the day the vendor changes the terms, your edge changes with it.
The technical depth ends here. The decision does not require it, but the people evaluating you on it deserve to know you have thought about it.
The hybrid answer most 20-person firms land on
Almost nobody should build everything. Almost nobody should buy everything either. The pattern that works, over and over, looks like this:
Buy the commodity. Scheduling, reminders, review requests, basic phone answering, standard CRM: these are solved problems with mature tools. Pay the subscription, connect them properly, move on. Do not pay a developer to rebuild a calendar.
Build the edge. The one or two workflows that are genuinely yours. For the HVAC company, we bought the answering service and built the dispatch triage, because how they route techs across three counties is the thing their competitors cannot copy. For a real estate team, buy the CRM, build the speed-to-lead layer on top of it, because response time is their edge and the CRM's built-in version was not fast enough. For the distributor, buy the accounting software, build the quoting logic, because their pricing rules live in one estimator's head and that is exactly what should be captured in a system they own.
Connect them like you mean it. The hybrid only works if the bought pieces and the built pieces actually talk to each other. A bought answering service that does not write into your CRM is not a system, it is a gadget. The connection layer, webhooks, shared data, one source of truth for the customer record, is where these projects succeed or fail. It is the same foundation-first thinking behind Why We Don't Sell Websites: the website was never the product, the connected systems underneath it are.
Notice what is missing from this picture: the $40,000 everything-custom quote and the $149-a-month does-everything tool. Both were wrong for that plumbing company. They bought the scheduling and review tools, and we built the after-hours triage that routes emergency calls to the on-call tech with the customer's history attached. Total first-year cost came in well under either extreme, and the thing they built is the thing their competitors do not have.
Your Monday-morning move
Pick the single most painful workflow in your business right now. Not the top five. One. Run it through the five questions above and write down the answers. If three or more say build, get a build quote AND a buy quote, and compare total cost over two years, not month one. Include the maintenance, the per-message fees, and the Sunday nights. If three or more say buy, buy the best-reviewed option, connect it to your CRM properly, and move on to the next problem.
And if anyone tells you the answer is obvious before they have seen your workflow, get a second opinion. The red flags to watch for, in both directions, are the ones we catalogued in Your AI Vendor Is Lying to You. The answer is never obvious. That is why there is a scorecard.
Questions owners actually ask
Should a small business build custom AI or buy software? It depends on the workflow, not the company. Buy the commodity stuff every business does the same way. Build the one or two things that make you different. Most 20-person firms end up doing both, and that is the right answer more often than either extreme.
When does custom AI automation make sense? When the process is unique to your business, touches sensitive data, connects a lot of systems, or gives you an edge competitors cannot buy off a shelf. If none of those are true, buy. Run the five questions above and trust the score.
What does custom AI cost vs. off-the-shelf? Off-the-shelf usually runs tens to low hundreds per month per seat. Custom builds start in the five figures and climb with complexity. But the honest comparison is total cost over two years, including maintenance, usage fees, and who fixes it at 9 PM, not the month-one number on the quote.
Can you start with tools and move to custom later? Yes, and it is often the smartest path. Start with off-the-shelf to prove the workflow matters and learn exactly what you need. Then build custom once the requirements are clear instead of guessing. Just make sure your data can come with you when you move, because data trapped in a vendor's system is the most expensive part of switching.
---
Bring your toughest workflow to the free 30-minute diagnostic at krastor.com/diagnostic. We will tell you straight whether you should build it, buy it, or leave it alone, and we will show our work.