Founders ask me most weeks whether they should build a custom AI tool or buy something off the shelf. I won't run that decision for them, because that's strategy work I don't sell. But I've watched enough of these go both ways that I've got a view, so I'll write it down here instead.
Fit is the thing that matters most. Most businesses have at least one workflow specific enough that nothing off the shelf quite handles it. They've also got a dozen that are generic enough for a fifty dollar a month SaaS product to do a better job than anything custom. It's not a yes or no thing, it's a percentage. If an existing tool covers more than about 80% of what you need, buying almost always wins. Under about half, custom almost always wins. In between it genuinely depends.
Cost is where I see people miscalculate most often before a build starts. A four hundred dollar a month tool over three years comes to around fourteen thousand dollars. A custom build starts around twenty-five thousand and grows once you account for maintenance. That maths only tilts toward custom when you're looking at multi-year use and you think the tool is going to be a real advantage over competitors.
Time matters more than people account for. Something off the shelf can be running in days. Custom takes weeks. The builds I take are four to six weeks for a reason, which is roughly the point where custom starts beating configuration for the right sort of problem. If something is on fire next week, buy now and think about building later.
The trap I see most often is assuming you'll build it cheaper than you can buy it. That's almost never true, because the cost isn't really the build. It's the maintenance, the integrations, the edge cases and the hours your team spends thinking about it afterwards. A founder once walked me through a spreadsheet justifying eighty thousand dollars of custom work to replace a four hundred dollar a month tool. We bought the tool and spent the time on something that mattered.
The trap going the other way is buying something that doesn't really fit and then bending your workflow around it for the next three years. I've seen teams paying ten thousand a month for tools that solve most of one problem and create a decent chunk of another. At that point building is cheaper and faster and lasts longer.
In the builds I take on, some mix of the two nearly always wins. Buy the foundations, so API access to OpenAI or Claude, a workflow tool like Zapier or n8n, a vector database you don't have to run yourself. Then build the thin custom layer that makes those generic pieces specific to how your business actually works. That's most of what an AI feature build turns out to be.
None of it is permanent either. Most of the clients I've worked with started one way, proved the concept, and then switched. Buying and then building later is completely fine. Building and then realising you should have bought is also fine. The expensive option is sitting in the decision for six months while the problem keeps costing you money.
If you've already worked through this and landed on building something with a defined scope, that's what the scoping call is for. If you haven't, the cheapest thing you can do this week is try the off-the-shelf option and see how far it gets you.