Build vs Buy AI: Buy the Tools, Build the Seams
Every founder I talk to is stuck on the same decision, and it's the wrong one. Build vs buy AI gets argued like a fork in the road — do we license the platform or do we build our own? The answer is both, and neither, because the valuable part of your system isn't the tools. It's the joints between them.
Nobody is going to out-build Klaviyo's send infrastructure or a frontier lab's model. That war is over and you lost it before you started.
But nobody is going to build the specific way your business decides what to do next, either. Not a vendor, not a model, not a consultant.
Why Build vs Buy AI Is the Wrong Question
Because it assumes the unit of decision is a tool. It isn't. Tools are commodities now — the same three vendors show up in every stack in your category, priced within twenty dollars of each other, shipping the same features six weeks apart.
The buy camp ends up with eleven subscriptions that each do their job perfectly and refuse to talk to each other. The build camp spends nine months rewriting a worse version of something that costs $99 a month.
Both camps lose in the same place: the handoffs. The moment where the support tool learns something the product page should have said. The moment where the ad platform's data should have changed what your buyer sees. Nothing owns that moment, so nobody builds it.
Buy the verbs. Build the sentences.
The Reframe: The Seam Stack
A seam is where two systems meet and a decision gets made. Vendors sell you the systems. Nobody sells you the seam, because the seam is shaped like your business.
I build every stack around four of them — the Seam Stack: Ingest, Judgment, Action, Swap. Buy everything else. Build these four, always, in-house, on purpose.
- —Seam 1 — Ingest. Every vendor holds a piece of your business hostage inside its own dashboard. Pull it out on a schedule into storage you control — orders, tickets, ad spend, events — in a shape you defined, not the shape their export gave you. This is the cheapest seam to build and the one that makes the other three possible.
- —Seam 2 — Judgment. The thresholds, the rules, the taste. When a return gets approved without a human. What margin makes a SKU worth restocking. Which customer gets the apology and the credit. Vendors ship generic defaults here because they have to serve everyone. Your defaults are the business.
- —Seam 3 — Action. A decision that doesn't write anywhere is a dashboard. Build the layer that pushes back into the tools — update the listing, pause the ad set, fire the flow, email the supplier — with every write scoped by what a wrong call costs you. Read widely, write narrowly.
- —Seam 4 — Swap. Assume every vendor in your stack gets acquired, repriced, or outclassed within eighteen months. Put a thin layer of your own code between your logic and their API so replacing one is a week, not a quarter. You are not avoiding lock-in for ideology. You are keeping the option to move fast when the market moves.
Ingest gives you the inputs. Judgment gives you the call. Action gives you the effect. Swap keeps all three yours.
The Stack That Holds the Seams
What you buy, without guilt
Shopify for the store. Stripe for money. Klaviyo or Resend for sends. Gorgias for the queue. The model from a frontier lab. Anything where the vendor's roadmap is already ahead of yours and always will be. Paying for these is not a compromise — it's how you afford to build the parts that matter.
What you build, always
A Postgres you own — Supabase or D1 — as the place every vendor's data lands. Cloudflare Workers on a cron for ingest and for the writes back out. Your rules as actual code in a repo, versioned, reviewable, with the model calls behind a strict JSON schema so the output is data your system can act on instead of prose someone has to read.
The test before you build anything
Ask one question: if a competitor bought the identical tool tomorrow, would they get the identical result? If yes, buy it and stop thinking about it. If no, that difference is the thing only you can build — and it's almost always sitting in a seam, not in a feature.
How This Runs Across My Ventures
At Bayani Brands I buy the entire commerce stack and build almost none of it. What I do build is the layer that reads across it — what the ads promised, what support heard, what the returns say — because no vendor can see all three at once. That view is the asset. The subscriptions are rent.
Marky AI sits on someone else's models by design. The defensible part was never the inference. It's the judgment seam: what good output looks like for a specific brand, encoded well enough that the system holds the line without me in the room.
Across 200+ websites, the ones I could still change two years later were the ones where the logic lived in my repo and the tools were interchangeable underneath it. And the most common thing I untangle for members inside AI Systems Club isn't a bad tool choice. It's a founder who bought nine excellent tools and built zero seams, then wondered why the stack feels like a filing cabinet instead of a machine.
The Takeaway
Stop auditing your tools. Audit your joints. Open the map of your stack and find every place a human currently carries information from one box to another — that human is standing in an unbuilt seam.
Build one this month. Ingest first, because everything else needs it.
Anyone can buy the same stack you have. Nobody can buy the way it's wired together.
We build these seams — the ingest jobs, the rules, the scoped writes — with 500+ founders and operators inside AI Systems Club. Come build with us.
Want the full playbook? Join 500+ founders building real AI systems.
Join AI Systems Club