Vendor, build, or wait: AI tooling decisions in 2026

The single most expensive AI mistake mid-market companies make is buying tooling they do not need yet. The second is building tooling that should have been bought. The third is waiting on something they should have shipped six months ago. All three look identical from the inside.

Here is the framework I use when a client asks me to help them decide.

Wait if the capability is moving fast. If the underlying model improvement is on a quarterly release cadence, anything you build today will be a worse version of what is shipping at GPT-X-plus-one or Claude-X-plus-one. Wait one or two cycles. Use the gap to clean up your data and write the evaluation suite. The system you build on the next-gen model will be better and cheaper than the one you would have shipped on this generation.

Buy if a vendor does the boring 90% well and the rest is configuration. Resend, Clerk, Stripe, Supabase, OpenAI, Anthropic, Cloudflare, Railway. None of these are AI tools per se but the principle is identical. If the vendor solves the unglamorous middle of the system and you can configure the last mile, buy. Building your own auth system in 2026 is almost always a mistake. Building your own LLM API gateway is usually the same mistake one level up.

Build if the differentiator is in the configuration the vendor will not let you control. Custom retrieval, custom evaluations, the prompt logic that encodes your operational know-how, the schema that reflects how your business actually works. These are the parts where building is justified, because the vendor's defaults erase the thing that makes you different.

Build the seam, not the system. The pattern I keep coming back to: build the integration layer between vendor capabilities and your business logic, not the underlying capabilities themselves. The seam is where your competitive advantage lives. The vendor commodity is, well, commodity.

Wait if the team cannot maintain it. This is the silent killer. The decision is not just whether the system is right today; it is whether the team will recognize when the model provider ships a regression in six months. If nobody on your side will own the system, do not build it. Do not buy it either if the vendor's support model assumes a technical owner. Wait until you have hired the person who will own it.

The least useful answer in this whole framework is "use AI". The most useful is one of: wait, buy, build, or hire first. I will tell you which on the call.