Build vs. buy in the agent era: a CFO-friendly framework
The question has shifted
Three years ago, "build vs. buy" in enterprise software was mostly about integration effort and maintenance overhead. You bought your ERP, your WMS, your CRM. You built only where you had genuine proprietary advantage. The calculus was stable and well-understood.
Agentic AI breaks that model. The substrate is changing fast enough that buying a vertical AI platform today means you may pay a full SaaS tax on last year's architecture. But building from scratch means carrying the full cost of model research, eval infrastructure, and orchestration — work that is not your core business.
The three-zone model
- Buy the commodity. Foundation models, hosting, observability tooling, vector databases — this is infrastructure. Buy it. The switching cost is low and the innovation rate is high enough that owning this layer would be actively harmful to your roadmap.
- Configure the middle layer. Orchestration frameworks, retrieval pipelines, workflow automation — configure, do not code. You want control over the logic without owning the runtime.
- Build the proprietary layer. Your evals, your domain-specific prompts, your process logic, your human-in-the-loop contracts. This is where your defensible advantage lives. No vendor can build this for you because it requires deep knowledge of your specific operations.
The crossover point
The crossover between build and buy almost always arrives at the eval layer. A vendor can give you a general-purpose agent. They cannot give you an agent that understands your exception taxonomy, your confidence thresholds, your escalation rules, your team's override preferences. That has to be built — and it is more valuable than any feature on a vendor roadmap.
The CFO framing
Buy where the commodity is maturing faster than your use of it. Build where your ops knowledge is the moat.
In practice: your AI spend should be heavily weighted toward configuration and eval infrastructure, with a smaller allocation to foundation model costs, and a minimal amount to proprietary runtime ownership.
The organizations that will regret their 2024–2026 AI spend are the ones that paid full SaaS prices for platforms they could not customize, and the ones that built full-stack AI when they should have been configuring. The winners are building narrow, deep proprietary layers on top of fast-moving commodity infrastructure — and holding that line year after year.