Build, buy, or orchestrate: a PM's framework for enterprise AI agents

Every enterprise AI roadmap eventually runs into the same fork: do we build this capability ourselves, or buy it from a vendor? With agents, that framing is already out of date. There's a third option that is usually the best one — orchestrate. Wire together models, tools, and services you don't own into a workflow that you do.

“Software is eating the world.”

The reason the third option matters: in an agentic system, the model is rarely your differentiator. Five competitors can call the same foundation model. What you uniquely own is the orchestration — the retrieval, the tools, the policy logic, the evaluation, the way it all fits your customer's workflow. So the strategic question isn't "build or buy the model." It's "which layer of this stack is actually ours to win?"

A test for what to build

Build only where the capability is both core to your differentiation and not available at the quality you need. That intersection is smaller than engineers want it to be and larger than finance wants it to be — which is exactly why it's a product decision, not a purely technical or budget one.

Be honest about the second half of that test. Teams love to build because building is fun and feels like progress. But every component you build is a component you maintain forever, through every model upgrade and every security review. The 'Not Invented Here' reflex has sunk more AI roadmaps than any vendor ever has.

What to buy

Buy the commodity layers without ego: foundation models, vector databases, observability, the undifferentiated plumbing. The mistake here is buying a monolith that promises to do everything, then discovering it does none of the parts you care about especially well — and that you've handed your roadmap to someone else's release schedule. Prefer components with clean interfaces you can swap out over suites you can't escape.

Why orchestrate is usually the answer

Orchestration is where most enterprise AI value actually lives in 2026. The foundation models are improving on someone else's dime; the moment you hard-wire your product to one, you've signed up to rebuild when the next one leaps ahead. Orchestrate instead, and you can swap the model underneath without touching the workflow your customer relies on. Your differentiation moves to the seams — the routing, the guardrails, the evaluation, the domain logic — which is precisely where a deep understanding of the customer's world is hard to copy.

Putting it to work

1. 1Decompose the agent into layers: model, retrieval, tools, policy/decisioning, evaluation, workflow UX.
2. 2For each layer, ask: is this core to our moat, and is it available at the quality we need?
3. 3Build the rare intersection. Buy the commodity. Orchestrate the rest — and keep the interfaces clean so you can change your mind.
4. 4Revisit yearly. The 'build' list shrinks every time the ecosystem matures.

The teams that win the agent era aren't the ones who build the most. They're the ones who are honest about which layer is theirs to own — and ruthless about composing the rest.