Build vs. buy a decision tree for the 200–5,000 person company sitting on the agentic AI fence.

The question nobody answers honestly

Every vendor has the same answer to “build vs. buy”: it depends. That answer is technically true and practically useless. Here is a more honest framing.

For mid-market companies — between 200 and 5,000 employees, between $20M and $1B in revenue — the build vs. buy question for agentic AI is almost never about cost. It is about control, capability, and how fast your competitive advantage changes.

The three buy scenarios

Buy when the workflow is horizontal and non-differentiating. Invoice processing, meeting notes, scheduling, basic customer support triage — if dozens of companies in your industry do the same workflow the same way, a vendor has probably already built and hardened the agent. Buying makes sense.

Buy when you cannot absorb the learning curve. Building production-grade agentic systems needs skills your engineering team may not have: eval frameworks, agentic evals, observability for non-deterministic systems. If you are a 300-person manufacturing company with three engineers, the build path may add eighteen months to your timeline.

Buy when speed matters more than fit. Some windows close. If you need an agent in production in eight weeks, you may not have time to build.

The three build scenarios

Build when the workflow is your competitive moat. If the way you process claims, qualify leads, or manage operations is why customers choose you, you cannot outsource that agent. The vendor’s agent will look like your competitors’ agents.

In my experience, the workflows that are hardest to describe to a vendor are usually the ones most worth building yourself.

Build when your data is the asset. Agentic systems trained or fine-tuned on your proprietary data — customer interactions, domain knowledge, historical decisions — have a compounding advantage bought systems cannot replicate.

Build when you need to own the reliability. If the agent sits in the critical path of your operations — if it fails and your business stops — you need to own the codebase, the infrastructure, and the on-call rotation. Dependency on a vendor for a mission-critical system is a risk most mid-market companies underestimate.

The decision tree

Before answering build vs. buy, answer these four questions:

1. Is this workflow differentiating? If yes, lean build.
2. Does the agent use proprietary data as its core asset? If yes, lean build.
3. Is the agent in the critical path? If yes, lean build.
4. Do you have the capability to build and maintain it? If no, reconsider the build path or plan for capability acquisition.

If you answered no to all four, buying is probably fine. If you answered yes to any one of them, the build path deserves serious consideration — and the cost of building is probably lower than you think, especially with a senior-led boutique partner rather than a large consultancy.

The hybrid path most companies miss

The choice is not always binary. A pattern that works well for mid-market companies: buy the horizontal infrastructure (orchestration layer, memory system, eval tooling) and build the domain-specific agents on top.

That gives you vendor leverage on the commodity layer and proprietary advantage on the layer that matters. It also gives you a migration path if the infrastructure vendor changes direction or gets acquired.

Companies that move fastest on agentic AI usually are not building everything from scratch or buying everything off the shelf. They make surgical build vs. buy decisions at the layer level, not the system level.