The buy-vs-build conversation around agentic AI gets stuck within ten minutes, every time. Someone says "we should just buy something off the shelf." Someone else says "nothing on the market does what we need." Both are right, because they're answering different questions.
There are actually three.
1. Buy or build the model?
This one's settled. You're not building the model. Frontier models are a commodity input — you'll consume Claude, GPT, Gemini, and whatever comes next via API, and you'll switch between them based on price and capability. Anyone proposing to fine-tune a foundation model for your close process is solving last year's problem.
Verdict: buy, and assume you'll be swapping providers every 12-18 months.
2. Buy or build the platform?
Platform means the orchestration layer — the harness that runs prompts, manages tools, stores eval results, handles human-in-the-loop, logs decisions, integrates with your stack. This is where most teams burn six months reinventing infrastructure that already exists.
If you don't have an opinionated platform engineering team with capacity to spare, buy the platform. The differentiation you're chasing isn't in the orchestration plumbing — it's in the workflows you build on top of it.
3. Buy or build the workflow?
This is the only one that's genuinely a decision. The workflow is your rev rec policy, your reconciliation logic, your close runbook, your controls. The off-the-shelf products that promise to do these "out of the box" are selling you someone else's interpretation of your process — which is exactly the thing you spent ten years differentiating.
The honest answer is build the workflow on a bought platform, with the people who actually run the process today. Not the platform vendor. Not a generalist consultancy. The people who own the close.
Why this matters
Most failed agentic AI programmes failed because they answered question 3 with "buy" and ended up with a tool that handles the easy 40% of their process and creates more work for the rest. Most successful ones bought aggressively at layers 1 and 2 and invested their scarce internal time at layer 3.
If you're being sold a single product that claims to handle all three layers, ask which layer they actually built. The answer is almost always "the platform" — which is the one layer you probably shouldn't have built yourselves anyway.
The decision isn't buy or build. It's what to buy and what to build, and getting that boundary right is the difference between a programme that compounds and one that gets quietly shelved after the pilot.