Every build-vs-buy meeting I walk into starts wrong. Someone wants a verdict. Build or buy. Pick a lane. Ship the slide.
That is the wrong question. And a recent Lyzr CTO framework actually names it. Ask a room of engineering leaders build or buy, and you get answers rooted in ego, prior experience, or whichever vendor pitched them last quarter. Ask instead which layer of the stack should we own, and the conversation changes.
Good. Now here is the answer that holds up when the demo becomes a dependency: build the IP, buy the substrate.
Substrate is rented. IP is owned.
Substrate is everything that makes your agent run but nothing that makes it yours. The layers break down roughly into foundation models, orchestration and agent frameworks, retrieval and context infrastructure, evaluation and testing, observability and guardrails, and governance and compliance. Models are substrate. Commodity MCP servers are substrate. Orchestration runtimes, tracing, the eval harness plumbing: substrate. You rent it, you swap it, you refuse to fall in love with it.
IP is the opposite. It is your agent’s decision logic, your domain-specific eval suite that encodes what good actually means in your business, your proprietary context graph, and the MCP servers that wrap your systems and your data. Nobody can buy that off a shelf because it does not exist on one. That is the whole point.
The confusion persists because the industry keeps selling you a volume calculation dressed up as a strategy. The answer changes at approximately one million conversations per year. Below that threshold, buy wins on speed and integration simplicity. Above it, build wins on per-token economics. Useful math. But that is a substrate question, per-token versus per-conversation pricing. It tells you how to pay for the engine. It tells you nothing about what to own.
Where teams get it backwards
The failure mode in production is not building too much. It is building the wrong layer and buying the wrong layer.
Teams buy the IP: they license a packaged platform to run their differentiated workflow, then spend six months bending the platform to behavior it was never designed to express. The result is predictable. Roughly 50 percent of enterprise agentic AI projects are still in POC or pilot stage, and a meaningful share are stuck there because the vendor platform they bought could not flex to the workflow they actually needed.
Then those same teams turn around and build the substrate. They stand up a bespoke model gateway. They hand-roll a protocol layer. This is pure waste. The December 2025 donation of MCP to the Linux Foundation AAIF stabilized the spec and removed the vendor-coupling risk that held back enterprise adoption. The connective tissue is now a standard. Do not rebuild the protocol. Buy the substrate and spend your engineering budget on the twenty percent that is actually yours.
There is one honest caveat, and it is an optimization, not a mandate. Remote MCP servers add network round-trips to every tool call, and for agents that make ten to twenty tool calls per task this compounds, so several teams reported rewriting their most-called MCP integrations as native tools for latency reasons. Fine. That is a performance decision on a hot path, not a reason to own the whole substrate layer.
The CODN math nobody runs
The Cost of Doing Nothing is the number that should drive this decision, and almost nobody computes it. Doing nothing here means stalling in the build-or-buy debate while the window closes. Gartner projects that more than 40 percent of agentic AI projects will be canceled by the end of 2027, mainly from underestimating the real costs. Most of that is teams who bought the wrong layer, or built a layer they could have rented, and ran out of patience before they ever owned anything differentiated.
The operators getting it right already made the call. According to the McKinsey State of AI 2026 report, 32 percent of organizations have decided against buying off-the-shelf software, opting instead to build their own solutions using agentic coding tools. And the appetite skews toward the winners. The 6 percent of respondents who attribute at least 5 percent of their EBIT to AI are the high performers, and nearly half of them are skipping software purchases, compared to 31 percent of their peers.
They are not building because building is noble. They are building the IP and renting the rest, because the rest is a liability you do not want to carry. The dependency risk is already baked in regardless. A Zapier survey found 81 percent of enterprise leaders are concerned about AI vendor dependency, yet only 6 percent believe they could switch providers without material disruption.
So own the layer that makes you different. Rent everything that does not. The substrate will keep changing underneath you, and that is exactly why you should never pour your moat into it.
Buy the substrate. Build the IP. Ship the agent before the window closes on the people still debating the lane.