Buy-vs-build in the age of foundation models

The old buy-vs-build calculation assumed software was expensive to build and cheap to buy. Foundation models flipped both sides. Building agents on a platform is cheaper than it has ever been; buying vertical AI products is riskier than it looks because the incumbents haven't caught up and the challengers are mostly one release cycle away from being commoditized. So the calculus that worked for the CRM era doesn't work here — and enterprises that apply it literally end up with the wrong contract, the wrong vendor and the wrong internal capability.

The three questions that decide every case

Before we open a vendor short-list, we ask three questions. Is the workflow differentiating to your business? Is the data proprietary or commodity? Do you already own — or are you willing to own — a platform team? The answers point to buy, build-on-platform or (rarely) ground-up. If a leader can't answer all three, we stop and go do that work first.

Buy

Buy when the workflow is horizontal, the data is commodity, and the incumbent product is a real business — not a Series-A demo. HR service desks, generic marketing tooling, standard sales development, expense automation, meeting transcription. In these categories the incumbents have distribution, integration surface and a support model your CIO can defend, and the underlying LLM contribution is a small part of the total value.

Buy — where it gets dangerous

Buying gets dangerous when the vendor is 'AI-native' but the category isn't. If the workflow is differentiating to you, the data is proprietary, and the vendor's moat is a wrapper on the same foundation model you can access directly, you are paying rent on a capability you should own. Twelve months later the model provider ships the same thing natively and the vendor has a repositioning deck.

Build on a platform

Build on Bedrock, Azure AI Foundry, Google Gemini or Kore.ai Agent Platform when the workflow is differentiating, the data is proprietary, and you already own a platform team. This is where most enterprise AI value gets created in 2026. You get the leverage of the platform (models, safety, evaluation, observability) without giving up control of the workflow, the data or the roadmap.

Build on a platform — the org prerequisite

The unit of success for build-on-platform is not the model, it's the team. You need a small, senior AI platform team — 6 to 12 people — that owns the reference architecture, the eval harness, the guardrails and the FinOps rhythm. Without it, every agentic workflow becomes its own snowflake and the cost of ownership crushes the ROI. If you don't have that team and won't fund it, don't start here.

Ground-up

Almost never. If you are training your own foundation model, you are either a hyperscaler, a research lab, or you have made a mistake. Fine-tuning a small open-weights model on your proprietary data for a bounded task is not ground-up and is often the right move; training a general model to compete with GPT-5 or Claude 4 is a five-year, nine-figure decision with a very small addressable outcome.

The recontract question

The buy-vs-build decision now has a lifespan question attached: what happens at renewal? For AI-heavy vendors, 24 months is a long time. Structure contracts so you can walk — export your prompts, evaluation sets, historical traces and workflow logic. Don't sign a three-year deal that leaves you with nothing you can move.

What we recommend to CIOs

Portfolio-manage this. Bucket your AI investments into buy, build-on-platform and (very rarely) ground-up. Set explicit budget ratios — most Fortune 500 programs we advise land somewhere near 60/35/5. Review the mix every two quarters, because the market moves faster than annual planning cycles can absorb.