Over the past decade every enterprise has debated build-vs-buy for cloud. I saw several private cloud projects die on the vine because companies grossly underestimated the time and capital investment required to bring them to production and keep them running.

The same discussion is happening now for AI. When compute became commoditised, the real question was which orchestration layer to adopt and who would run it. The same is true here. The model is increasingly commoditised; the orchestration layer (retrieval, memory, permissions, guardrails, data connectors) is where the real decisions live.

  1. Your productivity suite vendor’s solution. Microsoft has Copilot, Google Workspace has Gemini. These options are turnkey; they already have connections to your data and it’s fast to deploy, but make sure you don’t end up with a patchwork of point solutions rather than a coherent AI strategy.

  2. A model provider’s solution. Model providers are racing to build tooling wrappers around their models, increasingly referred to as the ‘harness’. It’s the model plus the harness that creates the real productivity gains. Claude Cowork by Anthropic is a great example of this. The downside is a closed source harness backed by a closed weights model, in a space where the leading model can shift within a year. The upside is speed to value - these are often very polished end-to-end experiences, valuable for teams that want to move fast without building infrastructure.

  3. Build your own. Maximum flexibility on models and data sources. But it’s not a trivial undertaking. This is the private cloud path. For the right companies with the right investment, it works, but if you go this route, ensure you have sufficient capital and human resource allocations, and a multi-year plan. The recently released LangChain Fleet is worth a spike for those considering this route.

  4. Use a hosted model-agnostic harness. For most enterprises today, probably the sweet spot. You keep full optionality on models and data sources without the operational overhead. Glean and Kore are examples of this approach. But enterprises should be aware of new areas of potential lock-in. Who owns the index? Where do embeddings live? What happens to your knowledge graph if you switch providers?