Satya Nadella recently named a problem every serious AI buyer will recognize: the company may have to reveal its most valuable operating knowledge just to make the AI useful.
In his post on the Reverse Information Paradox, Nadella argues that AI changes the old market for information. The buyer is not only paying for intelligence with money. The buyer is also supplying the prompts, corrections, traces, evals, decisions, and business context that make the system smarter.
That is a sharp way to frame the enterprise AI architecture question. Vendor promises about "no training on your data" matter, but they do not answer the deeper question: who controls the learning loop around the work? That loop includes the memory, the tools, the private scorecards, the approvals, the model routing, and the audit trail.
The model is not the boundary
A single hosted model endpoint is a capability. It is not a strategy. Model rankings move, pricing moves, policies move, and workloads change once the agent touches real systems.
Some tasks need the best frontier model available that week. Some need a cheap, fast model that can classify 10,000 routine records. Some need a model that never sends a token outside the client's environment. A production agent has to treat those as routing decisions, not rewrites.
The durable asset is not one model choice. It is the client-owned orchestration layer that can move work across models without losing the workflow.
Choice has to be designed in
Foundation agents are built around that separation. The agent owns the workflow: what tools it can call, which approvals it needs, what context it can retrieve, how it records its decisions, and how the client evaluates whether it did the job correctly.
The model is one part of that system. We can route work to leading cloud models when the job benefits from their reasoning or multimodal strength. We can use smaller models when cost and latency matter more than raw capability. We can host local models on hardware the client controls when privacy, data residency, uptime, or internal policy requires it.
That portability matters. If one provider changes terms, raises prices, retires a model, or stops fitting the workload, the agent should not lose its memory or tools. The company should be able to move the intelligence layer while keeping the operating layer.
Local models are part of the same system
Local inference is useful well beyond regulated edge cases. Law firms, manufacturers, logistics teams, warehouses, and marketing firms all have moments where the context is too sensitive or too operationally important to hand to a third-party endpoint by default.
We have written before about local models on hardware built for agents. The practical point is simple: the same agent architecture should be able to run against a hosted model today and a local model tomorrow. The client's workflow, evals, approvals, and memory should not have to be rebuilt because the model moved from an API to a machine in their own controlled environment.
What clients should own
A client should own the artifacts that make its agents better at its work: private evals, replayable traces, corrections from experienced staff, tool permissions, memory, and the decision rules that define what "good" means inside that company.
Those artifacts are the company's operational advantage. They should compound for the client, not leak away as exhaust. They should also be usable across model providers, deployment modes, and hardware choices.
The Foundation approach
We build agents as managed software systems, not prompt wrappers. That means scoped credentials, audit logs, human review surfaces, private evals, retry behavior, and model routing are part of the design from the start.
For some clients, the right answer is a hosted agent using frontier models for high-value reasoning. For others, it is a hybrid deployment. For sensitive workloads, it may be a local model running completely on hardware the client controls. The important part is that the company can make that choice without throwing away the agent.
If your AI plan depends on keeping your learning loop, model options, and sensitive context under your control, talk to Foundation. We build agents that can run where the work and the trust boundary require.