Field Notes · Sep 25, 2026 · 6 min read
Design around the workload, not the model
Why we don't pick an AI provider first — and how voice, reasoning, privacy and volume each pull a system toward a different kind of model.
The first question most companies ask about AI is which provider to use. We think it’s the wrong first question. In real estate, the work itself pulls a system toward very different kinds of models — and sometimes toward different buildings entirely.
Four forces that shape the choice
Speed. A live phone call needs an answer in under a second or the conversation falls apart. That pulls toward realtime voice models built for turn-taking. A monthly owner report can take a minute to write and nobody will notice.
Sensitivity. Rent rolls, owner agreements and tenant records carry obligations. When data shouldn’t leave your environment, the model has to come to the data — an open-weight model on infrastructure you control, not an API on someone else’s.
Volume. Sorting ten thousand inbound messages is a different problem from reading one lease. At volume, a small, specialized model does the routine work quickly and cheaply, and hands only the uncertain cases to something larger.
Depth. Some work genuinely needs the strongest reasoning available — a long document with subtle clauses, or writing that a client will read. That’s where frontier models earn their cost.
What this looks like in practice
One organization can reasonably use three or four model classes at once: realtime voice for calls, a retrieval layer over its own documents, a private model for sensitive questions, and a frontier model for drafting. The architecture — which request goes where, under which data policy, with which human checkpoint — matters more than any single model choice.
It also ages better. Models improve and prices fall every few months. A system designed around the workload can swap the model underneath without being rebuilt; a system designed around one provider has to start over.
How this shows up in our own work
The MaintenanceTech platform was designed to be model-agnostic and capable of using open-source models, so the model underneath can change as better or cheaper options appear. We bring the same principle to every system we build for clients: the work and the data rules come first, and the models are chosen to fit them.
Published by K3YHOLD, Sep 25, 2026.