The AI stack for a team of forty

Most published AI architecture assumes a platform team. If you have forty people, no ML engineer, and an existing product to keep running, almost none of it applies.

What we reach for

• A hosted model API rather than anything self hosted. The savings from self hosting do not cover one engineer month until you are far larger than this.
• Your existing database. A dedicated vector store is a second system to operate and rarely earns that at this scale.
• Background jobs on the queue you already run, instead of a new orchestration layer.
• Plain logging and a scheduled evaluation run, instead of an observability product.

The pattern underneath

Every one of those is the same trade: fewer moving parts, at the cost of a ceiling you will not reach this year. Small teams die of operational load long before they die of insufficient throughput.

When to change your mind

Move off any of these when you have a measured problem, not an anticipated one. "We might need to scale" is not a measured problem. "Retrieval latency is now the slowest part of the request and we have the trace to prove it" is.