What causes unexpected cost spikes in AI assistant usage?

Yes. If conversations get longer or retrieval pipelines expand the context window, token usage per message increases. That multiplies quickly across thousands of sessions.

It can. Larger knowledge bases often mean larger context payloads unless retrieval is carefully optimized. Some systems inject more content than necessary.

Another common factor is retry behavior. If downstream APIs fail and the assistant retries automatically, each attempt still consumes tokens.

Exactly. Model cost is predictable. Poor orchestration isn’t. Systems that compress prompts, cache retrieval results, and validate API calls before executing them tend to produce much more stable cost curves.