An AI-native DMS has a real per-question marginal cost — Anthropic and OpenAI bill us per token. Without caps, one runaway agent loop or one power-user team could push spend past your subscription, and we’d either eat it or quietly throttle you. Both are bad outcomes.
Caps make the math legible. The Team tier’s 200 questions/seat/month is sized so a typical-usage seat costs us about $5–6 against the $30 you pay — comfortable margin without gouging. Heavy-use seats (the AEC firm running structural analysis through Opus) move up to Business, where unlimited reasoning is on the menu and the price reflects it.
Caps are load-bearing. We surface usage on /settings/billing and the dashboard; over-cap requests actually refuse rather than silently degrade — except for Opus reasoning, which falls back to Haiku so you keep getting answers on the cheaper model. The dashboard banner warns at 80% so the upgrade signal isn’t a surprise.
Your spend is also transparent end-to-end: every tool call the agent makes is logged with its cost in microcents on the audit trail, and /costs (owner / admin only) rolls up the 30-day total by kind. If a customer disputes a number, the audit-log row pairs 1:1 with the cost-event row that drove it.