AI features (dunning drafts, visit summaries, inbox reply suggestions, message translation) call Claude under the hood. Each call costs us real money; we pass through a monthly cap per tenant to keep the economics reasonable.
AI-drafted dunning is gated to the Estate plan. Visit summaries, reply suggestions, and translations run on Growth and Estate. None of these are available on Seedling.
The monthly cap
Default caps by plan:
- Seedling: AI features not included
- Growth: 500 AI calls / month (visit summaries, reply suggestions, translations)
- Estate: 2,500 AI calls / month (everything Growth has, plus AI-drafted dunning)
A "call" is one AI-generated output: one dunning draft, one visit summary, one set of reply suggestions, one message translation.
What happens when you hit the cap
Every AI entry point checks the cap first. When the cap is reached:
- AI-driven features short-circuit to their deterministic fallback (a simple template-based version).
- An audit row is still written, marked
source=fallback. - You see a banner in Settings: "AI credit cap reached for [month]. Resets [date]."
Fallbacks are not placeholders. They are functional — a fallback dunning email is still a sensible dunning email, just not AI-tuned to your specific situation. Nothing breaks when you hit the cap.
Using AI economically
The features with the highest per-call value are: visit summaries (auto-attached to every completion, so "free" in owner attention terms), and inbox reply suggestions (saves 3-4 min of writing per inbound email). The features with the lowest per-call value are: one-off dunning drafts (the automated cadence is usually enough) and translations you can do in 10 seconds mentally.
Monitoring your usage
Settings → AI shows a progress bar of your month-to-date calls vs cap. Green under 60%; amber 60-90%; rose 90%+. The Audit log has every call with tokens used and latency.
Each AI call counts once, even if it fails. A timed-out Claude call still burns one credit (from your perspective). Our platform absorbs the actual cost for failures, but the cap is simpler to reason about when every attempt counts. If you have repeated failures, contact support — it usually indicates an API key or quota issue on our side.