The safety gate is the differentiator between Summit and "AI auto-replies to everything" tools. Every inbound tenant email runs through a deterministic classifier before the AI drafts a reply. If the classifier fires, the agent drafts a reply but holds it for your review — it never ships to the tenant without you clicking Send.
The 9 categories
- Distress. Self-harm signals, suicide language.
- Emergency. Fire, gas leak, carbon monoxide, active flooding, electric shock, smoke alarm.
- Legal. Mentions of attorney, lawsuit, small claims, court, fair housing complaint, HUD, discrimination, rights-based defenses (withholding rent, repair-and-deduct, fit premises, constructive eviction).
- Habitability. No heat, no water, mold, bedbug / cockroach / rat infestation, asbestos, lead paint, raw sewage.
- Immigration. ICE, deportation, USCIS, green card, detained.
- Hostility. Scam accusations, threats, profanity directed at the operator, racism / sexism / bigotry claims, review-damage threats ("one-star on Google").
- Financial. Bankruptcy, garnishment, foreclosure, debt collector mentions.
- Evidence collection. Soft-language patterns like "you said", "as we agreed", "I have your text", "for three weeks the heat hasn't worked", "this is the third time" — signals a paper-trail buildup before legal action.
- Negotiation. "Can we just waive the late fee", "would you consider", "let me out of the lease", "free month" — operator-discretion territory that the AI shouldn't unilaterally agree to.
Soft-language manipulation
The hard categories above catch obvious red flags. The soft-language patterns catch the dangerous case the hard rules miss: a tenant collecting operator quotes before suing. Examples that fire:
- "You said the heat would be fixed last Tuesday." (Evidence)
- "As we agreed, the work should have been done by now." (Evidence)
- "Per our conversation, I expected the rent to be reduced." (Evidence)
- "Since you can't fix the bathroom can we agree I just don't pay April." (Negotiation)
- "I'll withhold rent until you address the mold." (Legal — conditional non-payment)
- "This is your responsibility to fix not mine." (Legal — liability admission probe)
- "For three weeks the heat has not worked." (Evidence — Fit Premises buildup)
- "My girlfriend is moving in next week." (Legal — unauthorized occupancy)
Reviewing held drafts
Held drafts land on the dashboard's Escalations panel and simultaneously in your email inbox. The email contains the tenant's message, the AI's draft reply, and the specific reason the gate fired. From the dashboard you can Send (releases the draft as-is), Edit (modify then send), or Dismiss (no reply will be sent, you handle this tenant directly).
What if the gate fires on something benign?
False positives are cheap — you just click Send. The classifier is intentionally conservative because false negatives (auto-replying to something risky) are the expensive failure mode. If a pattern fires repeatedly on a tenant you trust, email us — we'll look at whether the pattern can be tightened.