Three agents that run your inbox, cited to your own rules
Boards are volunteers, and most of the inbox is the same six questions. PonyHOA runs three agents against it: one answers residents from your governing documents, one reads every message that arrives and proposes what should happen to it, and a third, in preview, turns a violation report into a tracked case. Each one shows its work, and a board member releases everything that reaches a resident.
There’s been a boat parked in the driveway at 41 Alder for about three weeks. Is that even allowed? And who do I tell?
Dana R. · Lot 27 · +1 (512) 555-0184
Thread context: 2 messages · SMS · Lot 27
- search_governance_docsCC&Rs Art. VII §3 “Vehicles & Parking”, p. 12
- find_similar_past_inbounds4 prior threads · matched the board’s wording
- get_member_contextLot 27 · dues current · no open violations
Thanks for flagging it, Dana. Boats and trailers can’t be kept in driveways for more than 72 hours under Article VII, Section 3 of our CC&Rs. I’ve passed 41 Alder to the board so they can follow up directly.
↳ CC&Rs Art. VII §3, p. 12 · Board minutes, Mar 2025, p. 2
Held for review. Every proposal waits for a person.
The same five questions, forever, and the notice nobody sent
Boards spend hours re-answering “what time does the pool close”, “can I rent my unit”, “when is the next meeting”. The answers all sit in documents residents already have access to. Meanwhile the violation somebody reported in March never became a notice, because the person who meant to raise it at the next meeting forgot. Agents grounded in your own documents fix both ends: residents get answered in seconds, and the things that need to become cases actually do.
Capabilities
Answers agent
A resident texts “what is the pet policy” and gets the rule back with the article, section, and page it came from. If no section covers it, it says so and hands the thread to the board.
Inbox agent
Email, text, and portal messages land in one queue. The agent sorts each one, drafts the reply, opens a ticket, assigns a manager, escalates to the board, or marks it spam. One outcome per message.
Violations agent (in preview)
A photo and a sentence become a case: matched to the rule it breaks, opened against the right lot, assigned to your manager, with the notice drafted quoting the section. Violation cases, photo evidence, notices and escalation are live today and driven by hand; the agent that drives them is in preview.
Drafts in your board’s voice
Every draft is written against your last twelve months of announcements and the answers the board has already given, so replies stay consistent rather than restating policy from scratch.
A person sends everything, today
Every proposal waits in the queue for a board member, in every category, without exception. The controls that will let low-risk categories send themselves above a confidence line you choose are built and not yet switched on.
A transcript for every run
What it read, which sections it pulled, how sure it was, and what it proposed, written to the record on every run so an agent decision can be reviewed like a board minute. Browsing and exporting those transcripts from the dashboard is in preview.
From zero to running
- 01
Index your documents
CC&Rs, bylaws, rules, recent announcements, and meeting minutes are embedded into a private retrieval index for your community. Nothing leaves your tenant.
- 02
The agent gathers evidence
When a message arrives, the agent reads the thread, searches your governing documents for the sections that apply, checks how the board answered the same thing before, and looks up the sender’s standing with the association.
- 03
It proposes one action
Reply, open a ticket, assign a person, escalate to the board, mark spam, or do nothing. It picks exactly one, and the choice is recorded with its citations and a confidence score.
- 04
A board member releases it
The proposal waits in the inbox with its reasoning, its citations and its confidence, next to the thread that produced it. You send it, edit it, or throw it out. Nothing goes to a resident on its own.
What this looks like in a real association
Saturday-morning resident question
A resident texts the community number at 9 AM Saturday: “can I have a fire pit?” The Answers agent cites the fire-safety addendum and replies. Nobody loses a weekend morning.
The boat in the driveway
A neighbor texts a photo. The Inbox agent recognises it as a complaint, pulls the parking article, and puts a drafted reply and the escalation in front of the board. Opening the case and sending the notice is two taps today, and the subject of the Violations agent preview.
A week of email, triaged
Forty messages arrive over a week. By the time a board member opens the inbox, each one is categorised, twenty-two have a cited reply already drafted, and the rest carry a proposed ticket, assignee or escalation. The evening’s work is reading and sending, not writing.
Delinquency reminder draft
“Draft a firm but polite reminder for owners 30 days past due, mention the Friday cure window.” You get a draft in your community’s voice. Edit, send.
Questions boards actually ask
Where does an agent get its answers?
Only from documents you upload and announcements you publish. We build a private retrieval index per community using pgvector. Every answer cites the source section so you can verify it.
Can an agent send something without us seeing it first?
No. As it ships today, nothing an agent produces reaches a resident until a board member releases it, in every category. Unattended sending, opt-in per category and above a confidence threshold you set, is in preview and switched off.
What happens when an agent does not know?
It says so and routes the question to the board inbox. We tune for refusal over invention: “I do not see this in your governing documents” is the correct answer when it is the true one.
Can we see why it did what it did?
Every proposal arrives with the document sections it cited, a confidence score, and its reasoning, so you can judge it before you send it. The fuller run transcript, listing each tool the agent called, is written to the record on the tool-using agent path; browsing and exporting those transcripts from the dashboard is in preview.
Is our data used to train AI models?
No. Your documents and member data are never used for model training, and documents stay in your own tenant.
Can residents tell when an answer came from an agent?
Yes. Agent replies are labeled and carry the source citation. Residents can escalate to a human reply with a single tap.
What tier is this on?
Pro tier. Lower tiers get template-based compose without document grounding.
Put the agents to work in your association
Sign up free, invite the board, and see how this fits your community in an afternoon.