Collections AI pilot — rent reminders that don’t annoy anyone

Prepared 2026-07-23 · from the founders call + four research passes (market, AppFolio, platform, legal) · Full spec: docs/adr/0117-collections-ai-pilot.md (PR #4449, held for Fede’s approval)

Decided on the 07-23 call: while Yale is stuck in IT review, we build a basic delinquency agent and practice at Camellia. Every serious competitor ships this and sells it hard — EliseAI claims 40–52% delinquency reduction, Entrata bought Colleen AI, RealPage ships Lumina reminders, AppFolio has RealmX flows. The read-side of our platform is already built (renewals proved it). What’s left is the outreach loop, the guardrails, and the measurement.

Phase 0 — tenant language preference (ships first)

Fede’s call: this rolls out before any collections touch. Per-tenant languagePreference (en/es, extensible) sourced from application data, conversation inference, and PM-set overrides (Fede collected Spanish/English unit-by-unit on-site 07-21 — seed from that). Every outbound path checks it and translates before send; the mass-send composer shows the translated copy for PM review because raw MT is sometimes bad.

This isn’t just UX. English-only dunning at a heavily Spanish-speaking property is a national-origin disparate-impact risk (fair housing), and it’s the same gap that produced the English water-shutoff texts to Spanish-speaking units this week. It also benefits renewals, maintenance, and mass-send for free.

Phase 1 — who we contact (and who we never contact)

Population: Camellia residents who pay via the AppFolio online portal. Only them. Payment state is real-time, so no false “you’re late” pings while a money order sits in the office for a week. Fede’s estimate was ~40% of the building — measured 2026-07-24: 50.5% (46 of 91 current households pay majority-online); see the ledger audit below. Refresh monthly.

• Hard exclusions: money-order/check payers; Evicting / In Collections / Certified-Funds-Only flags (AppFolio disables their portal — never send them a portal link); Hidden tenants; verified-zero balances (applied payments, see gotchas); active disputes; per-unit PM opt-out (reuse the mass-send unit picker); channel opt-outs.
• Trigger: rent charge unpaid past day 8 (after CO’s statutory 7-day grace) and balance ≥ $50 (matches the platform delinquency predicate). All thresholds per-property config cloned from the RenewalPolicyThresholds pattern — we set Camellia’s values, JP configures nothing.

Camellia ledger audit — measured 2026-07-24

Every number below is 🟢 measured from the live AppFolio Data API (property 7), pulled 2026-07-24. Window: 2025-07-01 → 2026-07-24 (13 months). Sources: receivables_activity (1,618 payment txns), charge_detail (3,964 rows → 2,112 tenant rent-family charges), delinquency (79 rows). Raw pulls + analysis scripts archived in the session scratchpad; re-runnable read-only.

1 · Who actually pays online (the “~40%” check)

Payment-method mix across all 1,618 transactions in the window:

Per-household rollup, current-status occupancies with ≥1 payment in the window (n=91). The cohort definition matters, so here is every reasonable cut:

Verdict: Fede’s ~40% estimate was low — it’s 50.5% by the cohort definition (46 households), 53.3% of dollars. The brief’s “~40–45 tenant” cohort-size assumption holds almost exactly: the pilot cohort is 46 households (44 after the 2 certified-funds exclusions). 13 more households have an activated portal but still pay offline — the conversion opportunity below.

2 · How late Camellia actually pays (rent-family charges, 13 mo)

2,112 tenant charges on rent-family GLs (Rent Income, Parking Space Rental, Pet Rent), days from charge posting (≈ the 1st) to earliest applied receipt:

• 44.6% of all rent charges go past the day-8 trigger. Dollar-weighted: 56.7% of charged rent is collected by day 8, 74.9% by day 15, 82.1% by day 22 — i.e. the day-8→22 cadence window is exactly where ~25% of the rent lands today.
• Days-to-receipt percentiles (paid charges, n=2,042): p50 = 6 · p75 = 14 · p90 = 36 · p95 = 52 · max = 170. The tail is long: 12% of charges take 31+ days.
• Habit distribution (123 households with ≥6 charges): 30 never late past day 8 · 18 late 1–2× · 14 late 3–5× · 61 late 6+×. Lateness at Camellia is chronic-concentrated, not occasional.

The same distribution split by pilot-cohort membership — this is the number pair that actually sizes the pilot:

• Inside the cohort, the pilot’s addressable gap is the ~20% of charges (≈$139k rent-$/yr flowing late) that miss day 8 — and it’s concentrated in 11 chronically-late households (6+ lates each), while 20 of 43 established cohort households were never late once. Expect the cadence to be near-silent for half the cohort and a monthly ritual for a dozen.
• Cohort p50 is 1 day — the typical portal payer pays on the 1st/2nd. Non-cohort p50 is 9 days: the median offline payment would already be past the trigger.
• The behavioral split (20.1% vs 59.8% trigger rate) cuts both ways: (a) the pilot cohort is the better-behaved half of the building, so absolute delinquency-reduction upside inside the cohort is modest — success looks like moving day-8 collection from 80% toward 90%, not a 40% delinquency cut; (b) the money-order “control” group is wildly non-equivalent — any cohort-vs-control comparison must be within-tenant-baseline, never cross-group.

3 · Expected reminder volume (backtest of the day-8 trigger on the pilot cohort)

Distinct cohort households whose rent-family charge was still unpaid at day 8, per month — i.e. who would have received a day-8 reminder had the pilot been live:

• Average ~9.3 cohort households/month trigger day-8 (range 7–15) carrying ~$8–15.6k of late rent. Ops load is tiny; statistical patience must be large — at ~9 cases/month, “direction + case narratives” for the first 4–6 weeks is the honest bar (as the Measurement section already says).
• Cohort lateness is worsening: 7–8/month through Jan 2026, 10–15/month since Feb 2026. Baseline-lock must capture this trend (use the trailing-6-month per-tenant baseline, not the 13-month average), or the pilot will look like it’s failing against a stale baseline.
• Non-cohort (offline payers) triggers run ~35–53/month — that’s where 4× the delinquency lives, and the pilot deliberately can’t touch it. The realistic biggest lever on building-wide delinquency is converting offline payers to the portal (13 already have portals activated), which Phase 0’s language work directly serves. Worth stating to JP so pilot-cohort numbers aren’t mistaken for the whole-building opportunity.

4 · Delinquency snapshot (as of 2026-07-24)

• 27 current+active households carry a balance ≥ $50 right now. Total receivable $37,399, of which delinquent rent $33,447.
• Aging: $28,390 in 0–30 · $5,184 in 30–60 · $3,135 in 60–90 · $690 in 90+ · $100 prepaid. Most of the book is fresh — the recover-it-early thesis has room.
• Flags among current+active: 0 InCollections · 1 CertifiedFundsOnly · 0 PaymentPlan · 1 NSF. (The report only lists tenants with balances — 27 is the delinquent count, not the building population.)

5 · 10-day demands actually served (mailbox census, 2025-07-24 → 2026-07-24)

Source: read-only Microsoft Graph sweep of the camelliaapts@jp-co.com mailbox — the demand traffic with the eviction firm Tschetter Sulzer Muccio PC (TSM). AppFolio's Data API exposes no notice/eviction surface at all (33 candidate report endpoints probed, all rejected), so email is the only system of record. Tenant names withheld here (units only); the full machine-readable census lives in the session scratchpad.

• 6 demand batches in 12 months (~monthly), 2–6 demands each, ~22 line-items across ~16 distinct households — units 106, 107, 118, 122, 201, 202, 214, 217, 314, 410, 422, 508, 517, 612, 618, 620. Unit 517 was served three separate times; 122 and 217 twice.
• Cohort overlap — the founder’s question: 4 of the 46 online-payer cohort households were served a 10-day demand in the window (units 118, 410, 122, 620 — ~9% of the cohort). Three of the four are near-100% portal payers (29/29, 26/27, 15/16 online txns) — being an online payer does not mean never reaching a demand. Outcomes: 118 cured before filing (canceled); 410 filed; 122 and 620 escalated to actual Denver County eviction filings.
• Verified escalations to eviction across the whole building: units 612, 122, 517, 620, 107 (answers/trials in 2026). Verified cures after service: 118, 517 (Jan batch).
• The dominant failure mode is date math: an entire April 2026 batch of 4 demands was bounced by TSM as defective because the pay-or-vacate date gave 9 days instead of 10 (served 4/13, PVD 4/22) — every one had to be re-served from scratch, adding weeks. The same error bounced a 2024 batch. Timing: rent due the 1st → served ~the 14th → PVD ~the 28th → scanned to TSM early the next month.

6 · Data-quality findings that change the build

• delinquency.LateCount is 0 for every tenant — including the 61 households late 6+ times. Do not use it. Compute lateness from charge_detail receipt timing (as this audit did); the brief’s tone-adaptation logic must key off computed history, not AppFolio’s field.
• ChargeDate is the posting date (≈ the 1st for rent). Fine for day-8 triggers; the build’s bucket-aware daysDelinquent upgrade (net-new #4) stays necessary for mid-month charges like parking/pet added off-cycle.
• ReceiptDate is a comma-separated list on split payments — earliest-receipt parsing (already in parseEarliestReceiptDate) is mandatory or lateness is overstated.
• “Flexible Rent” (Flex) appears as its own PaymentType (9 txns) — classify as portal-adjacent but exclude from the cohort: Flex pays the landlord on time by design, so these tenants can’t be late in our ledger even when they’re struggling.
• receivables_activity carries per-tenant PortalActivated, OptedOutOfPortal, CertifiedFundsOnly, MustPayBalanceInFull on every row — the cohort query can come from this one report plus charge_detail; no extra endpoint needed.
• Caveats: “household” = AppFolio occupancy; window is 13 months; parking/pet rent counted as rent-family (they’re on the same statement); late fees excluded from lateness math.

Cadence — the “not annoying” contract

• Day 8Friendly nudge. SMS (email if no SMS consent), preferred language: balance + portal link + “reply here if something’s wrong.” No mention of late fees unless one is actually posted, lease-valid, and cap-compliant.
• Day 15Second reminder, slightly firmer, still service-toned. Explicitly offers help / a payment conversation.
• Day 22Final reminder + PM alerted in-app and in digest. Notes the team may follow up personally. The agent stops here.

Max 3 touches per resident per month — well under the Reg F 7-in-7 cap we self-impose. Stop instantly on: applied payment, promise-to-pay (one polite check-in if the date passes unpaid), opt-out, dispute (→ human), PM pause. Suppression re-checks the ledger at send time, not enqueue time. Quiet hours 8am–9pm resident-local enforced in the dispatcher. Tone adapts to history — first-time late gets the softest copy, chronic late gets neutral-firm, never harassment-adjacent — and the whole tone bank goes through promptfoo evals before launch.

Channels: SMS + email only. No outbound voice in the pilot — the FCC treats AI voice as “artificial voice” (highest consent bar), and the Camellia TFN has a T-Mobile block anyway. Disclosure-forward: first contact in any thread identifies Clara as Camellia’s automated assistant with an easy human handoff. EliseAI does the opposite (human name, disclose only if asked) and a National Housing Law Project litigator publicly called that “unfair and deceptive” — we make transparency the feature.

Human-in-the-loop boundaries

Escalate to PM on: dispute, hardship mention, opt-out, promise broken twice, day-22 reached, any eviction/legal question, any conversation the agent can’t hold in the tenant’s language.

The 10-day demand — as-is workflow, the law, and Clara’s role

Added 2026-07-24 from two parallel deep-dives: a read-only Graph sweep of the Camellia mailbox (the real demand traffic with TSM), and primary-source legal research (C.R.S. §§ 13-40-104/106/108, 38-12-105; JDF 99A rev. 2025-09-04; HB23-1120 / HB24-1098 / HB25-1240). Not legal advice; counsel sign-off required before any of Clara’s demand-adjacent behavior ships.

How the team does it today (reconstructed from the mailbox)

1. Decide: the PM team tracks delinquents in “deliq update” threads; Kenya owns the relationship with the eviction firm (Tschetter Sulzer Muccio PC, client ID 54784).
2. Prepare by hand: the on-site team fills the TSM-provided “Denver Demand for Compliance or Possession” (current rev 01.01ABDen 08/22/25 — a 9-page packet: fillable demand + return-of-service + Government Assistance Affidavit + the mandatory Denver Tenant Rights & Resources pages). It IS a fillable AcroForm — but today it’s filled with pen and scanned.
3. Serve — the property itself serves, not the firm: a staff member personally delivers (or posts after two failed attempts on separate days) and completes the return-of-service block.
4. Wait 10 full days; if uncured, scan the served demands and email them to TSM’s intake (Zendesk ticket, 3–4 business-day SLA).
5. TSM legal-reviews and files the eviction (Summons & Complaint, Denver County) — or bounces the demand as defective, which restarts everything including re-service.

Reality check from the census (audit §5): ~6 batches/year, 2–6 demands each, ~16 households. The single biggest failure: date math — one whole batch of 4 was voided for a 9-day pay-or-vacate window, weeks lost. And 4 of the 46 pilot-cohort online payers were served a demand this year (units 118, 410, 122, 620; two escalated to real eviction filings) — demand territory is not disjoint from Clara’s cohort, so the boundary rules below will actually fire.

What the law requires (verified against current statute + the official form)

• Service must be physical — a human has to do it. C.R.S. § 13-40-108 allows exactly three methods: personal delivery to the tenant; leaving with a household family member over 15; or posting conspicuously only after personal-service attempts on two separate days. Email/SMS/portal delivery is not service — a demand “sent” electronically is a legal nullity. No process-server requirement though: the PM, or any staff member, may serve it themselves (unlike the later court summons).
• The 10 days are full days from the day after service, and premature filing = dismissal. Camellia (~100 units) never qualifies for the 5-day “exempt small landlord” variant.
• Amounts are a minefield: § 38-12-105 bars demands/evictions based on late fees, and misstating the cure amount is the exact theory of the class action against TSM itself. The real served forms itemize “other fees/charges” (utility/late/parking) in a separate block per Denver’s county form — open counsel question whether anything beyond base rent belongs in a Denver demand; safe default is rent-only.
• Mandatory contents: all occupants named, missed-payment amounts+dates, the alternative pay-or-vacate structure, tenant-rights advisement (mediation + victim-survivor), signature by landlord/agent/attorney — and service in the tenant’s primary language (HB24-1098), which Phase 0’s languagePreference directly feeds.
• Status branches that change everything: tenants who disclose SSI/SSDI/Colorado Works → mandatory pre-filing mediation (HB23-1120; Camellia is not size-exempt) — the packet’s Government Assistance Affidavit swears to this. Subsidy tenants on covered dwellings → 30-day notice, not 10 (HB25-1240). Section 8 → housing authority notified at demand time.
• Filling out the form is not practicing law. Software populating the official form from landlord data, for the landlord’s agent to review, sign, and serve, is scrivener work under Colorado’s UPL doctrine. The line: Clara must never select the legal theory, advise the tenant, or present itself as counsel.

Clara’s role — proposed, counsel-gated, and 100% behind the /review queue

1. Demand drafting (“Clara fills the PDF for the team”): Clara populates the current Denver demand AcroForm from ledger + tenant data — occupants, unit, base-rent-due with missed months, pay-or-vacate date, community fields — and emits a print-ready PDF into a /review item alongside the evidence (ledger snapshot, language preference, affidavit checklist). The killer feature is the date math: a computed, guaranteed-compliant PVD (≥10 full days from day-after-service, with the service-date entered by the human server), eliminating the failure mode that voided the whole April batch. PM approves → prints → signs → serves. Clara never signs, never serves, never emails a demand to a tenant.
2. Pre-demand warning (“should Clara warn them?” — yes, narrowly): lawful as a disclosed first-party agent IF factual, rent-only, non-certain (“a formal 10-day demand may be the next step”), and only when JP actually intends to proceed. Design: fires only at the day-22 escalation stage, drafted by Clara but held in /review for PM approval before send during initial rollout; suppressed entirely for any tenant with a disclosed SSI/SSDI/Colorado-Works status, subsidy tenancy, active dispute, or existing served demand. The word “eviction” stays out of Clara’s vocabulary — “formal demand” is the ceiling.
3. Cure-clock + packet tracking: once a human records service (date/method/attempts), Clara tracks the 10-full-day clock, reminds the team the day it expires, pre-fills the return-of-service data for the server to sign, and assembles the scan-to-TSM packet checklist (signed demand + return of service + tenant-rights pages + affidavit). Kills the second failure mode: premature or sloppy submissions.
4. Suppression feedback loop: a served demand immediately hard-stops all cadence messaging for that tenant (it’s an active legal process — reminder copy could contradict the demand); cure/payment restores normal state.

Not doing, ever (this pilot): electronic “service” of anything; auto-sending warnings without review-queue approval; stating amounts that include late fees; contacting anyone post-filing; drafting anything for the court stage (JDF 101/102 is TSM’s job).

Open items for counsel: rent-only vs. Denver’s “other amounts” block; first-party status of Clara’s messaging (the pivotal classification); HB25-1240 effective date + covered-dwelling scope. Open item for Fede: confirm who signs/serves on-site today (the served scans show a signer whose name doesn’t match our team roster — verify before wiring the signature workflow).

What the platform already has vs. what you’re building

Already built (reuse, don’t rebuild)

• Ledger-accurate delinquency — getDelinquencyProfile() (src/lib/domain/leasing/delinquency/index.ts:30) computes balanceMonths / pastDueRent / 12-mo latePayments live from AppFolio charge_detail; fails closed (null = unknown, never “not delinquent”). Battle-tested by renewals.
• Balance snapshots — TenantBalanceSnapshot with bucketedAR aging (types.ts:1557), hourly lambda/appfolio-sync, refuseZeroRegression guard; canonical read getCurrentBalance().
• Threshold config pattern — RenewalPolicyThresholds + per-property delinquencyThresholdMonths (types.ts:1031).
• Tenant balance tool — lease_and_billing.get_tenant_balance with server-resolved identity (never LLM-supplied), text + voice.
• Outbound rails — one consent-gated dispatcher (agents/clara/lib/messaging/dispatch-envelope.ts): Twilio SMS w/ TCPA consent lookup, Graph email, ElevenLabs voice. Temporal autonomous-send precedent from renewals.

Net-new build

1. Adapter reads: wire the AppFolio delinquency report — AppfolioDelinquencyRow (integrations/appfolio/types.ts:351) is fully typed (PaymentPlan, InCollections, NSF, LateCount, LastPayment…) but nothing calls it — plus payment history (receivables_activity/charge-receipt join). Through getPMSClient only; add to the domain interface so Yardi/RealPage can implement later.
2. CollectionsCase per tenant-month: ELIGIBLE → REMINDED_1..3 → PAID | PROMISE | ESCALATED | SUPPRESSED, full audit trail (message, timestamp, channel, ledger snapshot behind every stated balance). Keep ≥ 4 years. Plus a PromiseToPay record (amount, date, kept/broken).
3. Temporal cadence workflow — same doctrine as renewals (ADR-0108); remember the 05-26 SQS-bypass incident. Send-time suppression re-read.
4. Tool upgrades: real lastPayment* (currently hardcoded null, handle-get-tenant-balance.ts:37-41) and bucket-aware daysDelinquent (currently assumes due-on-the-1st).
5. Language preference field + translation layer + mass-send review UX (Phase 0).
6. PM surface: collections cases on a simple list + per-unit opt-out picker; a per-tenant AR view can come later.

AppFolio mechanics that shape the design

• The Data API is read-only. Every write (post/waive fees, credits, payment plans, notices, flags) is UI-only → browser automation or human. Pilot keeps writes human.
• Unapplied prepayments still read as unpaid — “already paid” suppression must check applied payments, not money-on-account.
• Status flags kill the portal — Evicting / In Collections / Certified-Funds-Only tenants can’t pay online; a portal link to them is a broken promise.
• Payment plan = edit charge → split into dated child charges; visible via the payment_plans report and PaymentPlan column.
• Hidden flag suppresses tenant ops — check visibility before acting.
• Collections handoff (Hunter Warfield, past residents, >$100, DOB/SSN) is post-resident FDCPA territory — out of scope.

Legal — verdicts and hard constraints

Research 2026-07-23; counsel sign-off required before go-live. Not legal advice.

Market notes worth stealing (and avoiding)

• Steal: operator-configurable cadence (nobody real ships a fixed ladder) · tone adaptive to payment history · answering balance questions inside any existing thread · PM alert thresholds with drafted-not-sent legal notices · 24/7 responsiveness.
• Differentiate: the industry’s weakest documented dimension is payment-plan guardrails — no vendor publishes approval authority, dollar/term limits, or promise-kept tracking. Our explicit PM-configured menu + PTP kept/broken tracking out-specifies all of them. Second differentiator: disclosed AI.
• Avoid: undisclosed human-named AI (EliseAI’s exposed flank) · post-resident FDCPA recovery (Pay Ready’s territory, different legal regime) · competing with Flex/Best Egg (flexible-rent rails are prevention, not recovery — complementary).
• Claimed numbers for the sales narrative (vendor-reported, un-audited): EliseAI ~40–52% delinquency reduction, Brookfield 97.6→99.6% collection rate 14 days faster; Colleen/Entrata −40% unpaid rent; Asset Living +600 bps on-time. Sean’s packaging idea: sellable add-on, ~$0.25/unit/mo.

Measurement — proving we reduce delinquency

Lock the baseline before the first send: 12 months of charge_detail history. The portal cohort is 46 households (44 after certified-funds exclusions) with only ~9.3 day-8 triggers/month (audit §3), so month-one truth is direction + case narratives, not p-values — report absolute numbers. Baseline must be the trailing-6-month per-tenant window, because cohort lateness has been worsening since Feb 2026 (7–8 → 10–15 triggers/month) and a 13-month average would flatter the pilot’s starting point.

• Primary: % of pilot-cohort rent collected by day 15 / day 30 vs. each tenant’s own trailing-6-month baseline, with the money-order population as a concurrent (non-equivalent) control.
• Secondary: avg days due-date→applied-payment · promises made/kept · % cases resolved with zero PM touches · escalation rate.
• Kill criteria (pause + review): opt-out >5% of contacted residents · any validated wrong-balance message · >2 complaints to the JP office · any protected-class disparity in the monthly audit.
• Track bad-debt %-of-GPR and delinquent-resident headcount as separate metrics against Camellia’s own trailing baseline; trust no trend before 4–6 weeks (pm-domain-knowledge benchmarks).

Suggested build order

1. Phase 0 language preference + translation review in mass-send.
2. Adapter reads (delinquency report + payment history) → cohort query with all exclusions. Portal-payer share is now measured (50.5%, 46 households — see audit); productionize that query (majority-eCheck/CC over trailing 13 mo, minus CertifiedFundsOnly/Flex) and refresh it monthly. Do NOT read delinquency.LateCount — it’s zero for everyone (audit §5); compute lateness from charge_detail.
3. CollectionsCase/PromiseToPay + Temporal cadence with send-time suppression; tone bank + promptfoo evals; opt-out tool.
4. PM surface (case list + unit opt-out) + 10DLC campaign registration + carrier smoke test.
5. Baseline lock → counsel review of tone bank → live on the cohort. Test Property (ID 45) or log-mode only until launch day — no real comms in test runs.
6. Demand assist (parallel track, counsel-gated): new /review item types (demand draft, pre-demand warning, cure-clock expiry) on Gera’s review-queue primitive → AcroForm fill of the current Denver demand + compliant-PVD computation → cure-clock/return-of-service tracking → TSM packet checklist. Ships nothing tenant-facing until counsel signs off on the warning copy and the rent-only amounts question. See the demand section.

Sources: founders call 2026-07-23 (raw transcript) · EliseAI/competitor research pass · AppFolio KB pass (help-articles + Data API) · repo capability inspection (paths cited inline) · first-party collections legal research (CO/TCPA/FDCPA/Reg F) · Camellia ledger audit 2026-07-24 (live AppFolio Data API: receivables_activity / charge_detail / delinquency, property 7, 2025-07-01→2026-07-24). Full citations live in ADR-0117 and the research reports behind it. Vendor metrics are marketing claims, not benchmarks.