Vision v4 vs. Reality — Gap Analysis
Superseded (2026-08-18): this page is kept as history. The living architecture reference is now Architecture — Source of Truth.
How far the system is today from the long-term technical vision ("Clara does the work"), what we already do well, where the gaps are, and the five opportunities that close the most distance. Evidence gathered by four read-only code investigations of the production repo on 2026-08-16; every claim below was checked against code, not docs. The vision is a multi-year north star, so "gap" here means distance still to travel, not a miss.
The vision in one line: "The PMS remains your system of record. Clara becomes your system of work." Clara is a third-generation AI operator: she has context, she has goals (measured on outcomes), she has judgment with boundaries — she acts within policy, asks when outside it, and the answer becomes policy so she never asks again. PMs delegate outcomes, not tasks. One brain across leasing, maintenance, turnovers, renewals, collections. Every property makes her better.
Scorecard — vision promise vs. what's in production
| Vision promise | Today | What we found |
|---|---|---|
| Leasing is a goal (book → follow up → application), not Q&A | Real | Multi-touch follow-up sequences (email → text at +24h/+48h/+7d/+14d, post-tour application chase, re-engagement after silence) run by default on every property with no switch, and stop the moment the goal is met.src/lib/domain/leasing/outreach-cadence-schedule.ts, outreach-stop-conditions.ts, reengagement-trigger.ts |
| Maintenance runs on its own (intake → triage → dispatch) | Real | Send-gate removed 2026-06-11; every work-order change dispatches to AppFolio unconditionally. Most autonomous loop we have. |
| "Your answer becomes policy" | Real | A staff reply to an escalation is written straight into that property's knowledge base, no review step, live on the next conversation.src/lib/domain/escalation/teaching-capture.ts |
| Judgment with boundaries, fair-housing escalation | Real | Fail-closed guards throughout, two-key arming on every send path that touches money, hard human gate on turnover charges and on every collections message, escalation state machine with reminders and takeover. |
| Intelligence: "which residents, what they said" | Seed | The renewal-risk signal names specific residents and cites the exact email where they said they're leaving. But it's a once-a-day batch, the AI-written narrative is staff-only, and nothing equivalent exists for vendors, maintenance, or collections. No "ask why" chat.src/lib/domain/insights/signals/renewal-risk.ts, lambda/portfolio-insights-snapshot |
| Measured on outcomes (occupancy, resolution), not volume | Half | Funnel-leak, tour no-show, vacancy days, leasing velocity are computed — but only rendered as a report. The Command Center headline is counts and dollars-saved. No metric ever changes what Clara does next. |
| "She never asks that question again" | Fragile | Only near-identical wording matches; no check of learned answers before escalating (relies on the model noticing); and the whole knowledge block is capped at 4,500 characters — as a property learns more, older answers silently fall off and Clara says "not on file."agents/clara/lib/agent/property-knowledge-block.ts (KNOWLEDGE_BLOCK_MAX_CHARS) |
| PM can call/text/email Clara and delegate an outcome | One flow | Only the move-out/turnover flow executes from a PM text end-to-end. A PM emailing Clara isn't recognized as staff. Dashboard "Ask Clara" has 23 lookup tools and zero that change anything. The permission table already authorizes staff for far more — the tools just aren't offered.agents/clara/lib/agent/capabilities/index.ts, src/lib/tools/role-matrix.ts, src/lib/domain/dashboard/chat/tools.ts |
| "Drop rents by $150" / "assistance letter" / "work collections" | Not yet | Bulk rent change is marked "planned" and never offered. No document generation exists. Collections is a background cadence, not something a PM can invoke. |
| One brain: renewal knows maintenance history | Not yet | Zero references to work orders or complaints across all 30 renewal files. |
| One brain: turnover knows the vendor's track record | Not yet | No vendor score/rating/reliability field exists. Job outcomes are written after completion but never read back into vendor choice. |
| One brain: invoice knows the work order | Real | Invoices live inside the work order record; charge decisions reason about estimate-vs-actual. |
| Every property makes her better (compounding) | Engineer-mediated | Learned answers are hard-walled per property (correct for privacy). The one cross-property path — promoting a "principle" portfolio-wide — is a data type with no consumer yet, and by decision always a human act. Today compounding = engineers changing the prompt. |
| Five loops run on their own for a new property | One of five | Out of the box only prospect follow-up runs unattended. Renewals shadow-only; turnover auto only under $500 after arming; collections can never send without a click. Each needs an engineer-flipped company switch plus a per-property switch.src/lib/domain/admin/arms-registry.ts |
| Logs into your PMS like a new hire, no migration | AppFolio only | AppFolio: self-serve, 1–3 hours. Yardi/RealPage: "coming soon" tiles; the shared PMS layer is ~¼ wired and the highest-stakes writes call AppFolio directly. Yale is a hand-built, months-long project. |
| Fully audited | Half | Tool-call log exists but expires after 30 days and covers ~39 of 74 action handlers. No single "why did Clara do this" record. |
What we do well
- The safety posture is engineered, not prompted. Guards fail closed, money-adjacent sends need two keys, and the team pulled two well-intentioned "honesty" features within hours on 2026-08-15 when they blocked correct answers to real Camellia customers — after replaying 177 real conversations to size the blast radius. "We review every conversation" is true.
- Leasing follow-up is genuinely goal-driven and on everywhere. This is the clearest existing proof of the "Gen 3" claim.
- Maintenance is autonomous end-to-end.
- The teaching loop exists and is live — the mechanism behind the deepest moat in the vision is already in production.
- Quality machinery is unusually deep for the team size: 30+ eval suites, real-vendor-call replay, live voice simulation tests, prompt-drift guard, real-time conversation grader paging Slack, Grading Playground turning human verdicts into a regression corpus.
- Identity spine and per-property config are clean; nothing in production code gates behavior on a hardcoded Camellia.
Top 5 opportunities
Ranked by distance closed toward the vision per unit of effort. Each is a direction, not a spec.
1. Make Clara delegable — a real "PM instruction" path
The vision's headline experience and the biggest gap. The authority already exists (the permission table lets staff do work orders, renewals, pricing reads, tenant lookups); the offer doesn't — the tool composer never hands those tools to a PM, and the staff-reply classifier only knows "decision" or "question," never "do this."
- Recognize staff on email (today phone only).
- Add an instruction intent to staff replies; give the PM channel and dashboard chat mutating tools behind confirmations.
- Ship one flagship command end-to-end. "Get unit 204 ready" is closest — turnover already executes; generalize the entry point. Then bulk rent change (already marked "planned").
2. A knowledge base that scales — and "never ask again" as a hard check
The 4,500-character cap is a time bomb: the more Clara is taught, the more she forgets. This directly protects the most defensible promise in the doc.
- Retrieve learned answers by relevance instead of dumping everything.
- Semantic (not exact-string) matching on "was this already answered."
- Consult learned answers before deciding to escalate; add a regression eval for "taught once → not re-escalated."
3. Wire the two "one brain" edges that matter most
- Renewal decisions and conversations pull the tenant's open/recent work orders and complaint signals.
- A vendor scorecard (on-time %, quote-vs-actual, callback outcomes — the data is already written after jobs, just never read) feeding turnover/maintenance vendor choice.
Turns two vision sentences from false to true and feeds opportunity 5's "slow vendor behind it."
4. Close the loop from outcomes to behavior; show outcomes, not volume
- Make the customer-facing scoreboard outcome-first: occupancy, days-to-lease, resolution time, loop-closed rate.
- Let a signal trigger action: a leaky post-tour stage escalates cadence; a renewal-risk hit triggers proactive outreach. Today the goal-chasing code and the outcome-measuring code never talk.
5. Autonomous-by-default for a new property (safely) — arming ladder + PMS layer
- A per-property autonomy profile set during onboarding (shadow → arm per loop with proof gates), self-serve rather than script-flipped.
- Finish the PMS abstraction so renewal/maintenance/vendor writes go through the registry — the prerequisite for Yale/CONAM being anything other than bespoke.
Without this, every customer is a project and "AI runs the routine" needs an engineer per property.
Honorable mentions
- Durable audit trail — drop the 30-day expiry, cover every handler, one queryable decision record. Cheap, and it's what "fair housing built in" rests on.
- Conversational portfolio intelligence — the renewal-risk engine and the anti-hallucination narrative exist; extend to vendors/maintenance and expose the "why" in the product.
- Redesigned honesty guard, gated on the 177-conversation replay before it ships again.
Method & caveats
Four read-only investigations over PropFlow-Technologies/propflowai at HEAD on 2026-08-16 (system inventory; PM-delegation + outcome metrics; cross-workflow context + learning + audit; intelligence + onboarding + arms). Claims marked "Real / Not yet" were confirmed by reading the code paths cited; onboarding effort figures are estimates. Shipping velocity is high, so flag states may have moved since. A related, deeper audit of the coworker gap is at Clara as a Coworker — Capabilities.