Where the market is today
Four jobs make up nearly all real-world deployments:
Leasing: after-hours inquiries, prospect qualification, tour scheduling, follow-ups
Maintenance: request intake, diagnostic triage, vendor routing, emergency de-escalation
Resident communication: ledger questions, rent reminders, renewal notices, portal support
Marketing and reputation: listing copy, outbound campaigns, review responses across many locations
Timing is the main driver. According to Respage, a multifamily marketing platform, most rental inquiries arrive after 5:00 PM, when leasing offices are closed. Some of this automation already ships inside the platforms themselves. In Yardi’s ecosystem, agents categorize invoices, check utility charges against past baselines, and draft renewal offers without human involvement.
So the real question for an engineering team isn’t “can this be built?” It’s “should we build it?”
Question 1: Is the agent the product?
This is the first filter. If the client is a property operator that wants to automate its own leasing office, a custom build almost never wins against established vendors, who spread their R&D, voice-model tuning, and PMS maintenance across thousands of buildings.
The picture changes when the client is a proptech platform, a marketing agency, or a software vendor, and the agent is a feature they sell. Respage is an example: its AI leasing agent is a capability it offers to its own property management clients. In that situation, owning the orchestration logic, the IP, and the unit economics matters. Reselling someone else’s white-labeled chatbot doesn’t build a defensible product.
Question 2: What will the agent read and write?
Every job maps to concrete integration points, mostly in the property management system (Yardi, RealPage, or AppFolio). Listing them early is the fastest way to estimate real scope.
Leasing
- Reads: live unit availability, unit-level lease rules, the tour calendar, income and credit thresholds
- Writes: booked tours, follow-up reminders
Maintenance
- Reads: the resident’s description, photos or video, the approved vendor list, contractor availability
- Writes: structured work orders, scheduled repair windows, self-help guidance for minor issues
Resident communication
- Reads: the resident ledger, the lease document, rent matrices, lease expiration dates
- Writes: answers (for example, about prorated charges or notice rules), renewal offers, signed addenda sent back to the PMS
Marketing and reputation
- Reads: floor plans, amenities, concessions, vacancy changes, reviews from Google, Apartments.com, and Yelp
- Writes: updated listings on syndication portals, drafted review replies, flags for recurring complaints
Add the channels on top: email, SMS, web chat, and voice. A project that sounded like “a leasing chatbot” quickly becomes a multi-system integration.
Question 3: Is the source data clean?
The agent can only be as accurate as the PMS behind it. If availability, concessions, pet weight limits, or vendor contacts are outdated, or kept in spreadsheets outside the system, the agent will give polished, confident, incorrect answers.
Put data cleanup on the project plan before the agent goes live. If the client can’t commit to that, the project is at risk regardless of how good the model is.
Question 4: What happens when nobody is watching?
After-hours is both the business case and the biggest risk. It’s when the agent captures tours that would otherwise be lost, and it’s when no manager is around to catch mistakes. The article highlights three scenarios to plan for: misreading a fair-housing question, quoting a rent concession that has expired, and failing to escalate an emergency like a water leak.
The requirement is clear: hard guardrails, deterministic fallback rules, and strict escalation triggers before the agent handles evening traffic. Escalation isn’t a failure state either. Respage’s agent resolves 62% of first-touch inquiries on its own and routes the other 38% (complex or unusual cases) to people.
Question 5: One operator or many tenants?
Off-the-shelf tools assume one company managing its own portfolio. Platforms serving dozens or hundreds of management companies need much more:
- rules and escalation paths configured per client
- brand voice guardrails per property owner
- logical or cryptographic data isolation between clients
- multi-tenant admin APIs
- token usage metered per tenant
- granular permission layers
These are rarely available in tools built for single operators, and they’re often the strongest technical argument for building.
Question 6: Will the volume pay for it?
Custom builds in this space range from five to six figures, so the transaction volume has to justify the spend. The reputation agent built for the Localyser multi-location review platform paid for itself within three months, because it deflected around 1,100 hours of manual triage every month across 1,200 locations. The same system across 20 locations would never earn back its cost.
Question 7: Who will actually work with it?
Most underperforming deployments fail on adoption, not technology. Each role interacts with the agent differently, and each finds its own workaround when it doesn’t trust it:
- leasing staff override the tour calendar and double-book units
- accounts payable manually re-checks every invoice the agent categorized
- maintenance supervisors leave dispatch rules unchanged, so technicians ignore the AI’s notes
- residents keep phoning the office if nobody introduced the new channel
Treat role-specific, ongoing training and post-launch measurement as deliverables. Set baseline metrics before launch and review them after the first month.
Benchmarks from three delivered builds
Respage: commercial AI leasing agent
- Team: 2 AI engineers, 1 full-stack developer, 1 MLOps engineer, 1 PM
- Timeline: 14 weeks
- Cost: $140K–275K
- Result: 62% of first-touch inquiries resolved without a human
Street Digital: multi-agent customer acquisition
- Team: 2 AI developers, 1 backend engineer, 0.5 PM, 0.5 QA
- Timeline: 10 weeks
- Cost: $115K
- Result: 73% less manual research, 3.2x outbound capacity with no new hires
Localyser: reputation agent across 1,200 locations
- Team: 1 AI specialist, 1 backend engineer, 0.5 QA
- Timeline: 6 weeks
- Cost: $45K
- Result: response time cut from 18 hours to 4 minutes, payback within 3 months
When the answer is “buy”
A custom build only makes sense when all three are true: the agent is part of a product sold to customers, the platform needs multi-tenant isolation, and volume is high enough to recover the engineering cost.
It’s a clear “no” if an existing tool already covers the need, the portfolio is under a few hundred units, or nobody can name the metric the system is supposed to move. In those cases, mature options fit better:
- EliseAI for multifamily leasing
- AppFolio Realm-X or Yardi Agents for automation inside those PMS platforms
- Showdigs for scattered-site portfolios
A basic chatbot can go live in a day, and a deeply integrated agent takes several weeks.
Wrapping up
The most valuable engineering work on these projects often happens before any code is written: mapping integrations, checking data quality, defining guardrails, and confirming that the business case holds. For proptech platforms where all three conditions are met, that groundwork leads into the multi-tenant architecture described above, the kind of work involved in AI development for proptech platforms.