What is real, what is hype, what is coming

As of August 2026, some of these loops have crossed the line from research project to operational product, and others have not. Here is the breakdown as we see it.

What is real today

• Voice AI agents that hold a structured conversation with a tenant or vendor and capture structured fields. Real-time speech models became fluent enough for genuine back-and-forth during 2025, and conversational latency is no longer the blocker it was. This is the load-bearing capability: it is what makes attacking the four loops possible at all, because the loops are phone calls.

  [Our position] Our voice workflows cover the four loops — tenant intake, owner approval, and the vendor loop among them. They are legally compliant (every opener discloses call recording), pinned against drift, and every call is traceable to a database-verified session and its scored evaluation rather than to prose. We hold outbound destinations to a fail-closed allowlist, and our own rule is that a workflow is heard again on a live call whenever the shared calling runtime changes underneath it — results earned against code that no longer runs do not carry forward. No third party has ever been dialed. That is not modesty; it is the actual stage, and you should ask every vendor for the equivalent sentence.

• AI triage that classifies severity, identifies emergency conditions, and routes by urgency. Combine structured input from voice intake with weather data, lease context, and a language model, and you get consistent triage at 11 PM — which is the hour that actually matters, because that is when the alternative is a tired human or a voicemail box.

• Inbound voice agents for after-hours coverage. This moved from “coming” to “shipping” over the past year and is now one of the most common first deployments, precisely because after-hours is where the pain is least defensible.

• AI vendor ranking based on historical performance. Not a research project — basic ML over closed-job outcomes works fine and ships value quickly.

• Structured summarization for owner digests and property manager reporting. Summarize a week of maintenance activity into a clean owner digest. This works well today.

• Standing rules engines for owner approvals. Not strictly “AI” but the surface that makes AI valuable. Codify “HVAC over $500 needs my OK; plumbing under $200, just do it” and let the system enforce it.

What is hype today

• Generic “AI property management assistant” chatbots that promise to do everything. The tell is not the technology, it is the scope: a tool that claims every workflow usually owns none of them end to end. Ask which single workflow it closes without a human touching it, and watch whether the answer is specific.

• “AI-powered” PMS feature bullets that don’t say what the AI actually does. If the vendor cannot tell you in one sentence what specific decision the AI is making, the AI is not really there.

• Any ROI number a vendor quotes without showing you the measurement. The market is currently full of “reduces costs 40%” and “increases conversion 33%” claims sourced to the vendor’s own marketing page. Ask what was measured, across how many properties, over what period, compared to what baseline. We hold ourselves to this too — it is why this playbook cites a source next to every number and gives you none of our own.

• AI-generated tenant communications without human review for fair-housing language. Real legal risk if you skip the review step. The technology can write the message; you still need a human gate.

• AI vendor sourcing that promises “best vendor every time” without showing you how the rankings are built or letting you override them. Black-box ranking will produce wrong outcomes in your specific market and you will not be able to fix it.

What is coming — through roughly the end of 2027

• End-to-end voice closure across sourcing, booking and owner approval — not just intake. Intake alone is the easy half. The hard half is a vendor who counter-offers a different window, an owner who does not answer, and a job that has to close anyway. Independent analysts put broad production deployment of this class of agent roughly 18–24 months out.¹

  [Our position] This is what we are building, and we have run the full loop on supervised calls to our own line. Running it against real tenants and real vendors is the next step, not a finished one.

• Write-back into the PMS. Today most AI overlays are read-only or hand off by export. Whether an AI tool writes back cleanly is the difference between it removing work and it creating a second place you have to keep in sync — which is the most common way these deployments quietly fail.²

• Voice + vision combined intake. Tenant points their phone at the broken thing; the agent sees it and asks the right structured questions in one pass. The models can do this now; the workflow plumbing around it is what is still being built.

What is far off (skeptical through 2028)

• Fully autonomous property management operations with no human in the loop. The technology might get there. Your regulatory + insurance + fair-housing risk surface will not.

• AI that replaces the property manager’s judgment on tenant relations, eviction decisions, or major capital decisions. These are trust + liability surfaces where AI augments, not replaces.

¹ Build.inc, AI Voice Agents in Real Estate: Current Deployments and Where the Technology Is Heading (2026). ² A recurring finding across 2026 vendor reviews: PMS integration quality, not model quality, decides whether an AI deployment reduces workload or adds a parallel data-entry burden.

---