Every leasing director has been pitched "AI" that turned out to be a scripted chatbot behind a new interface. Now the same vendors promise autonomous agents that qualify prospects, book tours, and follow up on their own. Sorting substance from marketing requires a maturity model: a clear account of what each stage of conversational AI actually does, and where human accountability must remain no matter how capable the software becomes.
The short answer: chatbots respond, agents act
A chatbot is reactive automation. It matches a question to an answer. An agent is something categorically different. McKinsey defines gen AI agents as AI-powered software entities that plan and perform tasks or deliver specific services on a person's behalf, orchestrating complex workflows and applying logic along the way . Harvard Business Review describes the progression in three phases: assistant, then concierge, then agent, with agents eventually completing entire flows of tasks proactively on a user's behalf . The distinction matters because most deployments today are shallow, with assistants bolted alongside existing workflows rather than deeply integrated agents, which is a major reason widespread AI deployment has produced minimal revenue impact so far .
In leasing terms: a chatbot answers "do you allow pets" at 11 pm. An agent qualifies the prospect, checks availability, books the tour, sends the confirmation, follows up after a no-show, and logs all of it to the guest card. Those are different categories of software, and they carry different categories of risk.
The three stages in practice
Stage one is the scripted chatbot. It deflects frequently asked questions and captures leads, and it has a hard ceiling. A documented case from banking illustrates the limit: ING's classic chatbot resolved 40 to 45 percent of chats, leaving roughly 16,500 customers a week still needing a live person .
Stage two is the guardrailed generative assistant. In the same case, a joint ING and McKinsey team built a gen AI chatbot in seven weeks that offered tailored answers while applying a series of guardrails before any answer reached a customer, with risk stakeholders involved from the start and domain-specific rules, such as refusing to advise on mortgage products . The leasing analogue is obvious: fluent answers about amenities and availability, with fair housing topics, pricing commitments, and concessions treated as controlled territory.
Stage three is agent execution: software that plans across steps, acts in connected systems, and escalates what it cannot resolve. Even at this stage, not every interaction should be automated the same way; some situations will still call for human involvement . And the timeline deserves skepticism. The contact center of the future is likely an AI-led environment, but the pace of getting there is far less certain than vendor predictions suggest, because prior technology waves stalled on system integration and change management .
The trust boundary: where accountability stays human
Guardrails can identify and filter risky outputs, but they do not guarantee an AI system is safe, fair, or compliant; they must be paired with procedural controls such as trust frameworks, monitoring software, and testing practices . Trust is the gating factor for scale: 91 percent of respondents in McKinsey research doubt their organizations are very prepared to implement and scale AI safely and responsibly, and trust is described as the foundation for adoption . Transparency belongs in that trust stack. Stanford HAI researchers have proposed that autonomous AIs identify themselves as AI when asked, citing tests in which Google Duplex explicitly denied being a robot .
The practical boundary for leasing: let AI carry volume and speed, but keep a named person accountable for disclosure policy, escalation paths, exception handling, and regular review of what the system said and did. An agent that books tours with no accountable owner is not mature automation. It is unaccountable automation.
Why this matters, and what it requires operationally
The pressure is real. Despite years of efficiency technology, customer care leaders still report rising call volumes, consistent attrition, and talent shortages, and roughly 80 percent of organizations anticipate increasing AI investment . In residential real estate, McKinsey observes a premium of up to 15 percent between the highest- and lowest-performing players in a market, with high performers adopting centralized leasing and renewal teams alongside digital touchpoints . Yet real estate has been a historically slow technology adopter, and many organizations struggle to scale gen AI beyond pilots; advantage does not come from simply deploying a model .
That last point is the operational lesson. An agent is only as good as the workflow around it: shared context on each prospect, clear handoffs when it escalates, and visibility into what it did. This is where centralized execution differs from bolting a bot onto a website. Accolade's Call Center approach brings phone operations into one command center, so AI-supported conversations, human follow-up, and escalations run on centralized call-handling workflows with shared context, and leaders keep portfolio visibility into what was said, promised, and left unresolved. The maturity model is not really about software climbing stages. It is about redesigning the operation so that when software does more, people can still answer for all of it.
