An AI that drafts a listing or summarizes a lease is now ordinary in real estate software. The change worth watching is software that can carry out work inside the systems a business actually runs on.
That change is easy to underestimate. In a demo, an agent that schedules a viewing or processes a renewal can look straightforward. In live operations, whether it holds up depends on things a demo rarely shows: how clean the underlying records are, which systems the agent can reach, and what it is allowed to change.
In practical terms, agentic AI is software that works toward a goal by reading data, using tools, coordinating steps across a workflow, and taking permitted actions inside business systems, rather than only producing an output for a person to use.
Key Takeaways
- Agentic AI works toward a goal using data and tools and can take actions within a workflow, while assistive AI primarily generates outputs for people to review or use. The boundary is not always clean, but the operational difference matters.
- Most real estate AI use today is still assistive. The change that matters is execution: an incorrect action can alter a record or trigger a workflow, a consequence an incorrect answer does not carry.
- Once software can act on live systems, the requirements shift to data, integrations, identity, workflow design, permissions, business rules, monitoring, and human oversight.
From AI Feature to Operational System
For years, much of real estate AI has sat on the side of the workflow: drafting a listing, summarizing a lease, answering a question. When AI only generates, the path is short. A person writes a prompt, the model returns an output, and the person decides what to do with it.
When AI acts, the path runs through the business. A goal leads the agent to read data, call tools, carry out steps in the operational system, and return a result the business relies on. The boundary between the two is not always clean, but the longer path changes the engineering question.
Agentic works as bounded autonomy: the agent handles repetitive, lower-risk steps within defined permissions and business rules, while higher-risk or ambiguous decisions are escalated to a person. Human attention moves to the decisions that need judgment.
Consider a leasing workflow. Before the agent sends a reply, it has to resolve who the prospect is and which unit they mean, confirm the unit is still available, check the viewing slot against the scheduling rules, update only the records it is permitted to change, and escalate to a person when a decision falls outside those limits.
That longer path changes what “AI readiness” means. Once an agent can read data, use tools, take actions, and change something in a business system, the surrounding system has to support every step of that chain: data, integrations, identity, workflow design, permissions, business rules, monitoring, and human oversight.
| Dimension | Assistive AI | Agentic AI |
| Core behavior | Generates an answer or output | Works toward a goal using data and tools |
| Execution | No change to live systems | Can make a permitted change in a live system |
| Example task | Draft a renewal notice | Prepare and route a renewal workflow |
| Data and tools | May use supplied or retrieved context | Typically connects to live data and operational tools |
| Human role | Reviews the output | Defines permissions and reviews higher-risk actions |
| Risk | Incorrect or weak output | Incorrect output and unintended action |
PropTech Platforms Are Moving Toward Agents
The change is already visible in the products PropTech teams use.
AppFolio is one example. In a September 2026 update, its Realm-X platform expanded support for multi-step work across leasing, accounting, and resident operations, and the company reported that AI actions completed on the platform were up sevenfold year over year.
In June 2026, AppFolio also announced a connector between Realm-X and Anthropic’s Claude. The integration is designed to keep actions within AppFolio’s accounting and compliance controls, with human oversight for consequential steps. It is a useful pattern: the model drives the interaction, while the platform keeps control of permissions, business rules, and execution.
For live use, what matters is how an agent’s actions are controlled and audited. Access is only one question. What the agent is allowed to do once it has access is a workflow-design decision as much as a security one.
The same pattern appears beyond property management software. In the UK, Reapit announced an AI agent marketplace and a no-code agent builder for estate agencies in April 2026, with components rolling out across the year and agent actions that Reapit describes as logged, auditable, and reversible.
Rightmove, meanwhile, has been adding AI-assisted experiences such as its Ask Rightmove conversational search, introduced in stages during 2026. Across these examples, the common thread is execution inside controls: software acting on live systems rather than only generating text.
Acting Raises the Stakes for Data
AI use is already widespread among real estate professionals. A February 2026 survey of 225 U.S. real estate agents by Realtors Property Resource (RPR), a subsidiary of the National Association of REALTORS®, found that 82% currently use AI and 92% use it or plan to. Accuracy was the top concern, named by 63% of respondents, with compliance also among the leading concerns.
The key difference: a generative assistant can produce a bad answer. An agent can produce a bad answer and act on it.
This is why data quality becomes more important as agents take on real actions. If the underlying records are fragmented, duplicated, stale, or in conflict across systems, an agent can execute the intended workflow using incorrect information. It can also act on the wrong entity when the same property or resident carries different identifiers in each system, even when each record is accurate on its own. The system works as designed, but the outcome is wrong. When two systems disagree, the agent also needs to know which one is authoritative.
The consequences also differ. A poor draft may need another review. An incorrect update to a lease, unit record, payment workflow, or resident message can create operational problems and, depending on the workflow and jurisdiction, compliance risk. McKinsey’s research adds the architectural dimension: data limitations are a major barrier to scaling agentic AI, while integrations, permissions, governance, and workflow design determine how reliably agents can operate.
What Has to Be True Before an Agent Acts
Moving an agent from a convincing demo into live use takes more than the agent itself. Several parts of the underlying system have to work together:
- Connected systems and clear data ownership, so an agent can retrieve and update information across systems with less reliance on manual reconciliation.
- Consistent entity resolution, so the same property, unit, resident, or lease is matched reliably even when each system uses its own identifiers.
- Defined business rules and metrics, so terms such as “overdue,” “vacant,” and “renewal” have explicit definitions and ownership rather than being inferred by the model.
- Workflow design, so the agent’s goal, authority, sequence of actions, decision points, escalation paths, and success criteria are defined before automation is expanded.
- Permissions and human oversight, set as levels of authority (read, draft, execute, escalate), so the agent operates within defined limits and higher-risk decisions reach a person.
- Monitoring, auditability, and error handling, so the data an agent read, the tools it called, the actions it attempted, the systems it changed, and the outcomes stay traceable.
Different failures need different handling: a technical error may retry or roll back, while a business-rule rejection or an uncertain call escalates to a person.
The same framework applies across common PropTech workflows:
| Workflow | Agent action | Critical dependency | Human control |
| Leasing | Qualify a prospect and schedule a viewing | Correct prospect and unit identity, live availability | Approve terms before they are sent |
| Maintenance | Triage a request and dispatch a vendor | Accurate unit and asset records, vendor and cost rules | Escalate high-cost or safety issues |
| Accounting | Code and post a routine transaction | Chart-of-accounts mapping, reconciled source data | Review entries above a set threshold |
Where Agentic Projects Actually Begin
The starting point is the workflow, not the agent. Teams need to define what they want the system to accomplish, which decisions require human judgment, and which actions can safely be automated. That determines which data, integrations, business rules, controls, and success criteria need to be in place.
The workflow comes first. The technical foundation follows from it: reliable data, connected systems, clear business rules, bounded authority for the agent, and the controls to run safely in production. Production-ready agentic AI depends on that surrounding system as much as on the model. It is the kind of AI enablement work ORIL does with real estate and PropTech teams.
