Property management has spent decades refining its processes. The workflows around maintenance, rent collection, lease administration, and compliance have been shaped by generations of operators who learned what works through repetition. Those processes are documented, staffed, and generally effective.

AI is forcing a reexamination of them, and not in the way most operators expected. The question is no longer whether a process works when people execute it. It is whether that process can be described precisely enough that software can execute it instead.

Most property managers are already running a stack of software. Adding AI to that stack is a different exercise than adding another platform, because AI automation needs to reach across the workflow rather than sit beside it. “Property managers all have a system of record, the problem is getting the information into the system of record depends on people and that creates lots of opportunity for mistakes,” said Tony Schollum, Chief Technology Officer at Re-Leased. That dependency is the gap. A system of record is only as current as the last person who updated it, and the people responsible for updating it are also responsible for doing the underlying work. Information arrives late, incomplete, or not at all, and the system reflects a version of the property that is always slightly behind the actual one.

Before any automation can function, the information flow has to be described in detail, which means identifying who supplies each piece of information and where it ultimately needs to arrive. Schollum frames this as building a property hierarchy. “You have to first create the hierarchy for each property, where every person that is affected by the property mapped out,” he said. “Anything that occurs in the real world, you have to understand how it needs to push information around that schema.” That includes owners, asset managers, property managers, tenants, contractors, accountants, and anyone else whose work depends on or generates information about the asset. 

A leaking pipe in a third-floor unit touches a resident, a maintenance coordinator, multiple vendors, an accounts payable clerk, and eventually the financial report. Automating the response to that leak requires knowing every one of those connections in advance. Most operators have never documented their workflows at that level of granularity, because people fill the gaps intuitively. Software cannot.

Once automations are configured, the next requirement is verification, and the approach matters as much as the effort. “We can trust agents to do everything but the key is to audit it, particularly at first,” Schollum said. “The best practice is to audit the outcome, not each data point.” Checking every field an agent populates defeats the purpose of automating it. Checking whether the end state is correct is both faster and more informative. If a work order was created, assigned to the right vendor, priced within tolerance, and closed with documentation, the intermediate steps almost certainly ran correctly. If the outcome is wrong, the audit points toward which step to examine. Early-stage auditing should be intensive and then taper as confidence builds.

The instinct is to build one capable agent that handles an entire process end to end, and that approach runs into a constraint in how language models work. They have no persistent memory between turns, which means the full conversation history gets resent with every new interaction. The context window grows with each step, and cost and latency grow with it. Vertical software with AI built into it mitigates this by storing state externally, so the agent does not have to carry the entire history in context. But the structural problem remains if a single agent is asked to do too much.

“It is important to keep the scope of work for an agent small,” Schollum said. “You can ask one agent to do all of the steps in a workflow automation but that gives the agent too much agency. You might not get the same outcome every time.” Consistency is the operational issue. An agent with broad latitude makes judgment calls, and judgment calls vary. In property management, where the same situation should produce the same response every time, variability is a defect rather than a feature.

The alternative is decomposition, where each agent gets one clearly defined objective and the agents work in coordination rather than one handling everything. “Once an agent has a single goal it uses fewer tokens, provides more consistent outcomes, and reduces hallucinations,” Schollum said. All three benefits come from the same source. A narrow scope means less context, which means fewer tokens, faster responses, and less room for the model to drift into territory it was not designed to handle.

The process required to break up a workflow is more extensive than it appears from the outside. A maintenance request looks like a single process to anyone who has handled hundreds of them. Broken into discrete steps, it involves intake, classification, urgency assessment, vendor selection, scheduling, tenant notification, access coordination, work verification, invoice matching, payment approval, and record updating. Each of those is a candidate for its own agent with its own defined goal. “If you completely automate the entire property management process the token will be astronomical,” Schollum said. “There is just too much context for every step of the way.”

Systems built this way can operate across platforms rather than within one. An agent responsible for vendor scheduling does not need to live in the same software as the one handling invoice reconciliation. Each pulls what it needs and pushes what it produces, and the property hierarchy defines where everything goes. That architecture is what allows automation to reach the complex, multi-party processes that have resisted it. The work of getting there is unglamorous. Mapping every stakeholder relationship, decomposing familiar processes into their actual components, and auditing outcomes until the pattern holds is slower than deploying a tool and hoping. It is also the difference between software that records what happened and software that handles it.

The post Your CRE Software Knows What’s Happening. Why Isn’t It Doing More? appeared first on Propmodo.

​