Bounding Agent Authority in Property Maintenance Dispatch
How to design agentic workflows for tenant maintenance requests that limit financial exposure, mandate human approval for costly repairs, and maintain rigorous unit economics per door.
Maintenance requests represent the highest-volume tenant interaction in property management, making them an obvious target for workflow automation. Tenants submit vague descriptions of leaks, broken appliances, or climate control failures at all hours, expecting immediate resolution. Meanwhile, property managers must interpret the issue, verify lease responsibilities, check warranty statuses, and dispatch the appropriate vendor without driving up operational costs.
Deploying an agent to handle this intake and triage presents immediate financial and physical risks. An unconstrained model might authorize a high-cost emergency plumber for a routine issue, committing the property owner to hundreds of dollars in dispatch fees without oversight. The engineering challenge is not understanding the tenant's text, but bounding the agent's authority to act on it.
Defining the limits of automated authorization
Agentic workflows interacting with physical operations must be governed by strict financial boundaries. The system must know exactly what it is allowed to authorize automatically and what requires a human property manager's explicit signature before external APIs are called.
This is achieved by separating the natural language diagnostic capability from the dispatch execution. The model evaluates the tenant's request and outputs a structured classification, which is then evaluated against a deterministic rules engine. This engine holds per-property budget limits, vendor service level agreements, and strict emergency criteria.
Rules engines must override model outputs
Models are probabilistic; vendor contracts are not. Even the most capable models will occasionally classify a non-urgent request as an emergency based on a tenant's emotive language. A deterministic rules engine must sit between the model's classification and the vendor API.
If the model suggests an emergency dispatch for a broken dishwasher at 2:00 AM, the rules engine must intercept the payload. The engine checks the property's policy matrix, identifies that dishwashers are never classed as after-hours emergencies, and forcefully downgrades the priority, regardless of the model's assessment.
- Emergency versus routine classification matrices mapped to specific vendor categories.
- Hard-coded financial limits for automated dispatch, typically capped at standard diagnostic trip charges.
- Lease-specific exclusion checks to determine if the tenant is financially responsible for the specific repair.
- Time-of-day logic that routes non-critical requests to morning queues rather than triggering expensive after-hours rates.
The human approval gate for vendor commitment
When a maintenance request exceeds the automated threshold or hits an edge case, the agent transitions from a decision-maker to a case preparer. Its job is to assemble all necessary context so a human manager can approve, modify, or deny the action in seconds.
This structured handover must include the tenant's original request, the agent's proposed vendor, the estimated diagnostic cost, and the specific property policy that triggered the manual review. By surfacing this data in a strict schema within the management dashboard, the workflow maintains velocity without sacrificing control.
Unit economics of automated triage
The success of an AI workflow in PropTech is ultimately measured on a per-door cost basis. If the cost of inference, orchestration, and human-in-the-loop fallback exceeds the marginal cost of a property manager handling the ticket manually, the system fails its primary business objective.
Practitioners must model the exact cost of the API calls required to process a single maintenance ticket, including retrieval, prompt evaluations, and retries. A multi-step agentic loop utilizing frontier models can easily consume fifty cents per ticket. For low-rent portfolios operating on razor-thin management fees, that inference cost is indefensible. Smaller, fine-tuned models tailored specifically for HVAC, plumbing, and electrical classification are required to keep unit economics viable.
- Calculating token consumption across multi-step reasoning loops for complex diagnostic trees.
- Caching common tenant queries, such as parking or trash schedules, to bypass expensive generation steps entirely.
- Monitoring the human escalation rate; high fallback percentages mean paying for both the inference and the human labor, destroying margins.
Auditing the dispatch trajectory
When a vendor invoices for a repair that the property owner believes was unauthorized, the property manager needs to prove exactly why the dispatch occurred. The system must log the exact state of the rules engine, the model inputs, the retrieved lease clauses, and the API outputs at the moment of authorization.
This trajectory log defends the platform against disputes. It proves that the agent operated within the defined policy boundaries established by the owner, or it highlights exactly where a human operator explicitly overrode the system's recommendation to authorize the work.