Superintelligence Prep for Property Managers: Repairs and Owner Reports
An AI agent can keep repair history and draft owner reports, but a property manager must set access and spending limits before any vendor is dispatched.

An AI agent for a property management company can keep a running record of repairs by unit, draft owner reports from that record, and prepare work orders for approved vendors. The useful version is not the one with the broadest access. It is the one whose memory, permissions and spending limits were written down by the manager before the first vendor was contacted.
Key takeaways
- A repair agent is only as useful as the unit-level history it can read, so the first task is defining that record and who maintains it.
- Owner reports should be drafted from logged facts that point back to their source, then approved by a named person.
- Spending ceilings belong in the system as written rules for each job type and unit, not in someone's memory of what was agreed.
- Give the agent narrow, revocable access to the maintenance system, the approved vendor list and the work order queue, and nothing else.
- Judge the agent by an observable criterion, such as whether every sentence in an owner report traces to a logged event.
What a compounding repair agent should remember
An agent improves over time only if it keeps useful context across cases. In property management, that context is specific to each unit: prior repairs, the vendor who did the work, warranty dates, recurring tenant complaints, access instructions, and the owner's preferences about which decisions need approval. The memory should be structured records tied to units and work orders, not a long pile of chat transcripts that nobody can search.
A common pattern is to keep the source records in the systems the company already uses and let the agent retrieve the relevant pieces when it needs them. The Google Cloud architecture guide on retrieval-augmented generation describes this approach as grounding answers in an external store rather than in the model's general knowledge. For a manager, the practical benefit is that every statement the agent makes can be checked against a record a person can open.
Tenant data needs its own rules. Keep only what a repair decision requires, set a retention period for closed requests, and decide who can see tenant contact details. If you operate in Brazil, the national data protection authority, ANPD, publishes guidance on personal data handling that is worth reviewing before the agent starts storing tenant information.
How a repair request moves from intake to vendor
The workflow should be written as steps with a clear owner at each one. A workable version looks like this:
- Intake: a tenant request arrives through an approved channel, and the agent records the unit, the date, the description and any photos.
- Triage: the agent matches the issue to a known category and checks the unit history for recent repairs of the same kind.
- Authorization check: the agent confirms the job type is on the approved list and that the estimate sits under the ceiling for that job type.
- Dispatch: the agent creates one work order with a unique reference, so a retry after a timeout cannot create a second order for the same problem.
- Vendor update: status changes come back through a verified notification and are written to the record, not pasted into a chat.
- Close-out: the agent requests the invoice and completion photos, compares the amount to the approved estimate, and flags any difference for a person.
The retry rule matters more than it first appears. Payment and webhook providers document idempotency keys and signed notifications for exactly this reason: a system that hears about the same event twice should act once. The same discipline applies when a repair agent talks to a vendor. For a closer look at how a dispatch flow handles callbacks and rescheduling, see how a plumbing dispatch agent handles callbacks.
Which access limits to set before any vendor is dispatched
Access should be decided before the agent is connected to anything. Start from the smallest set of permissions that lets the job happen, then add only what a specific failure shows you need. A starting checklist:
- Read access to maintenance records for the units under management, and no access to the full portfolio unless the role truly requires it.
- Write access limited to creating work orders and adding notes, with no ability to edit vendor bank details or approve invoices.
- An approved vendor list maintained by the manager, with trade, service area and insurance status recorded for each vendor.
- Tenant contact details visible only while a specific request is open.
- Credentials scoped to named permissions and revocable without touching the vendor records themselves. The OAuth 2.0 model of scoped, time-limited tokens is a useful reference for how to think about this.
- An audit log that records what the agent read, what it wrote, and which rule allowed each action.
- No direct access to rent collection, deposits or owner distribution accounts.
The OWASP authorization guidance makes a point worth repeating in this context: each request should be checked against the permissions of the caller at that moment, rather than assuming an earlier approval still holds. Tenant messages also deserve care. A message can contain text that reads like an instruction to the agent, such as a request to change a vendor or waive a fee. The OWASP guidance on prompt injection recommends treating that text as data to be evaluated, not as a command, and the approval rules above are what keep such a message from becoming an action.
Every permission the agent holds should be one you could name, justify and remove in a single change.
[FIGURE]
Setting spending limits and approval thresholds
Spending authority is where most of the risk sits, so the rules should be explicit enough that two people would reach the same decision from them. The manager sets the ceilings in writing and reviews them on a fixed schedule. The agent works inside those rules and does not decide what the rules should be. A decision guide that works for many portfolios:
- If the job is routine, on the approved list, and the estimate is under the ceiling for that job type, the agent may dispatch and notify the owner afterward.
- If the estimate exceeds the ceiling, or the unit has had a similar repair recently, the agent drafts the work order and waits for a named approver.
- If the issue involves water, gas, electrical or structural risk, the agent escalates to the on-call person immediately and does not dispatch on its own.
- If a vendor invoice differs from the approved estimate by more than the agreed tolerance, the agent holds the payment question and flags the variance for review.
- If an owner has not stated a preference for a unit, the default is that approval is required.
The agent should never hold financial authority in the sense of moving money. It proposes, drafts and records. A person with payment authority acts on the proposal, and the audit log shows who did what.
Owner reports that stay tied to the record
Owner reports are where a compounding agent can either save hours or quietly create liability. The safe pattern is a fixed report structure, drafted only from logged events, with each claim linked to its record number. Where the record is silent, the report should say so instead of filling the gap with a plausible explanation.
Illustrative scenario: an owner asks why a water heater in one unit was replaced twice within a short period. The agent pulls both work orders, the vendor notes, the tenant reports and the photos. It drafts a summary that states what each record says, then marks one gap clearly: no inspection note exists between the two visits. The property manager adds the missing inspection detail, corrects one sentence, and approves the report before it goes out.
The same discipline applies to monthly summaries. Each figure in the report should come from a source the manager can open, and the report should list open items separately from closed ones. For incident-heavy operations, a parallel example from another field shows how logs become a shared memory: one agent memory for security firms built around post orders and incident logs.
How Be Human handles this in managed operations
Be Human designs this kind of role in a first engagement by defining one useful function, such as repair coordination for a named portfolio, along with the memory it needs, the tools it may call, the approval boundaries and an observable success criterion. Managed operations then keep those boundaries honest over time: monitoring of dispatch logs, recurring evaluations of owner report accuracy, permission reviews when vendors or staff change, integration maintenance when the maintenance system updates, and incident response if the agent acts outside its limits. The NIST AI Risk Management Framework is a helpful reference for mapping these risks to owners and review points before anything goes live. Take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent dispatch repair vendors without owner approval? It can, but only for job types and amounts the owner has pre-approved in writing. Anything outside those rules, including emergency-risk issues and estimates above the ceiling, should wait for a named person. The agent's authority should be as narrow as the rules you can state.
What should a property manager never let the agent change? Vendor bank details, invoice approvals, payment and deposit systems, and the approved vendor list itself should stay under human control. The agent can read and draft in those areas, but changes to them should require a person and leave a record.
How do you know an owner report is accurate? Accuracy comes from structure: every claim links to a logged event, gaps are stated rather than filled, and a person approves the report before it is sent. A periodic sample check, where a manager traces several sentences back to their records, is a simple way to confirm the process still works.
Does the agent need to keep tenant records indefinitely? No. Keep only the tenant information a repair decision requires, set a retention period for closed requests, and limit who can view contact details. Review those rules with whoever handles data protection for your company, especially if you operate under Brazilian or EU requirements.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


