A tenant reports a leaking pipe at 8 p.m. An AI assistant sorts the request and prepares a work order. Helpful. But if it can also read every tenant file, change payment details, or send a contractor into an occupied unit, the software has more authority than the job requires.
Google Cloud’s October 22 agent governance announcement describes separate agent identities, access policies and activity records. Microsoft’s agent guidance recommends limited permissions and approval before sensitive actions. Those are useful lessons even for a small property manager with one AI assistant.
Can our AI handle maintenance tickets without seeing everything?
Yes, when the system lets you restrict what the assistant can read and change. Give it access to the request details it needs. Keep payment records, personal documents, lock codes and unrelated tenant files outside that access. If the software cannot enforce those limits, keep the work in a human-reviewed draft stage.
Here’s the deal
Every assistant needs a named job, an identity and a record of its actions. A shared employee login hides who did what. A separate agent account, where supported, makes its activity easier to trace. Google Cloud’s October 22 announcement specifically describes agent identity and observability as part of its expanded governance tools.
What this actually helps you do
Route routine maintenance requests faster while leaving resident privacy, vendor authorization and property access decisions with people.
One niche use case: apartment maintenance intake
A manager overseeing 120 apartments receives requests by email and web form. The assistant categorizes each request as plumbing, electrical, heating or general maintenance, drafts a ticket and suggests the next available approved vendor. A coordinator checks urgency, confirms the address and authorizes the dispatch. The assistant does not approve spending, unlock doors or promise an arrival time.
Three benefits
What the assistant may do
Do not grant payment changes, entry codes or unrestricted resident-record access for this job.
Step-by-step instructions
- Choose one intake channel. Start with a single web form or monitored inbox, not every communication channel at once.
- List the minimum data. Unit reference, issue description, callback method and reported urgency may be enough. Exclude unrelated records.
- Create a separate agent identity where supported. If your software cannot distinguish the agent, use a restricted integration and keep a separate activity record.
- Set read and write limits. Allow ticket drafts. Block payment updates, vendor approval, door access and tenant profile edits.
- Record every meaningful action. Log the ticket ID, timestamp, action, system, result and approving person. Protect the log itself.
- Set approval gates. A coordinator checks emergencies, dispatches and messages that could create a commitment.
- Test with made-up requests. Try an ordinary leak, a false emergency and a request containing sensitive information. Confirm the assistant stops where required.
- Review weekly. Compare approved tickets with corrections, missing records and unauthorized attempts.
Five fields worth keeping
Tips and tricks
Keep a simple list of approved tools and permissions. Give the assistant a short, written job description. Start with drafts rather than automatic dispatch. Review permission changes whenever a new integration is added.
Common mistakes
Using an employee’s full-access login, storing tenant details in public logs, assuming a prompt is the same as a technical permission, and calling a drafted ticket a completed repair.
Human review checklist
- Does this request belong to the correct unit?
- Was unnecessary personal information excluded?
- Did the assistant stay within its approved access?
- Does the ticket need urgent human attention?
- Has a coordinator approved any vendor dispatch or resident commitment?
- Can we trace the action back to the assistant and reviewer?
How to measure success
Track median time from request to reviewed ticket, the share of tickets requiring corrections, missing action-log entries, unauthorized access attempts and resident follow-up delays. Compare against your current process before claiming time savings.
FAQ
Is an activity log enough to protect tenant data?
No. You also need restricted access, secure storage and clear human approval rules.
Can AI send a maintenance vendor automatically?
Only when the business has explicitly approved that action and the system can enforce its limits. Start with human approval.
What if our software has no agent accounts?
Use the narrowest available integration permissions and keep sensitive actions in a human-owned system. Do not share an administrator login.
Glossary
Agent identity: An account or identifier that distinguishes the assistant from a person. Least privilege: Giving a system only the access its job requires. Activity log: A record of actions and results. Approval gate: A step that waits for an authorized person’s decision.
Sources and further reading
- Google Cloud: Agent Compute Engine and expanded agent governance, October 22, 2026
- Google Cloud: Gemini at Work and agent governance, October 8, 2026
- Microsoft Learn: Design autonomous agent capabilities
Practical closing note
Start with one job and one set of permissions. A maintenance assistant should make a coordinator’s work easier, not inherit the keys to the whole business.
2 Responses