Using AI in property management safely: approvals, limits and audit logs
Use AI safely by sorting work by how hard a mistake is to undo, letting the AI act alone only on low-risk work, and sending everything else to an approval queue with named reviewers and response times. Add limits on scope, hours, contacts and volume, review the AI's log every week, and practice stopping it before you need to.
Most guidance on AI in property management stops at turn it on carefully. That isn't a plan. A plan says which work the AI may do alone, who approves the rest and how fast, what limits apply besides money, how someone checks the record every week, and what happens when something goes wrong at 9 p.m. on a Friday. This guide covers each of those. Our separate post covers money limits in detail, so here they appear only where they fit the bigger picture.
Step 1: Sort the work by what a mistake costs
List the tasks you are considering for AI, then put each in one of four groups based on what happens if the AI gets it wrong:
- Internal and reversible: drafting a reply, tagging a work order priority, summarizing a lease, matching a bank line. A mistake costs a few minutes to fix and nobody outside your office sees it.
- Contacts someone: a message to a resident, an applicant, an owner or a vendor. A mistake can be corrected, but the person has already read it.
- Money: anything that charges, pays, refunds, waives or moves money.
- Legal weight: notices, lease terms, renewal offers, screening decisions, collections referrals. A mistake may not be undoable at all.
The groups decide the default. Internal work can usually run on its own once you've checked it for a few weeks. Contact work can run alone inside rules, with sampling. Money runs alone only under caps. Legal work waits for a person, every time, unless you have made a deliberate, narrow exception with counsel.
Step 2: Choose a mode per area, not for the whole company
Each area of work should have its own setting: off, drafts that a person approves, or acting alone within limits. Leasing replies might be ready to run alone months before collections messages are. Setting everything at once means the riskiest area sets the pace for all of them, or worse, the easiest area sets the risk level for the hardest.
A sensible starting point for many companies:
- Maintenance triage and resident updates: drafts first, then acting alone after a few weeks of review
- Leasing replies and showing bookings: drafts first, then acting alone inside showing hours
- Bookkeeping matching: drafts first, then acting alone under low caps
- Collections messages: drafts only for the first months
- Anything legal: off
Step 3: Design the approval queue like a job
An approval queue nobody owns becomes a backlog, and a backlog teaches staff to approve without reading. Decide:
- Who approves each kind of step. Approvals should match the authority the person already has: a leasing agent approves leasing replies, a property manager approves maintenance costs within their limit, an accountant approves bookkeeping entries.
- How fast. Set a response target by type, for example within one business hour for resident replies and by end of day for bookkeeping items. A queue item that waits too long either goes stale or reaches the resident too late to matter.
- Coverage. Who approves after hours, on weekends and when the usual approver is out?
- What the reviewer checks. A short list per type: facts match the record, tone is right, the amount is right, the recipient is right, nothing promised that the office can't deliver.
- What rejecting looks like. A rejection should carry a reason, so patterns show up in the log and the setup can be fixed.
If approvers routinely approve every item unchanged for a month, that type of step may be ready to run alone. If they change a large share of items, it isn't, and the setup needs work before the mode changes.
Step 4: Set limits beyond money
Money caps get most of the attention. These limits matter as much:
- Scope: which properties the AI may act on. Start with a few properties where staff know the residents and can spot problems.
- Hours: when it may act alone. Many offices allow independent work only during company hours at first, so a person is around to catch an error.
- Contact rules: quiet hours for messages, how often the same person may be contacted, which channels may be used, and respect for every opt-out.
- Volume: a maximum number of steps per day. A daily ceiling turns a runaway mistake into a small one.
- Bulk messages: messages to many people at once should always wait for a person, because one error is copied to every recipient.
- Steps that always wait: some steps should never run alone regardless of settings, such as assigning a vendor, which commits a cost, or referring an account to a collections agency.
Step 5: Decide whose authority the AI uses
When the AI acts, it should act as a named person, with that person's permissions, so it can never do something that person couldn't. Choose someone senior enough to cover the work and accountable for reviewing it. Avoid an admin account with every permission, because then the AI's limits are only the limits you remembered to set.
Step 6: Know what a good log entry contains
A log you can't read is not a control. Each AI action should record:
- What it did, in plain words
- When, and on which property, unit or record
- As whom, meaning whose permissions it used
- What it based the action on: the message, the record, the rule
- The outcome: done, sent to a person, or failed
- If a person approved it, who and when
Keep the AI's log separate from the staff audit log, or clearly labeled within it, so you can review AI work on its own.
Step 7: Review the log every week
Put 30 minutes on someone's calendar every week. A simple routine:
- Read a sample of about 20 actions the AI took on its own, across areas. Would you have done the same thing?
- Read every action that failed, and look for a pattern: the same property, the same kind of request, the same missing data.
- Read every step that was sent to a person, and check whether the reason was right. Too many hand-offs means limits are too tight or data is missing; too few means they may be too loose.
- Check resident and owner complaints against the log for the same week.
- Write down one change, if any, and make it. Change one setting at a time so you can see its effect.
Step 8: Plan for stopping it
Write a short plan before anything goes wrong:
- Who can stop the AI, for one area or everywhere at once. More than one person should be able to.
- When to stop it: a wrong message to many residents, an unexpected payment, a complaint that suggests unfair treatment, or anything nobody can explain.
- What to do next: stop first, then read the log for the affected period, contact anyone affected, fix the setting or data that caused it, and record what happened.
- Who decides to turn it back on, and with what change.
Run the plan once as a drill. Stopping it should take seconds, and the rest of the platform should keep working for staff.
Step 9: Protect sensitive data
- Keep screening reports, IDs, tax forms and bank details out of AI steps unless there is a specific need and the data is masked.
- Make sure messages to applicants and residents are checked against fair housing rules and your property facts before they go out. Fair housing rules apply to AI-written messages exactly as they apply to staff. This is not legal advice; follow your counsel.
- Ask your vendor in writing whether your data is used to train AI models, and by whom.
Step 10: Roll out over 90 days
- Days 1 to 30: one or two areas in draft mode, on a few properties. Approvers review everything.
- Days 31 to 60: low-risk internal work runs alone during office hours. Contact work stays in drafts. Weekly log reviews running.
- Days 61 to 90: contact work runs alone inside contact rules on the first properties. Expand scope one property group at a time.
- After 90 days: review every area's settings, approval times and log findings, and decide the next step for each area separately.
How Zrico helps
In Zrico, each area is set to Off, Assist or Autopilot: Leasing, Collections, Maintenance, Bookkeeping, Owner reporting, Renewals, and Resident messages and calls. Money caps work per step and per day and can be limited to chosen properties, legal steps are off unless allowed, and some steps always wait for a person, such as assigning a vendor or sending an account to a collections agency. Waiting steps go to Zrico approvals, where approving runs a step once, as you, with your permissions. Each step carries a risk badge, Low risk, Contacts someone or Money or legal, and everything Zrico did is kept in its own log with the outcome: Did it, Asked a person or Didn't work. Stop everything is one switch. Read more at /ai/trust and /answers/how-to-stop-zrico-ai.