Blog

How to set money limits for AI in a property management company

Zrico team ·
In short

Set AI money limits the way you set staff approval limits: a per-step cap no higher than your most junior approver's, a per-day cap you would be comfortable discovering the next morning, and a scope limited to properties and hours where a person can check the work. Start low, read the log weekly and raise one limit at a time.

Once AI can pay a bill, issue a credit or approve a repair cost, the question is no longer whether it is smart enough. The question is how much it is allowed to spend before a person looks. Property management companies already answer that question for staff every day. The best AI limits borrow from that habit rather than inventing something new.

Step 1: List the money steps

A money step is anything that moves money or changes what someone owes. Write down every one your team handles in a normal month, by area:

  • Maintenance: approving a repair cost, accepting a quote, adding a charge to a tenant for damage.
  • Bookkeeping: paying a vendor bill, recording a refund, moving money between accounts.
  • Collections: waiving or reducing a late fee, setting up a payment plan.
  • Leasing: applying a concession, refunding an application or holding fee.
  • Owner reporting: paying an owner distribution.

Some of these should never run on their own, whatever the cap. Owner distributions and moving money between accounts are good candidates for always waiting for a person.

Step 2: Start from your staff approval matrix

Most companies have a rule like: a maintenance coordinator can approve repairs up to $500, a property manager up to $2,500, and anything larger needs the owner. Use the lowest staff limit in each area as the ceiling for the AI's per-step cap. The AI should never have more authority than your most junior person who does the same job.

Step 3: Set a per-step cap

The per-step cap is the most the AI can spend on any single step without asking. It limits the damage from one bad decision. Pick a number where a single mistake is annoying but not serious. A good test: if the AI got this one step wrong, could you fix it with one phone call?

Step 4: Set a per-day cap

The per-day cap limits the damage from a pattern, such as a bad batch of bills or a vendor who submits the same invoice three times. It should be a number you would be comfortable discovering the next morning. Here is a worked example for one company:

  • Maintenance: $500 per step, $1,500 per day
  • Bookkeeping (bill payments): $750 per step, $3,000 per day
  • Collections (fee waivers): $75 per step, $300 per day
  • Worst-case daily exposure across these areas: $1,500 + $3,000 + $300 = $4,800

Look at the total. If $4,800 of mistakes in one day would be a serious problem, lower the daily caps until the worst case is something you can absorb. Over a five-day work week the same caps allow up to $24,000 on Autopilot, which is why a weekly review matters.

Step 5: Limit where and when

  • Properties: start with properties whose data is clean: correct vendors, current leases, no open disputes. Steps elsewhere wait.
  • Hours: limit Autopilot to business hours so a person is around if something looks wrong. Outside those hours, steps wait for approval.
  • A daily step count: set a maximum number of steps the AI can take on its own per day. After that, everything waits. It catches loops that stay under the money caps.

Step 6: Decide whose authority the AI uses

When the AI acts on its own, it should act as a named person, with that person's permissions, and its work should show under their name in the audit log. Choose someone whose permissions match the job. If the person can't pay bills from a trust account, neither can the AI acting as them. Review that person's permissions when their role changes.

Step 7: Run it in Assist first

Before any money step runs on its own, run the area in Assist for two to four weeks. The AI drafts each step, a person approves it, and you collect evidence. Count how many drafts you approved without changes, how many you edited, and how many you rejected. If you are rejecting more than a handful, the problem is usually the data, not the limit.

Step 8: Review the log every week

Spend 20 minutes a week on the AI's activity log for each area on Autopilot. Look for three things:

  1. Every step that didn't work. Was the cause bad data, a missing setting or a real edge case?
  2. Every step that asked a person. Were the approvals reasonable, or is the cap so low that people approve without reading?
  3. The largest steps that ran on their own. Would you have approved each one?

Step 9: Raise one limit at a time

Raise a cap only when the review shows that the steps waiting just above it are ones you approve every time. Change one number, in one area, and wait two weeks before changing another. If you raise the per-step cap, decide whether the per-day cap should move too, or whether you want fewer, larger steps per day.

Common mistakes

  • Setting the per-day cap at 30 times the per-step cap. That is a monthly budget, not a daily safety net.
  • Turning on every property at once, including the one with a vendor dispute.
  • Letting the AI act as an admin. It inherits every permission the admin has.
  • Never lowering a limit. If a review finds a bad pattern, lower the cap first and investigate second.
  • Forgetting the stop switch. Everyone who manages money should know where it is.

A one-page policy

Write the limits down in one page that an owner or partner could read: the areas on Autopilot, each per-step and per-day cap, the properties and hours, who the AI acts as, who reviews the log and when, and who can stop it. If you manage money for owners, this page is also a good thing to share with them. If any limit is set by an owner agreement or by your own trust account rules, follow those first. This is not legal advice.

How Zrico helps

In Zrico each area is set to Off, Assist or Autopilot. Money caps work per step and per day and can be limited to chosen properties and company hours, with a maximum number of steps per day. Autopilot acts as a person you name, with their permissions. Legal steps are off unless you allow them, and some steps always wait for a person. Shadow mode lets an agent draft everything while a person approves each change, and Stop everything is one switch. See the controls at /ai and /ai/trust, and the bookkeeping agent at /ai/agents/bookkeeping.

See Zrico run your portfolio.

Bring a property. We will show you, on your own data, what Zrico would have handled last month.