Blog

Why Zrico's AI waits for your approval on money and legal steps

Zrico team ·
In short

Zrico AI does routine work on its own, but anything with money over your limit or legal weight waits for a person, because those are the mistakes that are hard to undo. You decide where that line sits, area by area, and you can move it as you build trust.

Most of the work in a property management office is not dangerous. Answering a lead's question about parking, drafting a reply to a maintenance request, matching a deposit to a bank line, summarizing a month for an owner: if any of that is slightly wrong, someone notices and fixes it in a minute. A small share of the work is different. Paying a bill, refunding a deposit, waiving a fee, opening a collections case or sending a formal notice either moves money or carries legal weight. When those go wrong, the fix can take weeks, and sometimes there is no fix.

Zrico is built around that difference. The AI does the routine work inside the platform, and the steps that are hard to reverse wait for a person. This post explains why we drew the line there and how it works in practice.

What makes a step hard to undo?

We sort AI work by what happens if it is wrong, not by how clever the work is. Three questions decide it:

  • Does it move money or change what someone owes? A payment that went out has to be recovered from the person who got it. A waived fee is awkward to put back.
  • Does it contact someone outside your company? Once a resident, owner or vendor has read a message, you can't unsend it.
  • Does it carry legal weight? A notice, a case opened with a collections agency or a lease change can start a clock or create an obligation.

Inside Zrico each step the AI proposes carries one of three risk badges that follow this logic: Low risk, Contacts someone, and Money or legal. The badge is the first thing you see on an approval card, so a busy manager can tell in a second which cards need a careful read.

How does Zrico decide what runs on its own?

You decide, per area. The areas are Leasing, Collections, Maintenance, Bookkeeping, Owner reporting, Renewals, and Resident messages and calls. Each one is set to one of three levels:

  • Off: the AI does nothing in that area.
  • Assist: the AI drafts the work and a person approves each step.
  • Autopilot: the AI acts on its own within the limits you set, and anything over a limit waits for approval.

The AI as a whole is off until an admin turns it on, and Autopilot levels are off by default too. Most companies will start with a few areas on Assist and move one area to Autopilot once they have read enough of its drafts to trust them.

What limits apply on Autopilot?

Autopilot is not a blank check. Each area has its own limits, and a step that goes past any of them stops and waits:

  • Money caps: a most per step and a most per day. If money steps aren't allowed at all, every money step waits.
  • Properties: you can limit an area to chosen properties. Steps for other properties, or where Zrico can't tell which property is involved, wait.
  • Hours: outside the company hours you set, steps wait for approval, so the AI isn't acting alone when nobody is in the office.
  • A daily count: after a set number of steps on its own in a day, everything waits until tomorrow. It is a safety net against a loop or a bad batch of data.
  • Legal steps: off unless you allow them. With them off, something like opening a collections case with a late notice for staff to serve waits for a person.

Some steps always wait for a person no matter how the area is set. Assigning a vendor and sending an account to a collections agency are two examples. These are decisions where judgment about a relationship matters as much as the rules.

What does approving actually do?

Waiting steps go to Zrico approvals, which has three tabs: Waiting, Decided and Activity log. Each card says what Zrico wants to do, why, and what it will cost if money is involved. The card text is blunt on purpose: nothing happens until you approve.

When you approve, Zrico runs that step once, as you, with your permissions. That detail matters. If you can't pay bills from a certain bank account, approving a card doesn't let the AI do it either. The approval doesn't create a standing rule, so the next similar step asks again unless you change the area's limits.

On Autopilot, the AI acts as a person you name, with that person's permissions, and its work shows under their name in the audit log. You choose that person deliberately, the same way you would decide how much authority to give a new staff member.

How do you see what the AI did?

There are two records. The staff audit log is append-only and shows what people did. A separate log shows everything Zrico did, with one of three outcomes: Did it, Asked a person, or Didn't work. Statuses on each item read Done on its own or Waiting for approval.

The Didn't work outcome is important. When an automatic step fails, Zrico says so plainly and nothing is changed by it. You can do the step from the record's page or ask Zrico to try again. Quiet failures are worse than loud ones, so we made them loud.

Why not let the AI do everything?

Because the cost of errors is lopsided. Suppose an AI handles 400 bill payments a month and gets 399 right. The one it gets wrong might be a duplicate payment to a vendor who is slow to refund, or a payment from the wrong owner's trust money. The hours saved on 399 bills don't help much while you explain the 400th to an owner. A per-step cap of a few hundred dollars and a daily cap you are comfortable with turn that tail risk into something small and visible.

There is also a trust problem. Staff and owners need to believe the books. If they suspect the AI might do something large without anyone looking, they start checking everything by hand, and the time savings disappear. Clear limits make it reasonable to stop double-checking the small things.

Why not make the AI only suggest?

Because suggestion-only tools tend to become one more inbox. If every lead reply, every work order note and every bank match needs a click, people approve in bulk without reading, which is worse than a clear limit. Autopilot for low-risk work, with approvals reserved for the steps that deserve attention, keeps the approval queue short enough that people actually read it.

How do you stop it if something looks wrong?

  • Each area has its own kill switch, so you can stop Collections while Maintenance keeps going.
  • Stop everything is one switch that stops every area at once. Nothing runs on its own again until an admin turns each area back on.
  • Shadow mode lets an agent draft everything as usual while a person approves every message and change, which is a good way to try an agent on real work before it acts alone.

A sensible way to start

  1. Turn on one area on Assist, for example Maintenance, and read its drafts for two weeks.
  2. Move that area to Autopilot with a low per-step cap, a modest daily cap, business hours only and a few properties.
  3. Check the activity log weekly. Look at every Didn't work and every Asked a person and ask whether the limit was right.
  4. Raise one limit at a time. Add the next area only when the first one is boring.

How Zrico helps

The levels, caps and switches described here are on the Zrico AI page at /ai, and the full list of controls, logs and security measures is at /ai/trust. To see how the bookkeeping agent drafts bills and matches bank lines while payments wait for you, read /ai/agents/bookkeeping. For a walkthrough on your own portfolio, book one at /demo.

See Zrico run your portfolio.

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