On this page
At a company of a few hundred people, the work AI could take is in plain sight: booking emails keyed into the transport system by hand, repair requests retyped as work orders, payments matched to open invoices one at a time.
Each eats hours of someone's week, and one burst pipe logged as routine costs a flooded unit and a vendor called too late.
Take a property manager of about 250 people whose two maintenance coordinators turn some 140 repair requests a week into work orders, the task a vendor is sent to do. At under ten minutes a request, corrections included (our assumption), that takes about 22 hours of their week:
| Line | Today | With the agents |
|---|---|---|
| Coordinator hours a week | 22.4 | 11.2 |
| Coordinators' cost a year | $43,792 | $21,896 |
| Running the agents a year | $0 | $4,200 |
| Total a year | $43,792 | $26,096 |
Dispatchers, the Bureau of Labor Statistics' category for scheduling crews and service calls, earned a mean $54,740 in May 2025, and benefits are 30% of what private employers spend by the same bureau's count, so each coordinator costs about $78,200.
Put your own desk in. The calculator opens on that example, with AI agents taking half the work: reading, writing up and ranking the routine requests, while a coordinator still checks and releases every draft.
Repair Request Work Orders ROI Calculator
What Turning Repair Requests Into Work Orders Costs You, And What An Agent Saves
Saved Per Year
$17,696
582 hours handed back, after $4,200 a year to run it
- Hours Handed Back
- 582 hours a year
- Agent Running Cost (Estimate)
- $350 a month
- Payback Per $10,000 Of Build
- 6.8 months
On these numbers, the agent saves $17,696 a year after its own running cost, and every $10,000 of the build is earned back in 6.8 months.
On the example, the desk saves $17,696 a year of coordinators' time after running costs and hands back 582 hours; every $10,000 of the build is earned back in about 6.8 months. Running the agents, about $350 a month, is an estimate: a few cents of AI model use per request, plus hosting and upkeep.
Day 90 Ends in a Decision, Not a Deck
Spend your first 90 days on one recurring job, priced: what it costs your team today, what agents can take off it, and what checking their work costs. Sign the rule that will judge it first, and on day 90 let the rule decide: keep, fix, or stop.
A company this size learns what AI is worth in its own work one job at a time.
At your size, the strategy worth having is a list of the jobs AI does, who owns each one, and what each has proved, and this quarter adds its first row.
Sign the Rule That Decides Day 90 First
Write the rule before any result exists, because a rule written afterward tends to fit whatever the results turned out to be. The owner of the company and the owner of the job sign it, with the date it applies.
It's read from the top, and serious errors come first: one emergency logged as routine can cost more than a week of saved minutes, so no number of hours turns a stop into a keep.
| Outcome | Serious errors | Net hours a week | Anything else |
|---|---|---|---|
| Stop | Two or more reach a resident or vendor in the live month | Zero or less | |
| Fix | One reaches a resident or vendor, or more than one draft in 50 has one caught | Under a fifth of today's | Checking takes 3 minutes or more; exceptions pass a quarter; payback over a year |
| Keep | None reaches a resident or vendor; at most one draft in 50 has one caught | A fifth of today's or more | None of the above |
Stop and fix each apply when any one of their cells does; keep needs every cell. The thresholds are the example's: set yours by what an error costs your residents and what an hour costs you.
Those hours are capacity, not cash, until they go somewhere: more requests without a new hire, faster answers to residents, or the coordinators' time on the next job. On the example that is $17,696 a year, and day 90 is when you say where it went.
The Agents Get One Permission at a Time
A serious error that reaches a resident or vendor means the reviewer missed it, so that request type goes back to the coordinators the same day, until its cause is found and a fresh test on that type passes. A second is the stop.
If requests start going around the agents, find the request type behind them. If one type's drafts need so much fixing that typing is faster, hand it back to the team and keep counting it as an exception.
Before they touch the inbox, we agree in writing what the agents may read and do, and where their data is processed. They work on their own account, never a person's login, with every action logged. For two weeks nothing they make is used; only then do they get their one write:
| Stage | They read | They write | They never |
|---|---|---|---|
| Weeks 5 and 6 | The maintenance inbox and tenant portal; the unit and vendor lists | A sheet outside the system | Write to the system; send a message |
| Weeks 7 and 8 | The same | A draft work order in one queue, marked as theirs | Dispatch, edit or close a work order; message a resident or vendor; approve a cost |
What Our Agents Do With Your Repair Requests
At a maintenance desk the work starts in the shared inbox and the tenant portal, and so do the agents we build for it.
On autopilot, they read each request as it lands, look the unit and the resident up, rank its urgency by the rule you signed, and leave a draft work order in one queue of the maintenance system your coordinators already use, for one of them to release.
With us, the quarter opens on last month's requests, read beside the coordinators who handled them, until every exception and vendor habit is on paper. Last month's finished work orders are the answer key the agents are tested on, and a first working version is filling drafts within two weeks of access.
We train the coordinators who will release their drafts. Every week the agents run, we count from your inbox and maintenance log the coordinator hours they return, net of checking, and the wrong units, vendors and urgencies they stop before a work order goes out.
Your signed rule reads those numbers, and our price is fixed for a defined result.
What the First Two Weeks With Qvery Look Like
It starts from the repair requests your desk took in last month and the work orders made from them. The first two weeks go to reading them beside your coordinators and running the agents on that month, and on day 90 the rule you signed decides whether the desk keeps them, fixes them or stops.
Frequently asked questions
Which job should go first?
One that happens every working day, has a right answer someone could check, draws on your own systems, and has mistakes that can be caught before they reach a customer. Among up to three that pass, pick the one with the most hours behind it.
Who should own AI adoption in our company?
For each job, the person who runs it. The company's owner approves the rule in writing and can halt the job, and IT decides access.
How long should a first AI pilot run?
Two weeks to try a first version on finished cases, then four weeks connecting it beside the team, then four live weeks to learn whether it pays. A fix round adds two weeks and moves day 90 with it.
What if nobody on our team is technical?
The first version runs on exported cases with no live connection, and the people who run the job score it. Connecting it to live systems needs a builder, inside the company or outside.
What does Qvery build in the first 90 days?
Custom AI agents for the one job you pick, working on autopilot inside the systems your team already uses. A first working version runs within two weeks of access, we train the people who release their work, and we count the hours saved and errors stopped each week. The price is fixed for a defined result.
What should a company do in its first 90 days of AI?
Prove one recurring job can pay for itself. Price it before anything is built, sign the rule that will judge it, test a first version on last month's finished work, and let the rule decide on day 90: keep, fix or stop.
What should an AI strategy look like for a company of a few hundred people?
A list of the jobs AI does, who owns each one, and what each has proved. A company this size learns what AI is worth one job at a time, and the first quarter adds the first row.
Why write the success rule before an AI pilot starts?
Because a rule written afterward tends to fit whatever the results turned out to be. The company's owner and the job's owner sign it, with the date it applies.
What decides whether an AI pilot is kept, fixed or stopped?
The signed rule decides, with serious errors checked first. In the example, two or more serious errors reaching a resident or vendor in the live month means stop. So do zero or fewer net hours saved weekly. Any stop or fix condition applies; keep requires every condition.
Why do serious errors outrank hours saved?
Because one emergency logged as routine, like a burst pipe, can cost more than a week of saved minutes: a flooded unit and a vendor called too late. No number of hours turns a stop into a keep.
What happens after one serious AI error?
If a serious error reaches a resident or vendor, coordinators take that request type back the same day. They retain it until its cause is found and a fresh test passes for that type. In the example, a second such error in the live month means stop.
How much time does turning repair requests into work orders take?
In this post's example, two coordinators turn some 140 repair requests a week into work orders at under ten minutes each: about 22 hours a week, $43,792 a year at about $78,200 per coordinator with benefits.
How much can AI agents save on repair work orders?
With agents taking half the work, the example saves $17,696 a year of coordinators' time after running costs and hands back 582 hours. Every $10,000 of build is earned back in about 6.8 months.
What does it cost to run repair request agents?
About $350 a month in this post's estimate, $4,200 a year: a few cents of AI model use per request, plus hosting and upkeep.
What access do AI agents get in a pilot?
One permission at a time. In weeks 5 and 6 they read the inbox, tenant portal and unit and vendor lists and write only to a sheet outside the system. In weeks 7 and 8 they leave draft work orders in one queue, marked as theirs.
What do AI agents never do on a maintenance desk?
Dispatch, edit or close a work order, message a resident or vendor, or approve a cost. A coordinator releases every draft, and the agents work on their own account, never a person's login.
Are the hours an AI pilot saves real money?
Not until they go somewhere: more requests without a new hire, faster answers to residents, or the coordinators' time on the next job. Day 90 is when you say where they went.
What should you do if staff start working around the AI agents?
Find the request type behind it. If one type's drafts need so much fixing that typing is faster, hand it back to the team and keep counting it as an exception.
