The AI agent reliability report for property management
AI can read a shared inbox, sort complaints and plan technicians. The harder question for any owner is simpler: can you rely on it? The doubt is real and measured. In AppFolio's 2026 industry survey, 78% of property managers said they cannot yet rely on the AI features in their existing software. Yet the firms that did adopt AI expect their portfolios to grow 31% this year, nearly triple the growth expected by those that didn't.
WHAT STOPPED HAPPENING
Three kinds of manual work that quietly ate the payroll
Most vendors will tell you their AI is accurate. We would rather show you the count: how many cases, over which period, checked by whom, and where it missed. Every figure below has a sample size and a confidence interval in the appendix. Where a number is a projection, it says so.
Nobody reads the inbox anymore
In our property-management environment, we audited every single message the system handled over a four-day stretch, read back one by one by a person. The result: 95% needed no human answer at all. At that summer rate, a year of this inbox is 13,350 messages read, and only 623 that need a person. At one minute per message, that is roughly 210 hours a year of reading that no longer happens.
The system read, sorted and routed the rest on its own: invoices, supplier threads, noise. And it judged correctly. In a separate window of 876 messages, it correctly ignored about nine out of ten junk messages while catching the one genuine emergency and flagging it as urgent. Because it is software, the inbox is read at 3 a.m. the same way it is read at 3 p.m.: no shift, no backlog on Monday morning.
95.3%
of messages needed no human answer at all
13,350 → 623
messages a year, at the measured rate
Nobody plans the technicians anymore
In our field-service environment, the planning engine schedules the technicians: it knows every calendar, calculates travel times, and writes down, on every job, why it chose who it chose. One moment from the planning log: three technicians were available, 186, 56 and 34 minutes away. The engine sent the one 34 minutes away, 22 minutes closer than the next-best choice and 2½ hours closer than the worst. Multiply that choice by every job, every week: that is fuel, hours and mood you are not paying for anymore.
For a tenant who couldn't lock her own front door, the engine picked the technician an hour closer and four days sooner than the alternative, then bundled a later job at the same address, so one trip resolved three tickets. Nobody touched a calendar.
2½ hours
less driving than the worst available choice, on one job
1 trip
resolved three tickets at the same address
Nobody types the same fact four times anymore
When a report comes in, the system sorts it into the right category (heating, leak, appliance), and that sorting decides who pays and who gets called. We tested this the hard way: the same cases, fed to the system again and again, across every category we handle. It gave the same, correct answer every single time: 444 runs across all 15 categories, zero deviations. A person having a bad Tuesday does not do that.
From that one decision, the ticket, the tenant reply in the tenant's own language, the owner update and the work order follow automatically.
444
runs across all 15 categories, zero deviations
A support line doesn't need a night shift
In our customer-support environment, an entire WhatsApp support line runs on our AI employee: it answers day and night, sorts every complaint, opens the ticket, and when a customer deserves compensation, it issues the discount code itself. Compensation with a cost attached, handled by software, within limits it cannot cross.
Over one two-week stretch it handled about 215 conversations and tasks a day. Of roughly 3,000 runs, 3 stopped or stalled (99.9% run stability), and every one of those was caught by our own monitoring, not by a customer. In twelve days, 36 customers got exactly one code each, nobody got two, and in 120 deliberately tricky cases, not one genuine customer was wrongly refused.
When a better AI model became available, we switched, and cut the running cost of this system by a factor of twenty, from about 11 cents per conversation to well under one cent, measured over 4,939 conversations, with quality re-tested and held. Your automation gets cheaper as it matures, not more expensive.
99.90%
run stability of the support line
€0.11 → <€0.01
running cost per conversation after the model switch
HOW WE MEASURE
Nothing goes live that hasn't been right 19 times out of 20
- 01
Nothing ships until it passes 19 out of 20
Before any change goes live, we run it against the same cases twenty times over. If it doesn't give the right answer at least 19 times out of 20, it does not ship. Full stop.
- 02
Hundreds of automatic checks, on every change
When we adjust anything, the system re-tests itself against its own history. After one recent change we re-ran 588 past cases: not a single outcome changed that shouldn't have.
- 03
A human audits the results
Periodically, a person separate from the build reads the outcomes back one by one, like the four-day inbox audit above, to confirm the system did what it should have.
Across the three control layers we measured, the weakest scored 99.38% and the pooled result over 1,638 runs was 99.69%: an error rate under 1% either way, with the full accounting in the table below for anyone who wants to check our math.
THE LIMITS
You decide what it may send.
Not us, and not the AI.
Whether the AI may send messages or create work orders on its own is a choice you make, per process, never a default. Out of the box, a person reviews everything before it goes out. Where you deliberately choose autonomy, the sending is done by fixed, predictable software following an approved plan, with an automatic content check, never by the AI free-styling.
- 01
Unknown suppliers
An unknown supplier never gets a work order; you get a notification instead.
- 02
Above the mandate
Costs above the mandate you set trigger an approval request first.
- 03
Doubt
When the system isn't sure, it stops and hands the case to a person with a summary; it does not guess.
The work your team shouldn't be doing is already gone at ours.
FROM DAY ONE
What you can hold us to
If you work with Triad, these are not aspirations. They are the rules the numbers above were produced under, and they apply to your implementation on day one. This report is the first in a series; every next edition adds new implementations and fresh numbers.
- 01
Nothing goes live below 19 out of 20
Every process is run against your own cases, twenty times over, before it touches your inbox, your tenants or your suppliers.
- 02
Every number comes with its count
You get the same accounting you just read: sample size, window, misses. Monthly, for your own operation.
- 03
A person audits the output
Someone separate from the build reads real outcomes back, on a schedule, and the misses are reported to you, not buried.
- 04
You decide what it may send
Autonomy is switched on per process, by you, never by default. Unknown suppliers, costs above your mandate and low-confidence cases always go to a person.
APPENDIX A
How we counted
Every figure in this report, with its sample size, measurement window and 95% confidence interval (Wilson method). A figure without an interval is an estimate posing as a fact. Sample sizes below are the actual measured counts; the annual volumes in the body are these measured rates projected to a year.
- Inbox censusn=15095.3%95% interval 90.7 to 97.7%
- Intake, decision layern=32099.38%95% interval 97.75 to 99.83%
- Classification consistencyn=444100%95% interval 99.14 to 100%
- Re-run after a changen=588100%95% interval 99.35 to 100%
- Three control layers pooledn=1,63899.69%95% interval 99.29 to 99.87%
- Support line, run stabilityn=3,01599.90%95% interval 99.71 to 99.97%
| Figure, as used in the body | n | Result | 95% CI | How measured |
|---|---|---|---|---|
| Inbox census: share needing no human answer | 150 | 95.3% | 90.7 – 97.7% | every message over 4 days, read back by a person separate from the build; 143 of 150 needed no action, 7 did |
| Inbox census: routed correctly | 150 | 99.33% | 96.32 – 99.88% | same census; 1 genuine miss, rule fixed same week |
| Junk filtering | 876 | ≈ 90% | 87.79 – 91.77% | separate window; junk messages correctly ignored, 1 genuine emergency in the window, flagged as urgent |
| Classification consistency | 444 | 100% | 99.14 – 100% | the same cases repeated across all 15 categories, tested before rollout; 0 deviations |
| Re-run after a change (regression) | 588 | 100% | 99.35 – 100% | past cases re-run after one change; 0 outcomes changed that should not have |
| Reporter recognition (intake) | 59 | 88.14% | 77.48 – 94.13% | gross; 3 of 7 misses were external parties correctly left unmatched (net 93%) |
| Intake: agent decision layer | 320 | 99.38% | 97.75 – 99.83% | measured runs; both errors caught by automatic retry |
| Support line: run stability | ≈ 3,015 | 99.90% | 99.71 – 99.97% | consecutive runs over 2 weeks, 3 incidents (runs stopped or stalled), all self-detected; measures completion, not answer quality |
| Support line: end-to-end process | 697 | 99.57% | 98.74 – 99.85% | measured runs, 2-week window, 3 errors |
| Support line: customer lookup step | 621 | 100.00% | 99.39 – 100% | measured runs, 2-week window, 0 errors |
| Support line: compensation guardrails | 120 | 0 wrongly refused | n/a | two independent 60-run adversarial rounds; 36 codes issued in 12 days, 0 duplicates |
| Support line: cost per conversation | 4,939 | €0.11 → under €0.01 | n/a | running cost before and after the model switch, quality re-tested on the same cases |
| Pooled control layers | 1,638 | 99.69% | 99.29 – 99.87% | end-to-end process + agent decision layer + lookup step; worst-case error 0.71%, under the 1% stated in the body |
Where the outside numbers come from
- 01
AppFolio, Property Manager Benchmark Survey 2026. Figures verified against the live page, 26-08-2026. The 78% concerns AI features in legacy property-management software. appfolio.com
- 02
Digital Applied, Customer Service AI Agent Statistics 2026, an aggregation of research from Zendesk, Salesforce, Gartner, Forrester, Intercom and McKinsey. Verified 26-08-2026. digitalapplied.com
- 03
Towards a Science of AI Agent Reliability, arXiv:2602.16666, ICML 2026: the research position this report's repeated-measurement approach follows. Verified 26-08-2026. arxiv.org