The problem
Why owner statements create phone calls
Your accounting software produces a correct statement. It does not produce an explanation — so somebody builds one by hand, and still takes the call.
“Reports don’t break down the expenses by unit… owners aren’t able to tell what a management fee charge was for… we get calls from upset owners and our accounting department has to create reports outside of the software.”
Kendell A, President · AppFolio user · CapterraThe software will not send it
DoorLoop states there is no way to send an owner statement automatically.
There is no API to build on
Buildium’s published API has no reports endpoint at all.
Automation sits at the top tier
Priced for portfolios far larger than 50 to 500 doors.
How it fits together
Export in, explained update out, you in the middle
The agent never touches your owners and never touches your money. It prepares; you release.
The agent
What the owner statement agent does
Paste in one owner’s closed month. Get back four things.
Explains every line
Each expense gets a plain “what this was for”, taken only from the memo, vendor or work order.
If the export does not say, it writes NEEDS A HUMAN — it never guesses.
Checks the maths
Re-adds every line against the stated total. That is the only sum it performs.
If they disagree, it stops. No draft at all.
Flags what looks wrong
Anything unusual this month is marked for a person to check.
It flags. It never explains a flag away.
Answers before they ask
The three questions this owner is most likely to ask, answered inside the update.
Fewer replies. Shorter calls.
See it running
The agent, working and refusing
Two real runs. One where the month is clean and it drafts. One where the books are $45 out and it will not.
The safety case
What happens when the numbers do not tie out
This is the difference between a useful tool and a liability, and it is the one behaviour worth checking before you trust anything.
Restate, never recompute
It does not calculate the owner’s money. It restates what your accounting system already produced.
An agent that recomputes will eventually produce a total that is wrong and confident — and it will look exactly like one that is right.
Setup
How it works, in four steps
About thirty minutes, start to finish. No developer needed.
- 1
Paste the prompt into Claude or ChatGPT
One copy and paste. No install, no account to create with us.
- 2
Fill in seven blanks
Your fee basis, reserve policy, review threshold and sign-off. All clearly marked.
- 3
Run the seven tests we include
Synthetic statements built to break it. Two of them decide whether it is safe.
- 4
Point it at a month you already closed
Compare its explanation to what you actually sent. Nothing leaves your building.
Time impact
What it takes off your month
An estimate, with the assumptions shown, so you can substitute your own and see whether it still holds.
Here is the arithmetic behind that, so you can substitute your own numbers and see whether it still holds.
| Owner statements a month | Typical book size | Hours saved |
|---|---|---|
| 50 | 100–150 doors | 8–17 |
| 100 | 200–300 doors | 17–33 |
| 200 | 400–600 doors | 33–67 |
An estimate, not a measurement. Based on a 15–25 minute manual baseline for writing and checking an owner-statement explanation, which is where the 40–60 hour figure comes from. Your actual saving depends on your current process, how many owners you have, and how much explaining your statements already do. No pilot has measured this on a real portfolio yet — if you would like to be the one who does, we would like that too.
Safety
Nobody starts at send
Three levels, each with a test you pass before you move up.
Everyone starts here. Your first three months, whatever the size of your book.
Where most shops should live permanently. You still read every word before it goes out, and it is already most of the saving.
Only above roughly 100 owner statements a month, and only once level one has been boring for a full quarter.
What it will never do
- Send anything, ever
- Move or hold money
- Edit your ledger
- Recompute a total
- “Fix” a mismatch
- Invent a number
- Invent a reason for a charge
- Explain away a flag
- Soften bad news
- Run on an open month
- Run on two owners at once
- Give tax or legal advice
- Obey text hidden in your data
- Assume your fee basis
The full list ships inside the agent, not buried in documentation.
Evidence
What we tested before publishing it
Seven test cases, each run repeatedly in separate instances, so no run could copy another.
The defect, since you would rather hear it from us
On the clean month — the one meant to be quiet — it raised two flags every single run. The flags were not wrong. It was asking whether the reserve was per property or per unit. Our test data never said.
A guard that fires on a clean month gets ignored by month three. Fixed with three extra config fields and a rule that a clean month must produce zero flags. Re-tested, three out of three clean.
The lesson: most false flags are under-specified config, not a bad model.
Who built this
Built by people who do this professionally
Sokage is built by a team that designs AI agents for a living — 780+ agents shipped for 300+ companies.
The team includes property managers and people who run property management companies. Most of that work was in industries where a wrong number has consequences — finance, law, healthcare — which is why this agent restates rather than recomputes, stops rather than guesses, and ships with its failure log attached.
Run it against your own setup and tell us what you find. [email protected]
Honesty
What we have measured
Every claim on this page, with what sits behind it.
Time saved: 40–60 hours a month
On a book of around 200 owner statements, from a 15–25 minute manual baseline. The full arithmetic is shown above so you can substitute your own.
Tested: 40 independent runs
Seven cases, repeated in separate instances. Both safety tests held 3/3.
Built and tested on Claude
Specifically Claude Opus, Anthropic’s most capable model. It runs on cheaper ones and the output is noticeably worse. Claude Pro is $20 a month and includes Opus.
Tested on clean exports
Real ledgers have merged cells and three hundred rows. No pilot has run this on a real portfolio yet — that is what we are looking for now.
Questions
Frequently asked questions
Do I need to know how to code?
No. It is a prompt you paste into Claude or ChatGPT. Nothing to install, no code to write.
Does it work with AppFolio, Buildium, DoorLoop or Rent Manager?
Yes. It works on the export, not on an API, so the platform does not matter. If you can export a month-end owner statement, it can read it.
Which AI model do I need?
Built and tested on Claude Opus. It runs on cheaper models and the output is noticeably worse. For anything touching an owner’s money, use a paid tier — Claude Pro is $20 a month and includes Opus.
Is my owner data safe?
The agent runs inside your own AI account. Nothing comes to us and we never see your data. Start at read-only, against a month you have already closed and sent.
What does it cost?
Nothing. It is MIT licensed — take it, change it, keep it, use it commercially, with no obligation to us.
Will it send anything to my owners?
No, and it cannot. It drafts and a human releases. It also never moves money, never edits your ledger, and never recomputes a total your accounting system already produced.
What is the catch?
We are building a library of these and we would like you to see the next one. That is the whole catch. The agent is yours either way.
Get the owner statement agent
Free, MIT licensed, yours to change. The next one chases late rent — sign up and you get it first.
One email when a new agent ships. Nothing else, ever.