Try AI on resident replies: start with drafts that know the history
Start with AI drafts built on the full history, checked by staff before sending, and keep urgent repairs out.
Key points
- Before drafting, gather the full history
- Staff check each AI draft before sending
- Staff own urgent repairs, not AI drafts
- Compare, then decide whether to go on
Contents
Before drafting replies, gather the full history
A&A perspective
If you try AI on resident replies, gather the request and its past history before you let the AI write anything. A reply draft here means a reply the AI writes first after reading a resident's message, which a staff member then checks and sends.
A&A perspective
The history means the open requests on the same unit, messages to vendors and their replies, and a record of what the resident was last told. Two messages with the same words can call for different next steps, depending on whether it is a first contact or a follow-up on a repair already being arranged.
From the sources
AppFolio, which makes software for property management companies, uses AI from Anthropic, an AI developer, to suggest replies to residents. The case study is one Anthropic published on its own website.
From the sources
For complex messages, the replies are described as drawing on the resident's history and current situation, existing maintenance tickets, and property-specific policies.
From the sources
It also says that even a simple maintenance request may require checking for duplicate tickets, reviewing maintenance history, and coordinating vendor schedules.
A&A perspective
Automate & Augment (A&A), which writes this article, thinks this is the first thing a small management company should get right. In our view, a reply draws more messages on the same issue less because it is badly worded than because it was written without knowing what had already been promised.
Hypothetical example
Say the bathroom fan in one unit has stopped. Last week you contacted a vendor, but the visit date is still not set. Then the resident writes, "When will it be fixed?" A draft written without the history might say, "We will arrange a vendor."
Hypothetical example
With the history laid out beside it, the draft can say only what is certain: "We have contacted the vendor and are confirming the visit date." The staff member checks the date with the vendor and adds the answer to the same request's record. When someone else looks the next morning, nobody books the vendor twice.
A&A perspective
At some companies, request records are scattered across phone notes and email and are not tied to each unit. In that case, start by organizing the records before trying AI drafts.
Staff check each AI draft before sending
A&A perspective
Do not send AI drafts to residents as they are; a staff member checks each one before it goes out.
From the sources
The AppFolio case says the AI has suggested more than half a million responses, and mentions the rate at which those suggestions are accepted.
A&A perspective
To measure an acceptance rate, a person has to look at each suggestion and choose whether to use it. From that, A&A reads the setup as one where people choose which suggestions to use. A small management company trying this should start the same way.
A&A perspective
Keep the staff check to three points. First, which request the reply is tied to. Second, whether any dates or costs in the draft are backed by the history. Third, whether it states as settled anything that has not been decided.
A&A perspective
Decide vendor orders and who pays separately from checking the draft. A correctly worded reply still does not settle which vendor to use or who bears the cost. When the owner's approval is needed, say in the reply that you are waiting for it, and the staff member goes to get it.
From the sources
Anthropic publishes a guide for developers on how to build AI systems. The guide recommends starting with the simplest solution possible and only adding complexity when needed. It also says that for many applications, improving a single AI call with retrieved information and examples is usually enough.
A&A perspective
A&A sees one AI call that produces a draft, which a person then checks, as fitting that simple form. Whether to also hand sending or vendor booking to the AI is a decision to make after seeing what happens with drafts.
| What to check | Where to look in the history | Bathroom fan example |
|---|---|---|
| Which request the reply is tied to | Open requests on the same unit | Reply as a follow-up to the fan repair the vendor was contacted about last week |
| Whether dates and costs are backed | Messages to the vendor and their replies | The vendor's reply gives no visit date, so the draft gives none |
| Whether anything undecided is stated as settled | What the resident was last told, and the vendor's replies | Write "we are confirming the visit date", not "we will arrange a vendor" |
Staff own urgent repairs, not AI drafts: send them to a separate line
A&A perspective
Messages about urgent repairs do not go into the drafting process; they go to a separate line where a person picks them up at once. Damage could spread while the message waits for a draft.
From the sources
The AppFolio case says the messages a management company receives range from urgent maintenance requests to sensitive policy discussions. It also lists making quick decisions about maintenance dispatch as part of the staff's work.
From the sources
The developer guide mentioned above also describes routing different types of customer queries into different downstream processes. Its examples are general questions, refund requests, and technical support.
A&A perspective
A&A reads sending urgent repairs to a separate line as the very first step of that kind of routing. The guide itself does not use urgent messages as an example.
A&A perspective
A&A treats separating urgent messages first as a precondition for the trial. Decide what counts as urgent with the standard your company already uses. When you are unsure whether a message is urgent, treat it as urgent.
A&A perspective
The reason is that the cost of a wrong call is not the same both ways. If a person picks up a message that was not urgent, it only adds a little work. If an urgent message waits for a draft, damage can spread in the meantime.
Hypothetical example
Say a message that meets your company's standard for urgent arrives at night. Take it out of the drafting process at that point and alert the staff member on call right away.
A&A perspective
Urgent messages a person handles still go into the same unit's request record once the repair is arranged. That way, when a follow-up on that issue arrives later, the draft can use the history.
A&A perspective
How to decide who receives each type of inquiry is covered in "Sort resident messages with AI: decide owners before reply drafts."
Compare with before the trial, then decide whether to go on
A&A perspective
Decide whether to go on by comparing the effort per reply and the number of repeat messages on the same issue with the period before the trial. The number of drafts used does not tell you whether residents' problems were resolved.
From the sources
AppFolio's principal product manager says property managers spend up to half their workday on communications with residents, owners, and vendors. Beta customers (property managers) are reported to have saved about 11 hours a week on communications alone.
A&A perspective
But this is a case study published by the vendor, and it does not say how the comparison was made. The same hours will not necessarily be freed at your company, so measure your own numbers yourself.
A&A perspective
Measure two numbers: the time spent per reply, and the number of repeat messages on the same issue after a reply. For example, compare the two weeks before the trial with the two weeks after it, counted the same way.
A&A perspective
Count a repeat message by tying it to the original request's record. If you gathered the history as in the first point, you can pull this number straight from the records.
A&A perspective
Time saved is not profit by itself. It only becomes money when it shows up as more properties handled by the same staff, or less overtime. Also subtract the time spent checking drafts and the tool's fees.
A&A perspective
At a company with few messages, the difference before and after the trial will be small. If checking drafts takes more effort than that difference, deciding not to go on is also an option.
A&A perspective
If repeat messages on the same issue do not fall, regather the history before reworking the wording of the drafts. Review the links between requests and their history.
Summary
A&A perspective
Before drafting replies, gather the request and its past history.
A&A perspective
Staff check each AI draft on three points before sending: which request it is tied to, backing for dates and costs, and nothing undecided stated as settled.
A&A perspective
Staff own urgent repairs, not AI drafts: send them to a separate line and alert someone at once.
A&A perspective
Compare the effort per reply and repeat messages on the same issue with the period before the trial, then decide whether to go on.
What makes AI work on resident replies is less the writing than whether requests and their history are in order. Open one request that drew another message this week: if that unit's history is in order, try drafts on a small scale; if not, start by organizing the records.
FAQ
Can we try this without property management software?
Yes, as long as requests are recorded per unit. A spreadsheet or email works if you can see the unit's requests, vendor exchanges, and what the resident was last told in one place.
Are there messages we can send from an AI draft without checking?
Not while you are starting out. Have staff check every draft before sending, and keep a record of the kinds of drafts that could be used without edits.
Can we start if messages come in by phone, email, and an app?
Yes. You do not need to connect every channel at first; start with the one channel that gets the most messages.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- AppFolio streamlines property management communication with Claude in Amazon Bedrock
Anthropic · Publication date not stated
Accessed 2026-10-24 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-10-24