A&A INSIGHTS
Sort resident messages at intake before comparing reply drafts: what AppFolio's two-layer setup shows small property managers
For owners of small property-management companies: use Anthropic's AppFolio case and its engineering essay on building effective agents to design the intake-routing layer before comparing AI reply-draft products.
日本語で読むTHE STARTING POINT
When AI is considered for replying to resident messages, the natural first move is to compare candidate products by the quality of the reply draft. In a small property-management company, though, without a decision about which case at intake goes to which queue, a better draft rarely reduces the owner's decision workload. Using Anthropic's AppFolio case and its engineering essay "Building effective agents" as lenses, this article suggests the intake-routing design that should be settled before any reply-draft comparison. A hypothetical six-person management company illustrates a one-sheet layout of category, confidence threshold and downstream queue.
Comparing only the reply draft hides the intake burden
Hypothetical example
Picture a six-person residential-management company. It handles several hundred units. The owner and a lieutenant share the first line of email, phone and chat; a contracted coordinator handles repair dispatch; one salaried employee handles lease renewals. A vendor now proposes an AI tool that drafts replies to resident messages. This is a hypothetical setup. Vendor demonstrations tend to leave an impression through the politeness of a draft to a single representative message. What the owner is actually struggling with, however, is not the wording of drafts. It is that intake of an emergency water leak, a lease-renewal schedule question, and a landlord approval for repair cost all land in the same inbox with no priority applied.
A&A perspective
A&A's suggestion for owners of small property-management companies is to avoid placing the look of the reply first when comparing AI tools. A draft is a downstream step that presumes the correct handler and the correct information have already been assembled at intake. Making a draft high quality while intake sorting is vague leaves the harder burdens in place: missed emergencies, replies sent before the landlord has approved a cost, and lease-related cases starting on the wrong desk. Before an introduction decision, articulate in-house the axes by which today's messages should be sorted and which queue each axis maps to.
What AppFolio's two-layer setup really shows is the intake split
From the sources
Anthropic's AppFolio case page describes the company's Realm-X Messages platform as combining two Claude models. Claude 3 Haiku handles high-volume, rapid-response scenarios, while Claude 3.5 Sonnet manages complex communications. For reply generation, the page notes checking for existing maintenance tickets where applicable, incorporating resident history and current situation, and incorporating property-specific policies and amenity information.
A&A perspective
The point of this description is not that a small company should also line up Haiku and Sonnet. Numbers in the case, such as about eleven hours saved per week or about twenty-six seconds per message, come from AppFolio's scale and do not carry over to a six-person shop. What is legible is the operational outline of separating a case at intake, drawing a line between simple inquiries and complex coordination, and also separating the internal information injected into the draft. What transfers is the design view that cases are sorted before the draft, not the model layout itself.
A&A perspective
If a smaller property-management company borrows the same view, a workable first split has at least four categories: emergencies affecting the property or common areas; individual-unit maintenance requests; contract- or cost-related inquiries; scheduling and simple lookups. This four-way split is an example. The actual split should come from reading one to two weeks of your own inbox. Delegating draft generation to AI before the split is settled tends to send polite template replies to contract- or cost-related cases first, so that the resident hears from AI before the renewal handler or the landlord has been notified.
Use Anthropic's routing pattern at the intake door
From the sources
Anthropic's engineering essay "Building effective agents" defines routing as a workflow that classifies an input and directs it to a specialized followup task, and notes that this workflow allows for separation of concerns and more specialized prompts. The essay also gives the example of routing easy or common questions to smaller, cost-efficient models like Claude Haiku, and harder questions to more capable models like Claude Sonnet, rather than optimizing a single prompt across all inputs.
A&A perspective
For a small property-management company, routing is not about model selection; it is about deciding downstream destinations before intake. Taking the tentative four categories from the previous section, one workable mapping is: emergencies push a notification to the owner and lieutenant's phones; individual-unit maintenance goes to the coordinator's shared inbox; contract and cost inquiries go to the renewals handler; only simple lookups become candidates for an AI draft. Attaching priority icons inside a single shared inbox collapses quickly on a busy afternoon. The point of routing is to change where a handler receives a case, not only how it is labeled.
Feed the classifier with property-system signals, not message text alone
A&A perspective
If classification precision depends only on the resident's message text, the same sentence "the AC has stopped" will be misread as a mid-summer emergency in one case and as a routine inspection request in another. The existing maintenance tickets, resident history and current situation, and property-specific policies and amenity information that AppFolio's case cites for reply drafting are equally strong signals at the earlier classification stage. Even a small company, if these items live in its existing property-management tool, can pass those identifiers to AI alongside the message body as classification input.
Hypothetical example
Continue the six-person company example. A resident sends a short line: "the water is leaking again." The classifier should look not only at the words but also at whether the unit has any open water-related maintenance ticket from the last ninety days; whether other units in the same building sent similar messages within the last forty-eight hours; the resident's tenure and past emergency-contact count; and the owner-approval rules recorded for this property. With the same message body, an unresolved water ticket sends the case to "continuing case, urgent"; without it, the case becomes "new maintenance request, needs confirmation." These lookups do not require AI reasoning; the existing property-management tool's search APIs suffice.
A&A perspective
It is worth noting that passing these signals to the classifier does not automatically pass them to the draft step. When a draft becomes appropriate, the handler decides on the spot which of the same signals to inject into the draft, as a separate step. Streaming both stages with everything by default injects too much history into non-urgent cases, elongates AI responses and adds re-checking work for handlers. Keep classification lean and specific; keep drafts to items the handler explicitly attaches.
Draft one sheet of thresholds and queues before comparing products
A&A perspective
Once the axes and destinations are set, decide how much to trust the AI's classification output. Set the default as two-tier: if confidence is high, drop the case in the assigned queue; if confidence is low, return it to the owner-and-lieutenant intake queue. Choose thresholds by the weight of the recovery when you are wrong, rather than by the number itself. Sending a common-area emergency to a normal queue costs more to recover than sending a simple lookup to the emergency queue, so on the emergency side keep the threshold low and lean toward picking up; on the simple-lookup side keep it high and lean toward narrowing.
A&A perspective
Put the decisions on a single sheet. Category, the signals used for that category, the confidence thresholds, the downstream queue, the notification channel the handler receives, and where the classification result is written back into your system, all sit on the same page. When comparing AI products, hand this sheet to the vendor and evaluate against two criteria: can this classification logic and confidence threshold be implemented, and in what format can the classification result be written back to our property-management tool. Compare the polish of the draft only after those two criteria are met. In the reverse order, you pick a product that produces attractive drafts and then rebuild the receiving surface in-house.
A checklist for a small property-management trial
A&A perspective
The trial that supports the introduction decision starts with AI performing only classification for one to two weeks. Do not enable draft generation yet. Record the classification results next to a manual re-classification by the owner and lieutenant, and read the per-category agreement rates and the breakdown of cases that straddle the threshold. A ten-to-twenty-percent random sample is enough to see a trend while keeping added handler work bounded. If the axes themselves do not match how the desk really works, rebuild the axes before continuing.
A&A perspective
Only after the classification agreement and the recovery cost of misclassification are within tolerance, enable drafts, stepwise, and only for the simple-lookup category. Even after drafts are running, keep emergency, contract and cost categories on humans only. At three months, check that no emergency was missed, that no contract- or cost-related case leaked into the AI-draft flow, and that the simple-lookup drafts are actually reducing the handler's editing time compared with the pre-AI baseline. If any one of those fails, narrow the draft scope and return to the classification design.
A&A perspective
When A&A is asked about property-management operations, the same one-sheet classification layout and trial design are where we start. The boundary is not there to reject a given AI product; it is there to prepare the in-house queues and write-back destinations at the granularity the product classifies well. The related article "Overwhelmed by resident replies: how AppFolio's case helps split property-management work" takes up the post-reply state-continuity side. This article covers the pre-reply classification side; reading both in sequence lets you consider the design of ownership from intake through state continuity as one continuous span.
When deciding whether to bring AI into resident messages, the number worth comparing is not just the look of the reply draft. Put the classification axes and confidence thresholds, the downstream queues, and the write-back destinations on a single sheet before choosing an AI product. Design the classifier to consume not only the message body but also the structured signals from your property-management system, and draft generation can be introduced stepwise without pushing extra re-checking work back onto handlers.
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 on page
Accessed 2026-09-18 - Building effective agents
Anthropic · 2024-12-19
Accessed 2026-09-18
AI-assisted editorial production
A&A uses AI for research, writing, translation and editorial checks. Source facts, our analysis and hypothetical examples are labeled separately.
Editorial check: 2026-09-18