Automation · playbook
The automation ladder: what to automate first, second and never
Automation fails when it starts with the most visible task instead of the most repeated one. The ladder ranks every workflow in a rental business by payback and by risk.
Key takeaways
- Rank workflows by frequency × decision-cost, not by how annoying they feel.
- Anything touching money, safety or a complaint stays human-approved — permanently.
- An automation without a weekly review rhythm silently drifts within two months.
Every operator who has tried to automate has automated the wrong thing first. The wrong thing is usually the task that irritates you most. The right thing is the task you perform most often, where the decision is already fully determined by facts you have written down.
Rank by frequency, not by irritation
Take the last thirty days and count how many times each recurring task happened. Then score how much judgement each one required on a three-point scale: none, some, real. Frequency times judgement gives you the honest order of operations. High frequency and no judgement is the top of the ladder. Low frequency and real judgement is the bottom, and it belongs there.
| Rung | Workflow | Automation mode | Typical payback |
|---|---|---|---|
| 1 | Routine guest questions answered from documented facts | Full automation with escalation rule | Days |
| 2 | Turnover scheduling and cleaner dispatch | Full automation, human exception queue | 2–3 weeks |
| 3 | Maintenance triage and severity classification | AI-assisted, human approves the work order | 1 month |
| 4 | Pricing adjustments inside pre-set bands | AI recommends, human confirms weekly | 1–2 months |
| 5 | Owner reporting narratives | AI drafts, human signs | 1 quarter |
| Never | Refunds, safety incidents, complaint resolution, policy exceptions | Human decision, AI prepares the brief | — |
The guardrails that make it survivable
- Ground every automated response in a single verified source of property facts; never in a general model's memory.
- Log every automated action with the reasoning attached, so a bad outcome is diagnosable rather than mysterious.
- Set a hard boundary on money: no automated compensation, no automated refunds, no automated policy exceptions.
- Review the exception queue weekly. The exceptions tell you what to document next.
62%
of guest messages in a typical portfolio are answerable from documented facts
3.4 hrs
median weekly time returned per 10 units after rungs 1–2
0
acceptable automated decisions involving money or safety
Why automations drift
A property changes. A lock code rotates, a parking rule changes, a cleaner leaves. The automation keeps running with yesterday's facts and quietly becomes a liability. The fix is not more technology; it is a standing fifteen-minute review of the exception queue and the knowledge base every week, owned by a named person.
Start with the prompts that already work
The Prompt Library holds production prompts for each rung of the ladder, with the guardrail already written in.
How many units do I need before automation pays back?
Rung one pays back at any size because it is priced in interruptions, not hours. Rungs two and three usually need eight to ten units before coordination overhead is large enough to matter.
Should guests be told a reply was automated?
Disclose the channel, not the mechanism. What matters is that a human is reachable and that anything sensitive reaches one immediately.
Operations Practice
Operations & maintenance
Turnover design, vendor management, preventive maintenance and the unglamorous systems that decide whether a portfolio scales.