We let AI agents run 23 vacation rental properties. Here is what we never let them do

Описание и применение

AI‑агенты управляют 23 объектами аренды, выполняя задачи по ценообразованию, сообщениям и срокам бронирования с жесткими ограничениями, которые реализованы в коде, а не в подсказках. Это снижает ручную работу, повышает доверие и обеспечивает контроль над критическими решениями.

Задача

Managing 23 vacation rental properties required repetitive, time‑consuming tasks such as answering parking questions and setting nightly prices, which ate into the operator’s evenings.

Результат

The operator now manages 23 properties with AI agents that handle repetitive tasks while preserving human oversight for critical decisions, leading to reduced workload and consistent application of business rules.

Рабочий процесс

• 1. Deploy agents in Suggest mode for each task (pricing, messaging, expiry).
• 2. Agents generate proposals and return them to the operator.
• 3. Operator reviews, edits if necessary, and approves or rejects the proposal.
• 4. Once approved, the agent can be switched to Auto mode for that task.
• 5. Code enforces constraints (price floor, no irreversible actions, no money spent) after the model returns.
• 6. Agents act only when no human has intervened (silence) and in the safe direction (e.g., decline booking if no response within 15 minutes).

Доказательства

The guardrail goes in code, not in the prompt

Источник

Ship every agent switched off, and make the user turn it up

Источник

When it fails, work out which direction is safe

Источник