AI & autonomy
Bounded autonomy
An approach where an AI agent acts on its own only within limits you set, per-task spend and approval thresholds, and asks for a human decision above them.
Glossary
AI & autonomy
Bounded autonomy is the principle that an agent should act freely inside clearly drawn limits and stop at their edge. Rather than an on-or-off 'let the AI run everything' switch, you set typed thresholds per task: approve maintenance under a dollar cap, auto-send routine tenant replies, retry a failed rent payment. The agent handles anything inside those bounds automatically and routes anything above them to you. The limits are the landlord's dials, not the software's.
This is what separates useful autonomy from a liability. The bounds are enforced by a single central gate, they sit beneath a floor of actions that can never be automated regardless of settings, and they are recorded so you can see exactly what the agent was and was not allowed to do. As trust builds, the design intent is that you widen the thresholds you are comfortable with; autonomy is meant to graduate with evidence, not to arrive all at once.
The idea borrows from robotics and safety engineering, where a system is given freedom to act only within a verified safe envelope. Applied to property management, it lets a small landlord get the leverage of automation without handing over blanket authority over their money or their tenants.
Authority ladder and maintenance scenario
A practical authority ladder is observe, draft, recommend, execute reversible work, and execute consequential work. Eligibility does not grant authority: evidence quality, freshness, limits, and exception rules must pass.
Example: an agent may classify a no-hot-water request and schedule a contracted vendor inside an emergency rate card. It escalates disputed access, changing safety conditions, outside scope, or excess cost.
Boundary failures and pilot checklist
Failures include treating missing data as permission, widening scope after a pilot, retrying uncertain actions, and leaving temporary overrides in place.