The problem
Extended stays are priced differently from hotel nights. A 45-night booking in Dubai Marina has to be competitive against monthly rentals, serviced apartments and hotels at once, and the right price moves with demand, season, competitor supply and the guest's own history. A small team was setting these prices by hand across a fast-growing inventory. Pricing was slow, inconsistent, and left money on the table in both directions.
What we built
A pricing agent that sets the nightly rate for every stay request. For each request it reads more than 1,200 live signals: demand indicators, competitor rates for the same property on other platforms, seasonality, length of stay, lead time and guest context. It proposes a price change, then passes through a trust layer before anything is published.
The trust layer
- Step limits: no single decision may move a price by more than a bounded percentage per cycle. Larger proposals are clamped and flagged.
- Floors and ceilings: hard price bounds per property, set by the business, enforced in code.
- Parity checks: the agent cannot publish a price that breaks our positioning against the same home on booking sites.
- Reversal path: every published price carries a snapshot of the previous one and a rollback window.
- Audit trail: inputs, proposal, each guardrail result and the final action are logged for every decision.
How autonomy was earned
The agent started in suggest-only mode, scored against what the pricing team did. Once its proposals beat the manual baseline on a fixed evaluation set, it was allowed to act within tight step limits. The limits widened as the evidence grew. Today roughly 70% of pricing decisions execute without a human; the remaining 30% are exactly the moves that exceeded a limit and went to a person. That split is by design.
Results
Take rate more than doubled from 10% to 22.98%. Revenue per available room rose 12 to 18% and occupancy 8 to 12%, while the platform held 99.95% uptime. The pricing team moved from entering numbers to reviewing exceptions.
What this means for you
If a human in your business sets a number repeatedly, an agent can probably set it better. The work is not the model. The work is deciding the step limit, the floors, the reversal path and the evaluation set. That is what an AI safety review covers.