At ten properties, trying an AI tool is low stakes. Switch it on, watch a few conversations, switch it off if it's wrong. At 500 or 5,000, the same mistake repeats hundreds of times a week before anyone notices, across brands, regions and teams.
Large operators need a different way in. Not a longer demo or a thicker RFP, but a short, structured trial on their own guest requests that answers the questions that matter at scale. This guide sets out how to run one, and what to settle before and after.
Why enterprise is different
The error maths changes. A property typically brings 10 to 20 guest messages a week. Across 1,000 properties, a 1% error rate is 100 to 200 wrong replies a week, each one a promise your housekeeping or maintenance team didn't know was made. At scale, how the AI escalates matters more than how much it automates.
Nobody configures it by hand. A tool that needs a knowledge base written and maintained per property doesn't survive contact with a large portfolio. Door codes, parking, staff and building works change every week. The AI has to learn from what you already have and keep itself current. See why keeping the AI up to date is the real job at scale.
Policies multiply. Several brands often have several answers to the same question: early check-in fees, pet policies, how much discretion the team has on refunds. The AI will apply whatever it's told. Somebody needs to decide what that is.
Teams are specialised. Guest communication, field operations, owner relations and revenue sit with different people. An autopilot touches all of them, so all of them need a say in the trial.
Before the trial: settle four things
- Your standard. What review or communication score would you be unhappy below, and which brands or regions are under it today? That's the bar the trial is measured against.
- What the AI must never decide. Every operator has one or two: refunds over a threshold, anything involving a neighbour complaint, owner stays. Write them down first.
- Who reads the replies in week one. The single best predictor of a successful rollout is someone on your team giving feedback in the first week. Operators who go quiet for a month and then judge the output rarely get there.
- Where policies conflict. If brands disagree on a policy question, decide whether the AI follows each brand's rule or a single standard, and who settles disputes.
The trial: a slice, a shadow week, two weeks live
1. Pick the slice
A few hundred properties across one or two brands, chosen to represent the wider portfolio. Include the awkward ones: rural properties with key collection, buildings with tricky access, the brand with the highest guest expectations. A slice made only of easy city flats will flatter any tool.
2. Connect and watch (one week)
A vendor that learns from your history should be able to load months of past conversations on connection, work out your rules and property knowledge from them, and show you the rules it found in plain English for you to correct. Expect that within a day, not weeks.
Then a week of shadowing. The AI reads every incoming message and prepares its reply or escalation, with its reasoning, but doesn't send. Your team sees what it would have done and corrects it. This is the cheapest week of feedback you'll ever get.
3. Go live and run the checks (two weeks)
Switch it on for the slice and measure four things:
| Check | What to look at |
|---|---|
| Automation | Share of requests handled end to end with nobody stepping in, PMS update included. Not drafts approved. See how to measure automation rate. |
| Mistakes | Every wrong reply or missed escalation, logged. Does the list get shorter each week once each is reported? |
| How it reads | Read twenty full conversations as a guest would. Does it sound like your brand? Is it warm where it should be? |
| What reached your team | Were escalations the right ones, with context attached, to the right person? Did anything reach your team that the AI should have handled? |
Two weeks is enough for both sides to see how working together would go. If the numbers don't hold up in the trial, don't sign on the promise that they will later.
What good looks like after three weeks
- Around 70 to 80% of requests on autopilot for a well-documented slice, with a clear path for the rest. Figures far above that deserve a closer look at what's being counted.
- A shrinking mistake log. Problems fixed stay fixed.
- Escalations your team trusts, reaching the person responsible for each property.
- No drop in your guest experience. Track the share of stays with no complaint, not just review averages. See scoring every stay.
Widening the rollout
Expand by brand or region, not all at once. Each new brand brings its own policies and tone, so give each a short shadow period. Keep the same four checks running monthly, and keep someone accountable for the mistake log.
As guest messaging settles, look at the operations around it. Coordination with housekeeping and maintenance, and re-planning staff days, are often where the next gains are, especially for operators running field teams across many properties.
Questions to put to an enterprise vendor
- How deep is your integration with our PMS? Can you write changes back (arrival times, tasks, notes), or only reply?
- How quickly does new AI capability reach customers on our PMS?
- What did you learn from our history on day one, and how do we correct it?
- What does the AI never decide, and how do we change that list?
- Show us your mistake log from another customer's first month, anonymised.
- How do you handle different policies across our brands?
- Who on your side reads our conversations in the first two weeks?
For the full list of questions, see how to choose an AI autopilot. For how the main enterprise options compare, see the market comparison.