Ask any AI guest messaging vendor how much of your inbox it will handle and you'll get a number. Usually somewhere between 70 and 90%. Ask how that number is worked out and the conversation tends to get vaguer.
That matters, because the automation rate is the figure that decides whether the tool pays for itself. Accuracy and safety are the minimum you should expect. The automation rate is what gives your team time back. This guide sets out what the number should mean, why vendor figures can't be compared side by side, and a method for measuring it yourself on your own guest requests.
What the number should mean
The useful definition is simple. The automation rate is the share of guest requests that were fully dealt with without anyone on your team stepping in.
Three words in that sentence do a lot of work.
Requests, not messages. A guest who writes "Hi!", then "Can we check in early?", then "Thanks so much" has sent three messages and made one request. Counting messages inflates the figure, because the easy acknowledgements pile up on the automated side.
Fully dealt with. Replying is often the smaller part of the job. If a guest asks for a cot, the request isn't handled until housekeeping knows, the cot is in the flat and the guest has been told. If the AI wrote a lovely reply and your team then did the coordination by hand, that request wasn't automated.
Without anyone stepping in. A draft that a person read and sent is not automation. Neither is a scheduled message your PMS was already sending before the AI arrived.
Why vendor figures don't line up
Here's what the main vendors say publicly, as of October 2026. All of it is vendor-reported, and none of it comes with a full method.
| Vendor | What they claim | What the claim is based on |
|---|---|---|
| Guesty (ReplyAI Autopilot) | Up to 80% of incoming messages | Not published |
| Hostaway (AI Auto-Reply) | Up to 90% of guest communication | Not published |
| Hospitable (Inbox AI) | 90% of guest communication | Not published |
| Conduit (formerly HostAI) | 70 to 90%, with case studies up to 96% | Customer case studies |
| Akia | Up to 90% of messages never reach staff | Customer case studies |
| Boom | Up to 80% of guest messages | Not published |
| Enso Connect | Up to 70% of messages | Customer case studies |
| Aeve | 70 to 80% of messages and tasks, as a deliberate target | Average across customers, reviewed drafts excluded |
The few figures that look more like measurements than marketing are lower. In 2024 Guesty said its most active accounts used ReplyAI for about half of incoming messages. RentalReady reported that its Maia agent fully handled about two-thirds of conversations during its beta. One Conduit customer reported 40% in its first month. Hospitable reports that about 60% of its AI drafts are sent without edits, which is a useful number but a different one: it measures draft quality, not automation.
The gap between those figures and the headline claims usually comes from one of these:
- Drafts counted as automation. The AI wrote it, so it counts, even though a person approved it.
- A narrowed denominator. The rate is worked out only across the categories the AI is switched on for, or only on the properties where it's live. Everything it was never allowed to touch quietly disappears from the sum.
- Messages instead of requests. "Thanks!" and "Great, see you then" are easy wins that make the percentage look better.
- The best customer, not the average. A city portfolio with excellent listing data will score far higher than a mix of rural cabins with keys under the mat.
- Replies without the follow-through. The guest got an answer. The cleaner was still moved by hand.
- After months of tuning. A figure that only appears after a long training period tells you little about the guests who stay with you in the meantime.
How to measure it yourself
Whether you're already running an AI tool or trialling one, the test is the same. It takes an afternoon and a spreadsheet.
1. Take a random sample of about 200 requests
Pull guest conversations from the last two weeks, from your PMS inbox or an export. Pick them at random, not the ones you remember. Two weeks catches a full turnover cycle and a weekend or two. If you run very different property types (city flats and rural houses, say), sample enough of each to see whether the rate holds across both.
Two hundred is enough to be useful. At that size, a measured rate of 60% means the true figure is very likely somewhere between 53 and 67%. That's precise enough to compare two tools or watch one improve.
2. Split each conversation into requests
Most conversations contain one request. Some contain three ("Can we arrive at 1pm, is there parking, and can you recommend somewhere for dinner?"). Count each one. Ignore pure pleasantries.
3. Mark what happened to each request
Use one label per request:
| Label | What it means | Counts as automated? |
|---|---|---|
| Handled end to end | Correct reply, and any follow-on work (task, team update, PMS change) done without a person | Yes |
| Replied, work done by hand | The AI answered, but someone on your team did the coordination | No |
| Escalated correctly | The AI handed it to your team, and it was right to | No, but it's a good outcome |
| Handled by a person | The AI didn't touch it, or a person took over | No |
| Wrong or should have escalated | Incorrect information, a promise you couldn't keep, or a judgement call it shouldn't have made | No, and it counts as an error |
4. Work out the numbers
- Automation rate = handled end to end ÷ all requests.
- Error rate = wrong or should have escalated ÷ requests the AI acted on alone.
- Escalation quality = escalated correctly ÷ all escalations. If your team keeps receiving things the AI could easily have answered, the system is escalating too much. That's the safer failure early on, but it costs time.
Keep the error rate separate from the automation rate. Even one wrong reply in a hundred is enough to make your team go back to checking every message, and once they do, the time saving is gone.
5. Repeat it, and watch the errors
One sample tells you where you are. The more important signal is what happens next. Keep a list of every wrong reply and raise each one with the vendor. In a system built for accuracy, a problem fixed stays fixed, so the list gets shorter week by week. If new problems appear as fast as the old ones are dealt with, more training is unlikely to change that.
What a realistic number looks like
Public figures and operator reports point to a fairly consistent picture. Measured honestly, around 40 to 60% in the first month is common. Portfolios with well-documented properties can reach 70 to 80% once the system has their information.
The goal is not the highest number possible. A good system also knows which conversations are better handled by a person: the cancelled flight, the upset guest, the neighbour complaining about noise. Around three-quarters on autopilot, with your team on the rest, is a balance many operators find works. The reason is arithmetic: a property brings 10 to 20 guest messages a week, so across 300 properties one wrong reply in a hundred is 30 to 60 bad promises a week. Escalating correctly matters more than replying. A figure close to 100% should make you ask what's being left out of the sum, or what's being answered that shouldn't be.
What it's worth to you
Once you have your own measured rate, put it in the calculator with the size of your team and what an hour of their time costs.
Work it out for your operation
Costs you today
$20,800
a month · $249,600 a year
Comes back to you
$10,400
a month · $124,800 a year
Team hours back
347
a month
Counts recovered team time only, after upkeep but before the cost of the tool. It leaves out upsell revenue and any effect of faster replies on reviews or ranking. Use the share you measured on your own conversations, not a vendor's headline figure.
This counts recovered team time only. It leaves out upsell revenue and the knock-on effects of faster replies, from better reviews to stronger OTA ranking. That makes it a conservative figure, and a more honest one to take into a budget conversation. Remember to subtract what the tool costs, and any time someone spends maintaining its rules. If a person on your team has to keep reviewing and updating it, that's headcount, and it comes off the saving. On a large portfolio that upkeep can run to tens of hours a month, which is why how a tool keeps itself current matters as much as its headline rate.
Other numbers worth tracking
The automation rate isn't the whole picture. Alongside it, track:
- Time to first reply, especially overnight and at weekends. This is where autopilot shows most clearly, and Airbnb's Superhost criteria still ask for replies to 90% of new messages within 24 hours.
- Your communication score in reviews, month by month. A slow drift down is often the first sign that replies are correct but cold.
- Time from escalation to a person picking it up. An escalation that sits in an inbox for three hours isn't much better than a wrong answer.
- Errors per week, and whether that number is falling.
- The share of stays with no complaint at all. Read each stay's messages and review, and count the stays where nothing went wrong. If automation rises and this falls, something is being answered that shouldn't be. We set out the method in scoring every stay.
For the questions to ask a vendor before you get this far, see how to choose an AI autopilot. For how the main tools compare, see our comparison of AI guest messaging software.
Sources
- Guesty, ReplyAI Autopilot and AI Task Creation launch (April 2026): press release
- Guesty ReplyAI usage, 2024: Travolution
- Hostaway AI: hostaway.com
- Hospitable AI usage, H1 2025: ShortTermRentalz
- RentalReady Maia beta: ShortTermRentalz
- Conduit customer figures and pricing: conduit.ai
- Akia: akia.com
- Boom: boomnow.com
- Enso Connect: ensoconnect.com
- Airbnb Superhost criteria: Airbnb Help Centre