An AI concierge is not a feature you ship once. A property manager who buys one is not buying the version they see in the demo — they are buying whatever it becomes over the next two years, because guest questions do not stand still and neither does the model answering them.
So the useful question to ask any vendor is not “what can your AI do?” but “what could it not do last quarter that it can do now?” Here is our answer for this month, and the process behind it.
Three ways an AI concierge actually gets smarter
“Smarter” gets used loosely in this industry, usually to mean a newer underlying model. In practice, a concierge that handles real guests improves along three separate axes, and a bigger model only helps with one of them:
- What it knows — how much of your operational reality it can see and quote accurately
- What it does — the actions it can take on your systems, not just the sentences it can produce
- How it hands off — how reliably it recognises its own limits and gets a human involved
That distinction is worth dwelling on, because it explains why swapping in a larger language model is not the upgrade path people assume it is. A model with no access to your deposit policy cannot answer a deposit question no matter how capable it is — it can only produce a fluent guess, which is worse than saying nothing. Most of the work in making a concierge genuinely useful is not model work at all. It is connecting it to the systems where the answers actually live, giving it the ability to write back to those systems, and making the handoff to a human reliable enough to trust.
This month’s releases move one axis each.
1. It knows your deposit rules
Deposits generate questions, and they generate anxious ones. A guest who is about to have several hundred euros held against their card wants to know the number, when it comes back, and what could stop it coming back. They ask before they arrive, they ask in the same breath as everything else, and they ask outside office hours.
The Virtual Concierge now reads your property’s active deposit rule directly, so it can answer with your actual policy rather than a vague gesture at one:
- The deposit amount for that specific property
- The release and refund conditions attached to it
- The expected timeframe for the guest getting it back
Critically, it does this regardless of where the guest is in the check-in flow. This is the part that matters operationally. A deposit only becomes a payment step once a guest starts their online check-in — but guests ask about money long before they start any process, often within minutes of booking. A concierge that can only see the payment step is blind precisely when the question is most likely to arrive.
Working from the property’s rule instead means the answer is available from the moment the reservation is confirmed:
Guest: hi, is there a security deposit for this apartment? just want to know before we book flights
Concierge: Yes — there’s a refundable security deposit for this property. It’s held against your card rather than charged, released after check-out once the property has been checked, and typically back with you within a few days. You’ll see the exact amount and terms when you complete your online check-in.
The alternative, which is what most concierges do, is to tell the guest to wait for a human. That answer generates a new chat for you to answer, costs the guest a day, and teaches them not to bother asking the chat next time.
If you are already automating deposits and tourist taxes in the check-in, this closes the loop on it: the policy you configured once is now something the concierge can explain, in the guest’s language, at any hour.
2. The Virtual Concierge captures arrival and departure times from ordinary conversation
Guests volunteer their timing constantly, without being asked and without using any particular form of words. “We’ll get in around six.” “Flight lands at 11, so maybe one?” “We’re out by nine tomorrow.” Until now that information sat in a chat transcript, where your cleaning schedule could not reach it. Arrival times were only captured as a side effect of an early check-in request – so if the guest was not asking for a favour, nothing was recorded, and departure times were never captured at all.
The concierge now records expected arrival and departure times on the reservation whenever a guest states them:
- Both times can be captured from a single message — “we land at 11 and have to be out by 9 on Friday” writes both
- If the guest changes their mind later, the revised time replaces the old one
- A guest-stated time survives a later PMS sync, so it is not quietly wiped out by a routine data refresh
- It never overrides a time you have already confirmed
- An informational mention does not trigger an upsell pitch
How the times are prioritised
Three sources can disagree about when a guest is arriving, so the order of precedence matters more than the capture itself. It works top down:
- A time you have confirmed — an agreed early check-in, a negotiated late departure. This always wins. Nothing the guest says in chat can overwrite a commitment you have already made
- A time the guest stated in chat — better information than a default, and now durable enough to survive a sync
- The property’s standard check-in and check-out window — the fallback when nobody has said anything
The practical effect is that guest-stated times fill in the enormous middle ground where you previously had nothing, without ever putting your own confirmed arrangements at risk.
Two behaviours that matter more than they look
It does not pitch an upsell at an informational mention. A guest telling you when they arrive is not asking to buy early check-in, and the concierge no longer treats it as an opening. The fastest way to make an AI concierge feel cheap is to have it try to sell something every time a guest opens their mouth. Capturing the information and simply confirming it is the correct response — and it is what makes the genuine early check-in upsell land as helpful when a guest actually does ask for one.
It never claims to have recorded something it did not. If a guest says something genuinely ambiguous — “we’ll get there sometime after lunch, depends on traffic” — nothing is written to the reservation, and the concierge does not reply as though it has captured a time. This sounds like a small thing. It is the difference between a tool your team can plan around and one they have to double-check, and quietly it is the harder engineering problem of the two.
The operational payoff is a schedule that reflects reality: staggered arrivals your team can see, departure times your cleaners can plan turnovers around, and a sharp drop in “what time are they getting here?” messages between your own staff.
3. Escalations that cannot be lost in silence
Every AI concierge hands conversations to humans. Recognising the limit is the feature, not the failure. But a handoff is only as good as what happens next – and the weak point in every escalation system is not the escalation, it is the silence after it.
Previously an escalation notified your contacts once. If nobody picked it up, nothing chased it. The incident stayed open, the guest waited, and you found out when they complained, and that is the most expensive possible way to learn about a problem you were told about hours earlier.
The Virtual Concierge now runs re-alerts on unaddressed escalations — an escalation on the escalation. The design question was what “unaddressed” means, and the answer is that it means two different things, so there are two independent clocks:
- Not yet responded to — nobody has replied to the guest or picked the incident up. The guest is still sitting in silence
- Not yet resolved — someone replied, but the underlying problem is still open. The guest has been acknowledged and then forgotten
The second failure mode is the sneakier one, and it is the reason a single timer would not have been enough. A reply is not a fix.
You set the windows, per priority
Each clock is configured per priority — High, Normal and Low — so a broken boiler and a Wi-Fi password question do not share a deadline. You write the windows in plain terms: 15m, 2h, or 1d 4h. Leave any cell empty to switch that clock off entirely.
A reasonable starting point, to adjust against how your team actually works:
| Priority | Not yet responded to | Not yet resolved |
|---|---|---|
| High | 15m | 2h |
| Normal | 1h | 8h |
| Low | 4h | 1d |
One thing to factor in when you choose these: the clocks run on wall-clock time. They do not pause overnight. If you set a two-hour resolution window on Normal priority, a 23:00 escalation will breach at 01:00 whether or not anyone is rostered. Either set windows that survive your quietest hours, or point the re-alert at whoever genuinely is on call.
How to Configure your Re-alerts on unaddressed escalations
Step 1 — Open your AI Escalation Rules
In the left-hand menu, click AI & Messaging to expand it, then choose AI Escalation Rules.
Scroll down to the Re-alerts for unaddressed escalations section.
Step 2 — Switch on Enable re-alerts
Until you turn it on, this is the only control you’ll see — everything else stays hidden.
Click the Enable re-alerts toggle.
Step 3 — Set your time windows
Enabling re-alerts reveals a grid with two rows, because “unaddressed” means two different things, each with its own clock: not yet responded to (nobody has replied or picked it up — the guest is still in silence) and not resolved (someone replied, but the issue is still open — the guest was acknowledged then forgotten).
The three columns are your priorities, so a broken boiler and a Wi-Fi password question don’t share a deadline. Accepted formats are 15m, 2h, or 1d 4h.
Leave a cell empty to switch that clock off.
Suggested starting point: High 15m / 2h, Normal 1h / 8h, Low 4h / 1d.
Step 4 — Choose who gets the re-alert
Either the same contacts as the original escalation, or a different contact, since the first alert already went to whoever missed it.
Step 5 — Pick the delivery method and the contact
The method you choose in the dropdown does two things: it decides how the re-alert reaches your contact, either via WhatsApp, Email or Both.
Three details that make it usable rather than annoying
- Breaches are coalesced into one digest per person — “3 escalations need attention”, with a link back to your inbox, rather than three separate pings
- Each clock fires at most once per incident. There is no escalating spiral of reminders, and no way for one stuck incident to bury your notifications
- You can route re-alerts to a different contact — a duty manager or owner — by email, WhatsApp, or both. The second alert reaching the same person who already missed the first one is not much of a safety net
If an incident is resolved and then reopened, its resolution clock re-arms, so a premature “fixed” does not buy permanent silence. Alerts respect each recipient’s language.
Pair the whole thing with internal chat notes for shift handovers and notification preferences tuned to unanswered messages, and an escalation stops being something that can quietly evaporate between two shifts.
How we keep it improving
Features are the visible part. The reason they keep arriving is the part worth explaining, because “we’re always improving our AI” is a claim anyone can make and almost nobody substantiates.
Every change ships against an eval suite. The concierge’s behaviour is tested on a library of real conversation scenarios before a change goes out, and new capabilities arrive with new scenarios added to that library. This is the difference between improving an AI and merely changing it. Prompts are not code: a tweak that fixes one behaviour can silently degrade two others, and without evals you find out which ones from your guests, weeks later, in a review.
We read real conversations. The deposit improvement above did not come out of a roadmap workshop. A property manager told us the concierge could not answer a question it should obviously have been able to answer; we pulled the trace of that actual conversation, found the gap between what the guest asked and what the concierge could see, and closed it. That is the whole pipeline, and it is the fastest one we have. If your concierge gives a guest a weak answer, send it to us.
What to check in your settings
- Turn on re-alerts for unaddressed escalations. They are off by default, so this one needs a deliberate visit to your AI Escalation Rules page. Start tight on High priority and generous on Low
- Point the second alert at a different person than the first, especially if your first contact is the one most likely to be mid-turnover when it fires
- Check your windows against your scheduled staff, remembering the clocks do not pause overnight
- Check your deposit rules are active and current. The concierge quotes them now, so an out-of-date policy becomes an out-of-date answer
- Tell your team the arrival times are there. The capture is fully automatic, and ready to be used.
Common questions
What happens when the concierge does not know the answer?
It escalates to a human rather than guessing, which is the behaviour you want — a confident wrong answer about a deposit or a key code costs far more than a short wait. As of this month, if nobody picks that escalation up inside your window, it chases itself.
Will it try to upsell every guest who talks to it?
No. A guest mentioning their arrival time is treated as information, not as a buying signal, and is recorded without a sales pitch attached. Upsells are offered when a guest asks for something that maps to one.
Does a guest-stated arrival time overwrite what is in my PMS?
It takes precedence over a routine sync, so the better information is not lost — but it never overrides a time you have already confirmed. Your commitments stay authoritative.
Can I switch a new capability off?
Yes. New concierge capabilities ship behind switches, and escalation re-alerts have their own master toggle that preserves your configured windows when you turn it off, so switching it back on later does not mean re-entering everything.
Does the concierge answer in the guest’s language?
Yes, and the host-side alerts follow each recipient’s language too — a re-alert reaches your duty manager in the language they work in.
See these new features in action
If you are already using UpMarket, all three improvements are live in your account. Two need nothing from you at all; the third takes about a minute on your escalation rules page and is the one most worth doing today. Start using them today.
If you are evaluating an AI concierge for your portfolio, ask every vendor on your shortlist the question at the top of this post — what could it not do last quarter? The answer tells you more about the next two years than any feature list will. We would love to show you how UpMarket answers it. Contact us to book a demo and see these updates in action, or try UpMarket for free.
We seamlessly connect property managers, guests and local businesses to deliver a one-stop-shop for your guest needs.
