Nights are covered: agents for the work that happens when nobody is in

An after-hours agent takes the call or the message, writes the note into the log you will open at 08

, and pages a person only when the exception is real. The night is not staffed by a hope. It is staffed by a brief, a voice, a desk, and a list of things that are allowed to wait.

Vesper’s nights were a contractor and a hope.

The contractor took what they could. What they could not take sat in a voicemail, a WhatsApp, a missed note. By morning the exceptions had stacked. The first hour was archaeology: who called, who got a recording instead.

Customers were left with silence. Silence is also an answer. It is the wrong one.

The work we designed was smaller than a “hospitality AI” deck. Voice and chat on one agent. Automations that file the note. A page only when the brief says the night cannot hold it. The morning starts with a log.

Marcus Hale, who runs operations at Vesper, said the only sentence that matters:

Nights are covered. We come in to a log — not a pile of exceptions that waited until morning.

This is not a list of builders for hotels. It is the pattern: hotels, on-call SaaS, clinic intake. Same night. Same log. Same rule about who gets woken.

What actually happens on a night of exceptions?

A night is not a quieter day. It is a different mix of work.

The repeatable requests still come — late checkout, a door, a known outage, a morning slot. Mixed in with them are the things that cannot wait: a medical emergency, a fire alarm, a production incident that is not on the status page. Treat the whole night as one queue and you either wake someone for towels or you leave a real exception until 08

.

Vesper’s old night did both, depending on who was holding the phone.

The redesigned night sorts as it arrives.

The night is a sort, not a performance.

A late checkout is a note. A guest who cannot get in is a page. A known incident is a note plus the public line. Exposed payment data is a page. Tomorrow’s slot is a note. A caller who cannot breathe is a page, and the agent is not a clinician.

The brief is that list. Without it, the voice is timid or reckless. Both fail at 02

.

What did Google Cloud’s 2026 concierge framing actually say?

The chatbot era is being replaced.

In December 2025 Google Cloud published its 2026 AI Agent Trends Report, based on more than 3,466 global executives. Anil Jain’s recap is direct: “The era of scripted chatbots and reactive customer service is coming to an end.” Agents, he wrote, would establish “concierge-style” service as the new standard. The same recap is on video:

Jain’s March 2026 follow-up is more useful than the word concierge. For a decade, he wrote, customer-service automation meant “pre-programmed chatbots answering simple questions and deflecting support tickets.” Agents remember, act, and escalate when it matters. That is a night pattern.

The hospitality piece Google Cloud published in March 2026 names the hour. Joff Romoff and Lacey Duncan, writing about how hotel brands are using AI, put the guest-services problem in one line: “it’s smoothly handling a 2 AM message from a guest in Barcelona asking about late checkout — in Catalan.” Digital concierges, they write, handle the routine so staff can take the cases that still need a person.

Lukas Karrer, director of Direct Business at Hotelplan Suisse, is quoted there on the loop we actually run:

Any questions that stump the chatbot are escalated to our travel experts, who now have more time to respond in detail. Those answers are then fed back into the model, creating a feedback loop to continually improve responses.

We would not call the thing a chatbot. The loop is right. The night writes the exceptions into the knowledge, or the same exception will stump the line again tomorrow.

Danfoss, in the same recap, cut average email response from 42 hours to near real time by automating 80 percent of transactional decisions. Forty-two hours is two nights and a morning. None of this requires Google’s stack. It requires you to stop treating after-hours as a recording plus an apology.

How should a hotel night actually run?

Not as a concierge demo. As a desk.

A hotel night is a set of units of work. Late checkout. Extra towels. A noise complaint. A door. A reservation that was not honoured. A medical event. Some of those are notes. Some are pages. Some are a person walking down a corridor.

The agent can take the first group on voice and on chat, in the language the guest is using. Google Cloud is right about the Catalan message at 02

. The guest does not care that your night manager is asleep. They care that the late checkout is confirmed, or that a person is on the way.

What the agent must not do is invent a room, a rate, or a medical opinion. The live reservation system is the source for what is available. The on-call is the source for what is unsafe. If the property’s own data is not connected, you have a voice that describes last spring’s brochure.

A hotel agent that cannot write the note is a night receptionist with no pad.

The morning log is dull, and that is the point. Which late checkouts were granted. Which doors failed. Which pages went out, and whether the on-call picked up. The day team should not start by listening to voicemail.

This is why we will not write “best AI agent builder for hotels.” The builder is not the constraint. The brief is. A neighbouring-market builder conversation is best AI agent builder for real estate. A listing is not a night desk.

What does the same night look like for on-call SaaS?

A status page is not a night line.

Customers still write at 01

because their job is on fire, or because they do not trust the green dot. Some have a known incident. Some have a unique failure. Some have a billing question they decided was an outage.

The agent’s first job is to sort. Known incident: give the public line, file the note, do not page. Unique failure: file the case with the logs they already pasted, page if the brief says this severity wakes someone. Billing dressed as downtime: answer from knowledge, leave it for the day team.

Salesforce’s 2026 statistics say 79 percent of service leaders treat investment in AI agents as essential, and that representatives already spend only 46 percent of their time with customers. Nights make that worse. The agent is there so the on-call is woken for the exception, not for the known incident.

The morning log is a digest. Incidents handled against the public line. Cases opened. Pages sent, with the reason. Threads refused. The day team should read it in one pass. An agent does not get tired. It also does not get judgement you did not write down. Write the judgement down.

What should clinic intake do after the desk is closed?

Take the request. Refuse the diagnosis. Page the right person.

After-hours clinic work is mostly intake. A slot tomorrow. A prescription that cannot wait. A caller who should be told to use emergency services. The clinical lines are not a tone-of-voice problem. They are a list of things the agent is forbidden to do.

The agent can confirm hours, offer the next bookable slot the live calendar will honour, and take a name and a number for the morning. It can page the on-call against a short list of red-flag phrases you wrote with that clinician, not with a model card.

It cannot diagnose. It cannot suggest a dose. It cannot tell someone they are fine. A night line that sounds warm while doing any of those is worse than a recording.

The clinic test is the same as the hotel test and the SaaS test: did the morning open a log, or a pile.

The object is intake, a calendar, and a page. We will not recommend a widget that talks like a nurse.

What should a morning log contain?

The night, in a form a person can act on before the coffee is gone.

Not a transcript dump. A digest.

Vesper’s digest is this table, produced by automations rather than by a person listening to a tape. n8n files the note, pages, and assembles the morning. The agent is the ear and the voice. The workflow is the pad.

That split is the automation work and the agents work as one system. Split them and you are back to a contractor with a hope.

If you still think of this as a chatbot problem, read chatbot vs AI agent. A night that can talk and cannot file a note is the same failure as Aurelia’s first widget, on a darker clock.

When should the night wake a person?

When waiting until morning would be worse than a tired on-call.

Write that as a list. Do not leave it as a feeling.

1. Safety. A guest, a patient, a customer, a member of staff. If someone is not safe, page.
2. Access. A person locked out of a building, an account, a room they have paid for, when the agent cannot restore access from a live system.
3. Severity you named in advance. The P1. The fire alarm. The data exposure. Not the known incident already on the status page.
4. A legal or medical line the agent is forbidden to cross, where silence is also forbidden.
5. A page the on-call themselves asked for — a VIP, a contract clause, a property that is full and a guest at the door.

Everything else is a note. Towels are a note. A late checkout the system can grant is a note. A billing question at 03

is a note. A feature request is a note, and it is also a reminder that someone is using the night line as a suggestion box.

The first brief will wake people too often or too rarely. The morning team marks the pages that should have been notes, and the notes that should have been pages. The brief is edited. The clock is the only difference from a daytime review.

Where does the product sit, if the night is the test?

After the brief. After the on-call rules. After you know which system the note has to land in.

We put voice and chat on one agent so the caller is not sorted by channel. YourGPT is the surface we use for that — the same knowledge, a handoff that carries the transcript. We brief it as the ear, the voice, and the first sort. n8n writes the note and sends the page. The morning team owns the digest.

YourGPT is not the idea. The idea is the log. A prettier voice with no brief would still be a contractor.

The Vesper night line is the case we can stand in. After-hours service design. Voice and chat on one agent. Automations for notes and escalation. A digest for the morning. That is the work.

This page names Google Cloud’s report, quotes Hale and Karrer, and refuses a fake night-resolution figure. Buyers will still call you at 02

. The desk covers the night. The answer economy does not.

What should you refuse to automate at 02

?

Anything you would not let a careful stranger do with a phone and a login. Refunds above a line. Medical advice. Legal commitments. An unauthorised complimentary night. A production rollback. A prescription. A door opened for someone the system cannot identify. A public status change. A promise that a person is coming when no person has been paged.

Also anything the live system cannot confirm. If the reservation tool is down, the agent says so and pages, or it takes a note. It does not guess that room 412 is free. The night line should refuse the same things the day team would refuse, and write down that it did.

FAQ

What is an after-hours AI agent?

An after-hours agent is a voice and chat line that can take the night’s repeatable work, write a note, and page a person when the brief says the night cannot hold the exception. It is not a recording, and it is not a chatbot that apologises until morning. The test is the digest the day team open, not how natural the voice sounded at 02

.

How does a night line know when to wake someone?

It does not guess. It follows a list you wrote with the on-call: safety, access it cannot restore, named severities, legal or medical lines, and the pages the on-call asked for. Everything else is a note. The first month will be wrong in both directions. You edit the list from the digest, the same way you would edit a junior’s night book.

Are hotel concierge agents just chatbots with a new name?

They are if they can talk and cannot write the note or page. Google Cloud’s 2026 framing uses “concierge-style” to mean agents that remember, act, and handle a 2 AM message instead of deflecting it. A hotel that ships a brochure FAQ in a voice widget has bought a chatbot. A hotel that confirms the late checkout in the reservation system and logs it has bought a night desk.

Can one agent cover voice and chat overnight?

Yes, and it should. The guest or customer should not be sorted by channel. Vesper runs voice and chat on one agent, with the same brief and the same log. Two products, two memories, two handoff styles is how exceptions get lost. The channel is an ear. The unit of work is the job.

What should a morning log contain?

Time, identity, the unit of work, what the agent did, what it refused, whether a page went out, a one-line note for the day team, and any knowledge gap the night exposed. It should be readable in one pass. A zip file of transcripts is not a log. It is the pile you were trying to leave behind.

Is after-hours coverage only for hotels?

No. Hotels make the night visible because guests are in the building. On-call SaaS and clinic intake have the same clock and the same failure — exceptions that wait until morning and arrive as a mess. The pattern is a brief, a voice, a desk, and a digest. The industry only changes the objects: a room, an incident, a slot.

What should never be automated at 2am?

Diagnosis, doses, legal commitments, unauthorised comps, production changes, unidentified access, public status edits, and any promise that a person is coming when no person has been paged. Also anything the live system cannot confirm. Warmth does not make those actions safe. A recording that refuses them is better than a fluent agent that agrees.

Do we need to replace the on-call?

No. You need to stop wasting them. The on-call is for the exception. The agent is for the rest of the night, and for the writing that means the exception arrives with a transcript instead of a rumour. A night with no on-call and a fluent agent is a different, worse product. We will not sell it.

Send the current after-hours path — the recording, the contractor brief if you have one, and a week of what waited until morning. We will write back with what an agent should take, what should still page, and what the digest should contain. That is the start of the night line, and it is how we prefer to be reached.