Skip to content

1of three

Appointment confirmation texts when someone else must say yes

Appointment confirmation texts for jobs other people must agree to: three wordings to copy, what an AI scheduling assistant can really do when the calendars it needs belong to other people, why the standards already have a word for an appointment nobody has agreed to, why moving the time throws every yes away, and four tests before buying one.

Excelsior Mind23 min readUpdated

An appointment confirmation text should wait for every yes it needs, and this page gives three to copy. A job that needs someone else's yes can be booked on your calendar and still not be agreed by the people it depends on. We help local business owners set up an AI scheduling assistant that calls a time confirmed only after everyone who must agree has replied. Below is what such an assistant can and cannot do across calendars it does not own, four tests before buying one, whether the job is worth automating, and a three-week plan.

An AI scheduling assistant is software that proposes times, tracks who has answered and holds your calendar while it waits; on calendars it does not own, it can ask, and only the other person's reply can confirm. This page is for owners whose appointments need people outside the business: a contractor working in someone else's building, a dental office booking an interpreter, a salon booking a visiting stylist, a real estate office arranging a showing on another agent's listing.

An invented example, from a plumbing company. The repair was for eleven on Saturday, in an apartment in a managed building. Three people had to be willing: the owner of the apartment, who called; the building's manager, who controls access; and the tenant living there, who is entitled to notice. Two of them had not been asked yet.

The request went out on Thursday afternoon. That evening the owner heard that Saturday at eleven was confirmed. Nobody had objected, and the time sat in bold on the calendar like every appointment that has ever happened.

At twenty past nine on Saturday the manager replied: the tenant needs notice, and the super is off until one. The plumber was already in the van. The appointment had been booked, by one of the three people whose agreement it needed.

In short

  • A scheduling assistant ran for five months at Microsoft Research. It handled 39% of its requests inside its structured workflow, and needed a trained person for the other 61%. The three commonest reasons for calling that person in were all the same reason. An attendee answered in a way the system did not expect. An attendee could not take any of the offered times. Or an attendee never replied at all.
  • The calendar standards already have a word for an appointment nobody has agreed to. It is the value every invitation starts life in. The only thing that moves it is a reply arriving. Not time passing, and not the absence of an objection.
  • And moving the time is not free. The standard that governs how calendar servers do scheduling requires this. When the start time changes, every attendee's answer is thrown away and set back to unanswered. A reschedule does not carry the agreement with it. It asks for it again.

What should an appointment confirmation text say when someone else must agree?

An appointment confirmation text, when someone else must agree, says only what has been agreed so far: a request while any yes is missing, a confirmed time once every yes is in, and a new question whenever the time moves. Three texts from the plumbing company's invented example show the order, and a person approves each one before it reaches the customer.

  1. While the building's manager and the tenant have not replied: "Your repair request for Saturday has gone to the building's manager and the tenant. We will confirm a time once both have agreed, and by Friday at six either way."
  2. Once both have said yes to eleven, the time the owner asked for: "Your repair is confirmed for Saturday at eleven. The building's manager and the tenant have both agreed to that time."
  3. If the time has to move: "Saturday at eleven no longer works for the building. Would two o'clock suit you? If it does, we will ask the manager and the tenant about two and confirm once all of you have agreed."

The first names no time, so nothing has to be taken back. The second names who agreed, so the customer can see why the time will hold. The third treats the moved time as a new question, because every yes was given to eleven, as "Why does moving the time reset every yes in scheduling?" explains below.

What a scheduling problem is, once the calendar is not yours

Three things this is not

The hard part is not the booking.

Not whether they turn up

Getting someone from an inquiry into a time they have written down, and getting them to be there on the day, is a real problem with real evidence behind it. It is written up on its own. That question has two people in it. Only one calendar has to be true. This one starts at the point where a time is agreed. It asks a different question: agreed by whom.

Not the chain of steps

Wiring your systems so that finishing one thing starts the next is a different product again. Every step in that chain is yours. Here the important steps happen in an office that is not yours, on a calendar you cannot read. It is run by someone who owes you nothing and has a full week of their own.

Not the phone being answered

Answering fast matters and it is the smallest part of this. A reply inside ten seconds that proposes eleven on Saturday has not scheduled anything. It has made an offer on behalf of three people, two of whom have not been asked yet. The speed of it is what makes the offer sound like a fact.

A scheduling problem is a calendar problem only while the calendar is yours. If the only two diaries that must agree are yours and your customer's, a booking link can do the job, and this page has little to add.

This page is about an appointment that depends on someone else's customer, building or diary. A repair inside a managed building. An inspector inside an occupied home. A dental appointment that needs an interpreter. A showing on another agent's listing. None of those exists until people you may never have met agree to it, and none of them has given you their calendar.

That changes what the software does. With your calendar alone, it reads a fact and writes one: you are free at two, so two is taken. With other people, it sends a message and waits. A question can fail in several ways, and from your side every one of them looks like nothing happening.

So the unit here is the agreement you do not have yet. For every appointment that involves outsiders, two lists help from today: agreed, and asked.

The part nobody automates is the part that is other people

The part nobody automates is the other side of the conversation, and one study counted it. From April to August 2016 a research team ran a scheduling assistant for real subscribers, who copied it into their email threads; it proposed times, negotiated with invitees and booked the meeting. The 2017 paper reports 178 participants, 1,981 invitees, 1,626 meetings and 15,659 emails, and it publishes why the system had to hand a request to a trained person. It is the oldest source on this page, kept because it breaks those reasons down.

According to the paper, the three commonest reasons were an attendee replying in a way the system did not expect, no offered time suiting everyone, and an attendee never replying. They were 32, 27 and 26 percent of the escalations, and all three sit with the other people.

The evidence

Why a scheduling agent had to hand the meeting to a person

Why a scheduling agent had to hand the meeting to a personAn attendee replied in a way it did not expect: 32%. No offered time worked for everybody: 27%. An attendee never replied at all: 26%. It could not read the organiser's own calendar: 7%. Four of the eight reasons recorded when a request escalated from the automated workflow to a trained human, across a five month deployment in which 178 people scheduled 1,626 meetings by email. The first three are things an attendee did. The fourth is the only row on the published table that is about the organiser's own systems. A request can carry more than one reason at once. So the eight rows sum to more than a hundred.An attendee replied in a way it did not expect32%No offered time worked for everybody27%An attendee never replied at all26%It could not read the organiser's own calendar7%

Four of the eight reasons recorded when a request escalated from the automated workflow to a trained human, across a five month deployment in which 178 people scheduled 1,626 meetings by email. The first three are things an attendee did. The fourth is the only row on the published table that is about the organiser's own systems. A request can carry more than one reason at once. So the eight rows sum to more than a hundred. Source: Cranshaw, Elwany, Newman, Kocielnik, Yu, Soni, Teevan and Monroy-Hernandez, Calendar.help: Designing a Workflow-Based Scheduling Agent with Humans in the Loop, CHI 2017, Table 3.

These are meetings between office workers, not jobs in other people's buildings. The assistant worked by email rather than by text. What carries is the shape, and the shape is uncomfortable for anybody selling this category. The failures cluster almost entirely on the side of the table the vendor does not control. The failure the demonstrations are about, reading the organiser's calendar, is the smallest bar on the chart. Four other rows are not drawn. One is an uninstrumented bucket at 14%. Three at 8%, 7% and 2% are the system escalating internally, rather than anything a party to the meeting did. One thing the table cannot tell you is how many of those requests were recoverable. An escalation means a trained person picked it up, not that the meeting failed. The paper is about how much skilled attention a scheduling agent consumes, rather than about how many appointments it loses.

According to the same paper, 84 percent of the requests were meetings between two people, and 15 percent had three or more attendees, eleven at most. That was the easier version. A job inside an occupied home can start at three or four parties.

The evidence

How 1,626 real scheduling requests actually finished

How 1,626 real scheduling requests actually finishedFinished without the expert ever being called in: 39%. Needed a trained person at some point: 61%. Shares of the 1,626 meetings scheduled during the five month deployment. The first bar is the share handled entirely inside the system's structured workflow. The authors are careful to say that this includes small pieces of work done by non-expert people, as well as fully automated steps. The second is the share that reached a trained scheduling worker at some point in its life.Finished without the expert ever being called in39%Needed a trained person at some point61%

Shares of the 1,626 meetings scheduled during the five month deployment. The first bar is the share handled entirely inside the system's structured workflow. The authors are careful to say that this includes small pieces of work done by non-expert people, as well as fully automated steps. The second is the share that reached a trained scheduling worker at some point in its life. Source: Cranshaw et al., Calendar.help, CHI 2017, System Efficiency: 39% of requests completed entirely within the microtasking workflows, the remaining 61% requiring some intervention.

Two things stop this being a verdict on automation in general. The first is that it is a snapshot of one system at one point in its life, rather than a ceiling. The same paper reports that the earliest version of the service ran entirely on skilled humans. So this 39% is a position on a curve the authors were deliberately pushing. And they say plainly that getting there took continual observation of how people actually replied. The second is that nobody has run anything like this on property appointments. So the split above is evidence about the shape of the problem, and not a forecast for your Saturdays.

The system ran in tiers: software first, then a person doing one small, tightly defined task. The paper's 39 percent covers both tiers; it is the share of requests that never needed the expensive scheduler. That is a good result, and people were still in it.

A current figure points the same way. A 2026 paper on calendar conflict decisions, PEARL (revised 2026-06-23), reports that one 30-billion-parameter reasoning model in its experiments had an average error rate of 35 percent. Its experiments test conflict decisions, not your appointments, so it gives no rate for them; it does suggest a person should still own the decision. Before trusting any tool, it is worth asking where a reply it cannot read goes, and who sees it.

Who must say yes before an AI scheduling assistant can confirm?

How hard an appointment is to arrange tracks how many separate permissions it needs, and that count is rarely written down. They are not all people, and they fail in different ways. Three of the four below can say no to you. The fourth cannot say anything at all, which is how it stays off the list.

Before eleven on Saturday is real

Four permissions, and you hold one of them.

The person you are meeting

The only one everybody thinks about. The only one whose calendar the conversation can actually negotiate with, because they are in the conversation. If this were the whole problem then a booking link would have solved it years ago. For an appointment at your own shop it very nearly is.

Whoever controls the place

At your own shop that is you. That is why your own appointments feel easy and other people's do not. In someone else's building it is a manager with a full week. Behind that manager is an owner who has to approve the work, or a tenant who has rights about being given notice. None of those three is in your conversation. None of them owes you a reply this afternoon.

The way in

A code, a key, a super, an alarm, a doorman, a gate. This one is not a person, which is why it is the one left off the list. It silently carries a time window of its own. A code that works between ten and four is a constraint on the appointment. It constrains it exactly as much as an owner who works from home does.

The professionals, on a bigger job

A bigger job is the same problem with the volume turned up. An inspector, an adjuster, the owner and the contractor may all have to be on site in the same hour, and a real estate closing puts an attorney, a lender and a title company in one room. Every one of them is running the same week you are. These are the appointments where the cost of a wrong assumption is not an empty morning.

A long rank of tall painted signal levers in a railway signal box, receding away from the camera, their paint muted almost to grey, most of them a dark dusky tone with paler levers and black-and-white chequered levers standing among them, each with a curved catch handle at the top and a small numbered oval plate on its shaft, a row of brass instruments and dials mounted on the frame above them, a black lever bed and a plank floor below, and windows along the right showing track outside
A mechanical interlocking is the oldest working answer to this problem. Every lever in the frame is physically prevented from moving until the levers it conflicts with are set correctly, so a signalman cannot offer two trains the same piece of track even by accident. The bars are cut so that a wrong combination is impossible rather than discouraged, which is a higher standard than any calendar has ever been held to. Photograph by Steve Knight, CC BY 2.0, colour reduced.

An invented example: a contractor's site visit needs the owner, the contractor, the tenant living there and the building's rule on access hours. The rule never replies; it simply makes some times impossible. Beside any time you offer, it helps to write how many permissions it needs.

How does a scheduling standard name an appointment nobody agreed to?

The standard already has a word for an appointment nobody agreed to, and the word is NEEDS-ACTION. What your calendar does when it sends an invitation is defined in iTIP, RFC 5546, a standard in force. It splits everyone into an organizer, who owns the event, and attendees, who cannot change the master copy but can reply. The reply carries one value, PARTSTAT, short for participation status.

Section 2.1.1 says that when the organizer issues the first invitation, attendee status is typically unknown, so the organizer sets it to NEEDS-ACTION, and each attendee changes their own status in a reply sent back. Only a reply from that person moves it. Silence does not, and neither does bold type on your calendar.

The four answers

Every invitation is one of these, and one of them is silence.

Nobody has answered

The state every invitation is born in, and the standard names it out loud. When the organiser sends the thing out, the answer is set to needs action, because nothing has come back. This is not a failure state and it is not an error. It is the honest description of a proposal. It is where a great many appointments are sitting at the moment someone is told they are confirmed.

They said yes

Someone on the other end took a deliberate action. A reply came back to the organiser saying so. The important part is not the word yes. It is that it arrived as a message rather than as an absence. A reply is a thing that happened. Silence is a thing that did not happen. The two are only the same if you decide they are.

They said no

The cheapest outcome on this list, and the one everybody dreads. A no on Thursday costs a message. The same no at twenty past nine on Saturday costs a morning and a drive. It also costs part of a client's confidence. That is the part that comes from believing you when you tell them something is settled.

They said maybe

The standard carries a state for tentative, kept distinct from yes and from no. A system can hold a maybe there without rounding it up. One that does is telling you something true about your Saturday. A build with nowhere to put an answer that is not a decision will file it as one of the other two. It will not be the cautious one.

The same standard names what the manager in the story was doing: COUNTER, an attendee's way to negotiate a change, such as a different time. A counter is neither a yes nor a no, and a system with no place for it will file it as one of the two. An invented example: a cleaning firm proposes Tuesday at ten and the landlord texts back "cant do mornings thursday?"; that is a counter. A third column for counters helps, even on a sheet of paper.

Why does moving the time reset every yes in scheduling?

Moving the time throws away every yes you had, and the standard says so. RFC 6638 defines how a calendar server schedules on your behalf, and section 3.2.8 sets out what the server must do when an appointment moves.

Servers MUST reset the PARTSTAT property parameter value of all ATTENDEE properties, except the one that corresponds to the Organizer, to NEEDS-ACTION for each calendar component change that causes any instance to be rescheduled.

In plain words: if the start time, end time or duration changes, every attendee's answer goes back to unanswered on every affected occurrence, for everyone except the organizer. Nobody agreed to Saturday at two. They agreed to Saturday at eleven, and that time no longer exists.

That holds with or without software. A system that moves the time and keeps the old confirmations is keeping something the standard has cleared. An invented example: a dental office moves a patient and an interpreter from nine to half past nine and leaves both marked confirmed; the interpreter, who agreed to nine, is booked elsewhere. A moved time goes to everyone as a new question. The same holds for a real estate office: a showing moved from eleven to two needs the listing side to agree again, even if they agreed to eleven an hour earlier.

Your own system knows whether the message arrived

Your own system can know whether the message arrived. The CalDAV scheduling specification, RFC 6638, attaches a delivery status to each attendee, saying what happened to the message the server sent them. It publishes eight codes, because a message can fail to reach a calendar in several ways, and they collapse into three states a person can act on.

What your own system knows

Eight answers to did it arrive, and only one of them is yes.

It has not gone yet

One of the eight published codes means the server is still trying. That is a perfectly ordinary state to be in for a few seconds. It is a very bad one to be in for a day. The only way anybody finds out which it was is by looking. Nothing about a pending message looks different, from your side, from a message that landed.

It went, and nobody knows if it arrived

Two codes cover sent. One says delivered. The other says sent. It then says, in as many words, that the server has no explicit information about whether it was delivered. That is the ordinary case whenever the invitation travelled by email. That distinction is written into an internet standard because it is real. It is the difference between a confirmation and a hope.

It failed, in one of five ways

The other five codes are all failures. They are separated because the right response differs. The address was not a calendar user at all. You did not have permission. The delivery could not be completed this time. No route existed. Or scheduling with that recipient is not allowed. Two of the five are worth retrying and three are not. A system that treats them the same will quietly keep sending into a wall.

None of the eight, the good one included, says the person is coming. Delivered is a fact about a server; accepted is a fact about a person. An invented example: a contractor sees "delivered" beside a homeowner's name and treats the visit as agreed, though the homeowner never opened it. Delivered and accepted belong in two columns, both checked before a customer is told anything.

What an AI scheduling assistant can honestly do across calendars it does not own

An AI scheduling assistant can honestly do four things across calendars it does not own.

It can propose without promising. The first message to your customer either names a time or says a request has gone in, and those two messages lead to very different Saturdays, as the first two texts at the top of this page show.

It can hold your side properly. Your calendar is the one thing in the exchange your software controls, so a live proposal should block it and a dead one should release it. Blocking without releasing fills a diary with appointments that never happened.

It can chase, then stop: one second message, then a decision handed to a person. That hand-off is a feature.

And it can ask again when the time moves, for the reason the standard gives.

One more thing already exists that demonstrations rarely show. A major calendar platform's meeting-time call suggests times from organizer and attendee availability, and each suggestion carries a confidence percentage for the likelihood of all attendees attending; a matching input sets the minimum confidence for a slot to be returned at all. A tool built on top can drop that figure and print a bare time. The same reference describes that minimum as a required confidence from 0 to 100, so the platform itself treats a suggestion as a likelihood rather than a fact.

The mechanic

Thursday, half past four.

A staged illustration, not a real exchange. No address, name or number appears in it because none of them would be real. The sequence is the one the service page describes.

The exchange

  1. The owner of 4C has asked us to fix a leak under the kitchen sink. Saturday 11am or Saturday 2pm both work our side. Either possible?

    The assistant / 4:31 pm

  2. not 11, tenant needs 24h notice and super is off til 1. 2 is probably ok, checking with them

    The building manager / 6:02 pm

  3. Understood. Holding 2pm Saturday as unconfirmed on our side and not offering it elsewhere. Tell us either way by Friday 6pm and we will confirm to the owner then.

    The assistant / 6:02 pm

  4. tenant is fine with 2. super will meet you in the lobby

    The building manager / Fri 11:14 am

  5. Confirmed. Saturday 2pm. Sending an invitation to this number so it lands on your calendar too.

    The assistant / 11:14 am

What happened on your side

  1. Two times proposed, nothing promised

    The owner is told a request has gone in. The owner is not told a time. Nobody has agreed to anything yet, and the message the owner receives says exactly that.

    4:31 pm
  2. The counter is recorded as a counter

    Not as a rejection and not as a booking. One of the two proposals is dead. The other is alive and unanswered, and the difference is written down rather than remembered.

    6:02 pm
  3. Your own side is held

    The 2pm is blocked on your calendar the moment it becomes the live proposal. Nothing else can be offered into it while you wait. This is the only calendar in the whole exchange your software can actually control.

    6:02 pm
  4. The owner is told, once

    Nineteen hours after the request went in. It is the first time the owner has been given a time. That delay is the honest cost of not having promised anything on Thursday.

    Fri 11:14 am

The system

From somebody asking to everybody agreeing.

Six hops. The products in this category are sold on the first and the last. The middle four are the ones that decide whether a Saturday happens. Every one of them is a question about people who do not work for you. Note where the hold sits. It is the only step in the chain your software can carry out with any authority.

The path from a request to view a property through the access constraints, a proposal of two times, the replies from the other side, a hold on your own calendar, and one confirmation: The ask, The constraints, The proposal, The replies, The hold, The confirmationThe askSomebody wants inThe constraintsAccess hours, noticeThe proposalTwo times, no promiseThe repliesYes, no, or nothingThe holdYour side onlyThe confirmationOnce, and to everyone

Scroll to follow the chain

  1. The ask: Somebody wants in
  2. The constraints: Access hours, notice
  3. The proposal: Two times, no promise
  4. The replies: Yes, no, or nothing
  5. The hold: Your side only
  6. The confirmation: Once, and to everyone

An invented example: a tutoring agency sees two suggested evenings, one far likelier than the other, and offers the likelier one first. Any tool worth buying can show that confidence beside each time.

Why is an unheld slot worse than no slot at all?

For an AI scheduling assistant, offering a slot you cannot hold is worse than offering nothing, because the two failures differ in size. When a system feels slow, the instinct is to make it more decisive. Here the instinct is wrong.

In an invented example, a slot you did not offer costs a message. The customer waits until Friday morning, then hears a time, and loses a day of not knowing. Saying on Thursday that you are waiting on the other side covers that.

In the same invented example, a slot you offered and then withdrew costs something a message cannot buy back. The customer arranged their Saturday around it and may have turned down other plans. What they learn is that when you say "confirmed", it may or may not hold.

A demonstration tends to flatter the faster setup: a quick answer with a time looks better on a screen recording than a quick answer with a request. One is quicker and the other is accurate.

In your numbers

How many agreements are you carrying that nobody actually gave you?

Everything you have to set up with someone outside your own business: jobs in other people's buildings, inspections, site visits, interpreters, showings.

20

People who have to agree, other than you

Count the ones who can stop it happening. The person you are meeting is one. The building's manager is another. An occupant with a notice period is a third. On a bigger job it is more than that.

Be honest rather than aspirational. A voice message you did not keep. A nod at an open house. A thread that just went quiet. All of them count as no reply.

40%

The message. The second message. The call. And the two minutes afterwards, working out what you are now going to tell the client.

4

Agreements a year that were never actually given

192unconfirmed

Appointments a monthyour 20
20a month
Over a year12 months
240appointments
Agreements they need from other people2 other parties
480yeses
Where no yes ever came backyour 40%
192unconfirmed
At your chasing timeyour 4
768minutes
In hours60 minutes in an hour
13hours a year

The headline is the fourth row rather than the hours. The hours row underneath it is doing deliberate work. At the settings this opens with, the chasing comes to a number of hours that any owner would shrug at. That is the trap. The cost of this is not the time. It is that a share of those unconfirmed agreements are being described to a client as confirmed. Shares produce fractions, and two thirds of an agreement is not a thing. So read anything with a decimal in it as a rough count. Four things this deliberately refuses. There is no no-show rate anywhere in it, because whether somebody turns up is a different article on this site with real evidence behind it and this one has nothing to add to it. There is no figure for how often the other side declines a request, and that gap was looked for rather than assumed: we could not find one published. There is no money row, because the value of a Saturday morning depends on what was in it. And there is no second column showing what this becomes with a scheduling layer switched on, because we have not measured that on your appointments and the study on this page was run on office meetings between two people.

An invented example: a salon tells a bridal party their trial is booked before the visiting stylist has replied, then cancels; on the next booking the party asks twice whether it is real. Until every yes is in, the word for the customer is "requested".

How do you test an AI scheduling assistant before you buy it?

To test an AI scheduling assistant before you buy it, ask four questions; none of them is technical.

Ask to see the first text or email the person on the other end receives before anyone has agreed. If that first message carries a bare time with no qualification, that is the product, and no later setting undoes it.

Ask what happens to your calendar while a proposal is open, and then what happens when it dies. A system that blocks and never releases tells people you are busy on days you are free.

Ask what it does with a counter: a text in lower case, no punctuation, saying eleven is impossible but two might work. In the study above, an unexpected reply was the largest single reason the agent called in a person. Watch where it lands.

Ask how you find out that an invitation was never delivered. The standard has a real answer, so a vendor who built on a calendar server will recognize the question.

The honest read

Tell us how a job in somebody else's building gets set up in your business today, from the message arriving to the customer being told a time. We will send back where the unconfirmed gap is in your version of it, and which single message is the one that turns a proposal into a promise.

Send us your version

It is a short reply from a person, it costs nothing, we do not need access to your calendar to answer it, and the answer is useful whether you buy anything or not.

A row of six blue framed electronic departure screens mounted under a station roof, four of them headed Departures in white on blue with a yellow departure time and destination beneath, Wick, Edinburgh, Aberdeen and Kyle of Lochalsh, each of those four listing its calling points in small yellow type and carrying the operator name First ScotRail along the bottom, a fifth headed Subsequent Departures showing Dingwall and Aberdeen and carrying no operator name, a sixth cut off at the right edge headed Informat, and the words On time set beside every time on the board
Every time on this board was decided by somebody who is not standing in front of it, and the board's whole job is to say which ones are still true. On time is a claim being made now about a plan made months ago, and it is republished the moment it stops being true. That is the standard a scheduling system should be held to, and almost none of them are: they tell you a time once and then go quiet. Photograph by David Jones, CC BY 2.0.

An invented example: an owner types "maybe thurs" as the reply during a demonstration and watches the tool mark it accepted; that one answer settles the purchase. These four questions, written out, belong in every demonstration.

What does it take to run an AI scheduling assistant?

What decides the size of the job is how many kinds of people you have to reach and how each of them likes to be reached, more than the software.

An invented example: one contractor arranges visits with the same few building managers by text, and every one replies the same day. Another deals with a rotating cast: a manager who reads only email, a management company with a portal, two owners who want a phone call. Those are different projects, and the difference is in the reaching.

The work is in the decisions. Your week splits into clicks, the steps a system can repeat (sending, chasing, blocking, releasing), and decisions, the judgement you are paid for: which appointment types you run, the access rule on each, how long you hold your calendar waiting for an answer, and what happens when that wait runs out. Written down as defaults the system reads every time, they are worth having even if no software ever arrives.

A system can report who has confirmed only if someone put the confirmation in, so the people on your side have to forward the text instead of remembering it. Four defaults on one page is the week's job.

What does an AI scheduling assistant not do?

An AI scheduling assistant does not make anyone reply. The three largest reasons in the 2016 study were an unexpected reply, no time suiting everyone, and no reply at all, all on the other side, and nothing on your side changes what happens on theirs.

It does not read a calendar it has not been given. Inside one calendar system it may see the availability of people who share theirs, as the meeting-time call above does; for everyone outside it, a calendar is a message. A product that promises live availability across every party is describing the ones who shared.

It does not decide which appointments are worth having. Faster setup can mean more appointments, and with nothing between the request and the calendar that is a fuller week, not necessarily a better one.

It does not know an access rule, such as a tenant's notice period, unless someone told it. A tool without the rule keeps proposing eleven on Saturday until a person types it in.

And it inherits whatever your diary already gets wrong. Availability that lives only in someone's head is invisible to it, so proposals get withdrawn in front of a customer instead of quietly.

Three ways a working build produces nothing

None of them are the calendar.

A proposal is presented as a booking

The single most expensive thing in this whole category is not a setting. It is the wording of one automatic message. If the note that goes out when a request is raised says confirmed, then every unanswered request in your system is a promise, and you will only find out which ones were real on the morning.

The hold is never released

Blocking your own calendar while you wait for somebody else is correct, and a hold that nothing ever clears is a calendar that fills up with appointments that did not happen. Within a month the machine is reading a diary that is busier than your life, and offering nothing, politely.

The chase has no end

A request that has been sitting unanswered for two days needs a decision from a person, not a fourth reminder to somebody who has already ignored three. If nothing in the build ever says this one is not going to happen, the queue grows and the client who is waiting hears nothing at all, which is the one outcome worse than a no.

An invented example: a gym connects a tool while the owner still keeps her own lunch hour only in her head, and the tool offers that hour to clients. The gaps in your own calendar come first, before anything is connected to it.

Is an AI scheduling assistant worth it for your business?

Whether an AI scheduling assistant is worth it depends on three questions, and "leave it alone" is a real answer. These are the screening questions of our AI assessment, our own method rather than a published standard, run on one appointment type. Does it happen often enough that a failure would be noticed within a week? Can the rule, access rules included, be written down for a new hire? Does a wrong answer land where a person sees it before the customer does?

Then the stop question, which asks what breaks if the job stopped: if nothing breaks, stop doing it, and do not automate it. Then check whether a tool you already pay for does it; your calendar may already send, chase and hold.

Score impact and risk from 1 to 5. Four checks raise the risk: it binds the business, reaches a customer, feeds another system, or can damage trust. A confirmation to a customer reaches a customer and can damage trust, so a person approves it before it goes. An impact of 1 or 2 means leave it alone, because a rare job costs less to do by hand than to set up, watch and repair.

An invented example: a picture framer books a handful of consultations a month, each with one customer at the shop. Impact is low and nobody outside must agree, so a booking link covers it; leave it alone. The most frequent appointment type is the one to run the three questions on first.

Before anything is set up

Three questions, then the stop question

Often enough?

Frequent enough that a failure would show within a week.

Can the rule be written?

Access rules included, on one page a new hire could follow.

Seen first?

A wrong confirmation should reach a person on your side before it reaches the customer.

What breaks if it stopped?

If nothing breaks, stop doing it. An impact of 1 or 2 out of 5 means leave it alone.

Who an AI scheduling assistant fits, and who it does not

An AI scheduling assistant fits owners whose appointments need a yes from outside the business: contractors and inspectors working in occupied homes and managed buildings, dental offices booking interpreters, venues coordinating suppliers, real estate offices arranging visits on other people's listings. It fits less where only you and the customer must agree, such as a haircut at your own chair; a booking link does that job. It does not fit appointments where a wrong time costs a contract or a legal deadline, such as a real estate closing: a person arranges those by hand, with the two columns described below.

Who this is for

A good fit, a poor fit, and what to do instead

A good fit

Appointments that need a yes from outside the business: jobs in managed buildings and occupied homes, interpreters, visiting suppliers, showings on other people's listings.

A poor fit

Bookings with two calendars and no third party, like a haircut at your own chair.

Instead, for a poor fit

A booking link does the job.

Keep it by hand

Where a wrong time costs a contract or a legal deadline, such as a real estate closing: a person arranges it, with two columns, agreed and asked.

Common questions, answered honestly

What is AI scheduling, in plain terms?

AI scheduling is software that takes a request for an appointment, proposes times to the people whose agreement it needs, tracks who has replied, blocks your calendar while it waits and tells everyone once it is settled. The intelligent part is narrow: reading a request that arrives as three lines of lower case text, and keeping one negotiation straight across several threads. The rest is record keeping.

How is this different from AI appointment booking?

Booking is about getting one person from interested to a time they have written down, and whether they turn up; that is written up separately on this site. Scheduling, as used here, starts once a time is proposed and asks who has agreed. The two overlap when only you and the customer are involved, and come apart when an appointment needs a permission from someone outside the conversation.

Can it stop a double booking?

An AI scheduling assistant is built against one kind of double booking and not the other. Because it can read and write your own calendar, holding a slot the moment a proposal goes live is what guards against your own time being offered twice, if the hold is set up and working. It cannot stop the other side promising the same two o'clock to someone else, because it has no view of their diary. A product that says double booking cannot happen is describing the first kind.

What happens when somebody replies "maybe"?

A "maybe" belongs in a tentative state, which the calendar standards carry beside yes and no. A system can hold a proposal there without rounding it up. The thing to check is whether the product has a place for an answer that is not a decision.

Does moving an appointment need everybody to confirm again?

Yes, and that is the standard's rule, not ours. The specification for calendar servers resets every attendee's participation status to needs action when the start, end or duration changes. The agreement was to a time; change the time and only the appearance of agreement is left.

How does it know when I am free?

An AI scheduling assistant reads your calendar, which is the ordinary part. Two more things are worth checking: what level of access it asks for, covered in the appointment booking post, and whether it may write as well as read. Reading alone reserves nothing, and a proposal that has not been reserved is still open to whoever asks next.

What if the other side never answers at all?

A person has to decide, and the useful questions are when and who. A sound setup sends the request, sends one chase, then puts it in front of someone with the whole history attached. A system that chases forever lets the queue grow quietly while your customer hears nothing.

What should the first message to the customer say?

The first message should say that the request has gone in, who it is waiting on, and when you will be back to them. An invented example, for the plumbing company: "Your repair request for Saturday has gone to the building's manager and the tenant. We will confirm a time once both have agreed, and by Friday at six either way." It names no time, so nothing has to be taken back.

Who gets the chase, and when?

The person whose reply is missing, once, at a time set in your defaults; after that a person on your side decides. One chase keeps the request alive without turning it into pressure, and the decision after it is the hand-off described in "What an AI scheduling assistant can honestly do across calendars it does not own".

Is any of this different for a real estate closing?

A closing is the same problem with more parties and a much higher cost of being wrong, and nothing here suggests automating it. What carries over is the discipline: know which people have confirmed in writing, and treat a moved date as a new question.

Can we talk about scheduling on a call if my business is outside New York?

We are based in New York, and the call works for an owner anywhere.

Ten questions

What did you learn? Check it in two minutes

Ten questions on what this page taught. It takes about two minutes.

Take the quiz
  1. 1 of 10

    What moves an invitation's participation status out of NEEDS-ACTION?

    • A.A reply from that attendee
    • B.The appointment time arriving
    • C.A day passing with no objection
    • D.The organizer marking it in bold
    Show the answer

    A. A reply from that attendee Only a reply from that person moves it; see "How does a scheduling standard name an appointment nobody agreed to?".

  2. 2 of 10

    What happens to attendees' answers when an appointment's start time changes?

    • A.They carry over unchanged
    • B.They reset to needs action
    • C.Only the organizer's answer resets
    • D.They turn tentative
    Show the answer

    B. They reset to needs action A new start, end or duration resets every attendee; see "Why does moving the time reset every yes in scheduling?".

  3. 3 of 10

    In the 2016 study, where did the three commonest escalation reasons sit?

    • A.In the calendar software
    • B.With the other people in the conversation
    • C.In the subscriber's own calendar
    • D.In message delivery
    Show the answer

    B. With the other people in the conversation An unexpected reply, no time suiting everyone, and no reply at all; see "The part nobody automates is the part that is other people".

  4. 4 of 10

    What does the study's 39 percent measure?

    • A.Requests no person touched
    • B.Requests that never needed the skilled scheduler
    • C.Meetings with three or more attendees
    • D.Invitees who never replied
    Show the answer

    B. Requests that never needed the skilled scheduler It covers both tiers, software and a person on one small task; see "The part nobody automates is the part that is other people".

  5. 5 of 10

    What does a "delivered" status tell you?

    • A.The person is coming
    • B.The person read the message
    • C.The person accepted
    • D.The server got the message
    Show the answer

    D. The server got the message Delivered is a fact about a server, accepted a fact about a person; see "Your own system knows whether the message arrived".

  6. 6 of 10

    What is a counter?

    • A.A rejection
    • B.An acceptance
    • C.A request to change the appointment, such as a new time
    • D.A failed delivery
    Show the answer

    C. A request to change the appointment, such as a new time A counter is neither a yes nor a no; see "How does a scheduling standard name an appointment nobody agreed to?".

  7. 7 of 10

    Why is offering a slot you cannot hold worse than offering nothing?

    • A.It sends more messages
    • B.The customer learns your "confirmed" may not hold
    • C.It breaks the calendar standard
    • D.It slows the system down
    Show the answer

    B. The customer learns your "confirmed" may not hold A withdrawn slot costs something a message cannot buy back; see "Why is an unheld slot worse than no slot at all?".

  8. 8 of 10

    In the worth-automating test, what does an impact of 1 or 2 mean?

    • A.Leave it alone
    • B.Automate it first
    • C.Buy a new tool
    • D.Score it again next month
    Show the answer

    A. Leave it alone A rare job costs less to do by hand than to set up, watch and repair; see "Is an AI scheduling assistant worth it for your business?".

  9. 9 of 10

    Which of these is a decision, not a click?

    • A.Sending the request
    • B.Sending the chase
    • C.How long to hold your calendar waiting for an answer
    • D.Releasing a dead slot
    Show the answer

    C. How long to hold your calendar waiting for an answer Sending, chasing, blocking and releasing are clicks; the hold is a decision; see "What does it take to run an AI scheduling assistant?".

  10. 10 of 10

    What is the first step of the roadmap?

    • A.Buy a scheduling tool
    • B.Automate every appointment type
    • C.Pick the appointment type with the biggest gap and write its defaults
    • D.Ask other businesses for access to their calendars
    Show the answer

    C. Pick the appointment type with the biggest gap and write its defaults One appointment type, never the whole calendar; see "What to do first with an AI scheduling assistant".

What to do about scheduling that needs a yes from outside

The check needs nothing but your own calendar for the next two weeks. For every appointment that needs someone outside your business, one column holds how many people had to agree, and a second holds how many of those agreements you could produce if someone asked: not remember, produce, as a message with a time on it.

The gap between the two columns is your exposure, in your own hand, and it doubles as the list of calls worth making the next morning. A business with no gap is already doing the hard part by hand and has nothing to buy. A business with a gap has found the Saturday that may go wrong, with a week left to stop it.

An invented example: a dental office lists eleven appointments for the next two weeks, four of them needing an interpreter or a parent's yes. It can produce written agreements for two of the four, and the other two are the next morning's calls.

Open your calendar and find the next appointment that needs somebody outside your business. Then answer one question about it out loud: which of the people it depends on has actually replied, in writing, and where is that reply. If the answer is that you are fairly sure, you have found the thing this whole article is about, and you have found it while there is still time to send a message.

There is no price here because two things move it and neither is the calendar connection: how many parties a typical appointment of yours needs, and whether the people on the other side can be reached the same way every time or whether half of them are a phone call. The AI assessment is an hour, done with you, and for this topic it starts by walking one real appointment backwards from the confirmation to the first message.

What to do first with an AI scheduling assistant

The first step is one appointment type, never the whole calendar. Here is the order, week by week.

This week: the two-column check above, and the appointment type with the biggest gap. Its four defaults get written: the access rule, how long you will hold the slot, when a chase goes out, and who decides when the wait runs out.

Next week: the job written down so someone else could run it. When we run the AI readiness assessment, the interview asks one thing of every job: one job written in two minutes under input, do, output and the person's check, handed to a colleague who does not know it, noting where they stall. Where the colleague stalls is a decision nobody has written down yet. Then the three questions from "Is an AI scheduling assistant worth it for your business?" go to the job.

The week after: the decision, written down. Leave it alone, keep it by hand with the two columns, or take the four tests to a demonstration. Whatever the choice, a person approves every confirmation that reaches a customer.

An invented example: a home inspector picks inspections in occupied homes, writes the defaults on a card, hands the card to a part-time assistant, and finds the assistant stalls at "what counts as notice"; that line becomes the first default.

A suggested schedule

One appointment type, three weeks, a decision written down

  1. This week

    The two-column check; the appointment type with the biggest gap and its four defaults.

  2. Next week

    The job written down so someone else could run it; the three questions.

  3. The week after

    Leave it alone, keep it by hand, or take the four tests to a demonstration; a person approves every confirmation.

A schedule to start from, not a deadline.

Source: our method, the AI assessment

Ask us about AI scheduling for small business.

Sources

  • Justin Cranshaw and seven co-authors, "Calendar.help: Designing a Workflow-Based Scheduling Agent with Humans in the Loop", CHI 2017, arXiv 1703.08428, data from April to August 2016, the paper. Read November 11, 2026; kept for the argument.
  • PEARL: Self-Evolving Assistant for Time Management with Reinforcement Learning, a research paper, arXiv 2601.11957, submitted January 17, 2026, revised June 23, 2026, the abstract. Read November 11, 2026.
  • RFC 5546, iCalendar Transport-Independent Interoperability Protocol (iTIP), RFC Editor, a standard in force, the text. Read November 11, 2026.
  • RFC 6638, Scheduling Extensions to CalDAV, RFC Editor, a standard in force, the text. Read November 11, 2026.
  • findMeetingTimes and the meetingTimeSuggestion resource, a calendar platform's developer reference, live documentation, the method and the resource. Read November 11, 2026.

Sources checked November 11, 2026.

Written with an AI assistant and checked by a person.

Know somebody who would argue with this? Send it to them.