Four o'clock on a Tuesday and the chair is empty. The slot was booked three weeks ago, confirmed by email, and nobody came. You cannot resell the hour, you already paid for the space, and the only thing you can do about it is decide what happens next time.
Almost every tool sold into this problem promises prediction. The evidence says the biggest wins come from something duller and cheaper: what your reminder actually says, and when it arrives.
- Two randomised trials covering roughly 20,000 hospital appointments found that stating the specific cost of a missed slot in an SMS cut non attendance by about 2.6 percentage points against a standard message.
- General statements about cost worked less well than a specific figure, and an empathy worded message performed worst of the tested variants.
- A meta analysis of ten studies put the pooled benefit of reminders at a relative risk of 1.11 for attendance, so reminders help and do not solve the problem.
- Published no show prediction models reach useful accuracy, but nothing in the research tells you what to do with a prediction, which is your decision and carries the real risk.
- Change the message before you buy the model. It costs nothing and it is the intervention with the cleanest evidence behind it.
This is written for a business where an empty slot is the product: a salon, a clinic, a restaurant, a studio, a tradesperson with a day of appointments. The research below comes from healthcare because that is where the trials were run, and the mechanism transfers.
What does a reminder actually change?
Attendance, modestly and reliably. A systematic review and meta analysis of appointment reminders, covering twelve studies from nine countries with ten contributing to the pooled analysis, found a pooled relative risk of 1.11 for attendance across 8,236 participants, meaning roughly an 11 percent improvement in the likelihood somebody turns up.
The same review put SMS reminders at a relative risk of 1.14 with wide variation between studies, and telephone reminders at 1.11 with more consistent effects. Postal letters were tested once, raising attendance from 64.7 percent to 73.0 percent in an orthodontic setting.
Read those numbers as a floor rather than a disappointment. An 11 percent improvement on a 15 percent no show rate is not a transformation of your Tuesday, but reminders cost close to nothing and the effect is real across countries and settings. What the review also says, and this is the part worth carrying, is that effectiveness varies considerably with the population and the setting, so your own numbers will not match the average.
Which words in the reminder do the work?
The ones that state what the missed slot costs, specifically. This is the most useful finding in the whole literature for a small business, because it is free to implement and it was tested properly.
Two randomised controlled trials, reported in PLOS One as Stating Appointment Costs in SMS Reminders Reduces Missed Hospital Appointments, randomly allocated appointments to different reminder messages issued five days in advance. The first trial covered 10,111 participants and the second 9,848 with valid mobile records.
In the first trial, the control message produced an 11.1 percent non attendance rate and a message stating specific costs produced 8.4 percent. The second trial ranked four variants: specific costs at 8.2 percent, a recording message at 9.6 percent, general costs at 9.9 percent, and an empathy worded message at 10.7 percent. The authors concluded that missed appointments can be reduced for no additional cost by introducing persuasive messages into reminders, and estimated the reframing alone could prevent around 5,800 missed appointments a year in their trust.
| Message type | Trial two non attendance | What it tells a small business |
|---|---|---|
| Specific cost stated | 8.2 percent | Name the actual figure, not a category |
| Recording message | 9.6 percent | Process framing helps a little |
| General cost stated | 9.9 percent | Vague cost language loses most of the effect |
| Empathy worded | 10.7 percent | Warmth alone was the weakest tested variant |
| Standard control (trial one) | 11.1 percent | Your current reminder is probably this |
Translate it to a salon: not "please let us know if you cannot make it", but "this appointment holds a 45 minute slot we cannot fill at short notice". Translate it to a restaurant: the number of covers the table represents. The mechanism is making a hidden cost visible, and the finding is that precision carries the effect.
Say what it costs you, not what you will charge them. A stated cost is information. A stated penalty is a threat, and it belongs in your booking terms rather than in a reminder two days before.
How many reminders, and when?
More than one, and the second one closer to the appointment than feels necessary. The trials above sent messages five days ahead, which suits a hospital booked months out and is far too early for a haircut booked last Thursday.
The principle to work from is that a reminder has two jobs, and they need different timing. The first job is to let somebody cancel while the slot is still resellable, which means it must arrive far enough ahead that you can fill the gap. The second job is to defeat forgetting, which means arriving close enough to the time that acting on it is immediate.
For most small businesses that produces a simple pattern. One reminder at whatever notice period you actually need to refill a slot, phrased to make cancelling easy and free. One short reminder the day before or the morning of, phrased to make attending easy: the time, the address, and what to bring. Different jobs, different words.
Making cancellation easy is counterintuitive and it is the right call. An empty slot you learn about on Tuesday morning is worth something. The same slot you learn about at four o'clock is worth nothing, and the difference is entirely whether cancelling felt harder than not turning up.
Can a model tell you who will not show?
Reasonably well, and that is the easy half. Published work on no show prediction reports strong discrimination: one framework, in a decision analysis of machine learning approaches to no shows, reports test set areas under the curve above 0.90 on both of its datasets and sensitivity above 0.94, using resampling to handle the imbalance between attended and missed appointments.
The authors are clear about the intended use: accurate predictions enable targeted intervention strategies for the patients least likely to attend, including reminders and overbooking for high risk cases. They acknowledge that the two kinds of error carry different consequences.
Which lands you exactly where the research stops and your business starts. A probability is not a decision. What you do with a customer flagged as likely to miss is a policy question about your own tolerance for two very different mistakes, and no paper can answer it for you.
Reminder timing against your own refill window
Work backwards from one number: how long it realistically takes you to fill a freed slot. That figure, not a convention, sets your reminder timing.
If your waitlist can fill a gap in two hours, an early reminder buys you little and a same day one is enough. If it takes you three days of posting and asking, the first reminder needs to be at least three days out, because a cancellation that arrives inside your refill window is functionally a no show with a polite message attached.
Most scheduling defaults are set to twenty four hours because that is what the software shipped with. It is a coincidence rather than a finding, and it is worth overriding the moment you know your own refill time.
There is a second timing question nobody asks: how far ahead you let people book at all. Customer commitment decays with distance, and a slot booked eleven weeks out is a different proposition from one booked next Tuesday. If your no shows cluster among the longest lead times, shortening your booking horizon is a one setting change that costs you nothing except the illusion of a full diary.
The two mistakes, and which one you can afford
Overbooking a slot that was going to be honoured produces a double booked customer standing in your doorway. Leaving a slot single booked that was going to be missed produces an empty hour. Both are losses and they are not remotely the same size.
For a hospital with a waiting room, a double booking means a wait. For a two chair salon, it means turning away somebody who arrived on time, and that person tells people. The asymmetry is far more brutal in a small business, which is why the overbooking strategy that makes sense at scale usually does not survive contact with a shop where the owner faces the customer.
The same asymmetry runs through fraud work, where we set out why a wrongly blocked good order costs more than the fraud it prevents. Here the cheaper response to a high risk booking is not to double book it. It is to ask for confirmation, offer an easier reschedule, or take a deposit. All three reduce the loss without creating the worse failure.
Do deposits work?
They do, and the cost is paid in bookings you never receive. A deposit converts a soft commitment into a financial one, and the mechanism is not subtle: people turn up to things they have paid for.
What is harder to see is the booking that did not happen because a card was requested. That loss never appears in your calendar, which makes deposits feel free when they are not. If you introduce one, measure total attended appointments rather than no show rate, because the rate can improve while your revenue falls.
The middle option most small businesses skip is a card held without a charge, combined with a clearly stated late cancellation fee that you rarely enforce. The commitment device does the work and the goodwill survives, which matters for a business where every customer is also a local recommendation.
Build a waitlist before you build anything else
A no show only costs you the full amount when there is nobody to take the slot. A waitlist converts the same event into a schedule change, and most small businesses do not keep one because nobody ever asked them to.
It does not need software. A list of names, the slot each person wanted, and how much notice they need is enough to start. The discipline is in the last column: somebody who can come at two hours' notice is a completely different asset from somebody who needs three days, and mixing them together is why informal waitlists fail to convert.
When you do reach for a booking system, judge it on this feature above the calendar itself. Ask what happens automatically the moment a cancellation lands, and whether it contacts the shortest notice people first. Many products treat the waitlist as a signup form rather than as the thing that recovers the revenue.
Restaurant bookings deserve their own note here because the slot is not the unit, and they sit alongside a wider set of decisions covered in the ranking of which restaurant jobs AI should take first. A cancelled table for two at eight is easier to refill than a table for eight at eight, and a waitlist sorted by party size and flexibility is worth more than one sorted by who asked first.
Where an assistant genuinely helps
In the parts that are laborious rather than the parts that are clever. Three of them are worth setting up.
Filling the gap. The moment a cancellation lands, somebody has to contact the people who wanted that slot. This is time critical, repetitive and perfectly suited to automation, and it converts a cancellation into a filled hour rather than into a lost one.
Answering the booking question. Most enquiries are availability, price and parking. An assistant connected to your actual calendar answers those instantly, and connecting it to real data rather than to a copy is exactly the problem that standard connectors exist to solve, which is the argument behind a proper connection between an assistant and your own systems. An assistant guessing at availability is worse than no assistant.
Taking the call. A phone answering agent that books when you cannot pick up is genuinely useful for a business where the phone rings during appointments, though it is judged on responsiveness rather than realism, which we went into in what a phone agent is actually judged on.
One caution, and it is not hypothetical. Giving an agent write access to a booking calendar has gone wrong in public before, which we documented in the case of an agent that cancelled a stranger's class. Read access to answer questions is a different risk from write access to change other people's bookings, and the two should not be granted in the same afternoon.
What to measure
Three numbers, tracked monthly, none of which requires new software.
The first is your no show rate by day of week and time of day, because it is almost never uniform and the pattern usually points at something fixable. Early morning and the slot immediately after lunch are the usual offenders.
The second is your cancellation notice distribution: how much warning you get, in hours. If most cancellations arrive under two hours, your problem is that cancelling is too hard or too embarrassing rather than that people are unreliable.
The third is attended appointments in total. It is the only number that pays you, and it is the one that protects you from optimising a rate while shrinking a business.
What none of the research covers
Every trial cited here measured a population that had already agreed to attend something. None of them tell you what to do about the customer who books three times, misses twice, and books again.
That decision is yours and it is a commercial one rather than a statistical one. A small business can afford to lose an unreliable customer and often cannot afford to keep them, because the empty slots they generate are paid for by everyone else. Requiring prepayment from a specific person is not a punishment, it is a different set of terms, and saying so plainly is kinder than quietly hoping.
Write the threshold down before you need it. Two missed appointments inside a year moves that customer to prepaid booking. Applied consistently and explained once, it stops being a judgement about a person and becomes a rule about a slot.
The order to do things in
Rewrite the reminder first, this week, naming what the slot costs. Add the second reminder close to the appointment. Make cancelling one tap and explicitly free. Measure for two months.
Only then consider prediction, and when you do, treat the score as a prompt for a gentler intervention rather than as permission to double book. If you are choosing customers to prioritise for anything, the same reasoning about acting on a probability applies as in deciding which quiet customers are worth chasing.
The empty chair is not usually a technology problem. It is a message problem, a timing problem and an awkwardness problem, in that order, and all three are fixed with sentences rather than software.