AI Phone Agent for Coffee Shops
Coffee phone orders are five billable decisions in six words: size, milk, shots, syrup, temperature. Here is what a phone agent has to price and repeat back.
On this page
"Large oat milk latte, extra shot, one pump vanilla, not too hot."
Twelve words, five decisions, and every one of them is priced, timed, or both. The size changes the cup and the milk volume. Oat is an upcharge and a separate carton. The extra shot is a second pull at the machine. The syrup is a pump count the barista has to hear correctly over a grinder. "Not too hot" is a steam temperature, not a preference.
A person on bar takes that sentence while tamping. A phone agent either gets all five attached to the right drink or the bar makes it twice.
TL;DR
- A coffee order is 5 billable decisions inside one sentence, and an agent that hears the drink but loses the milk swap has made a drink the bar cannot price.
- Milk swaps and syrup pumps are 2 of the 5 decisions, and both are missing from most menus until someone adds them as priced choice groups.
- Coffee breaks the usual missed-call argument: at a $7 ticket, recovering one call is small change, and the value is the barista who stays on the bar.
- Kovo answers after close and on holidays, takes the drink from your own modifier groups, and starts at $99 a month for 750 connected minutes.
- It keeps no order history for "the usual", it reads your own allergen note back with a disclaimer rather than answering a safety question, and orders past the shop's own limit are declined.
Why a coffee order is more than one item
Most verticals agree on what an order is: a dish, a quantity, a modification. Coffee splits the transaction across the bar.
A twelve-ounce drink and a sixteen-ounce drink made with the same milk, the same shots, and the same syrup are two different products with two different prices, and the size has to be chosen first or every later answer is wrong. Milk is not a garnish; it is most of the cup, and it is a cost line the shop tracks. A shot is labor. A syrup is inventory. A temperature is a promise about a steam wand.
That is why a coffee phone order falls apart at exactly the point a generic agent feels finished: it heard a drink name and a size and moved to the next item.
The test is a drink with four modifiers, said fast, in the order a regular would say it. If the read-back comes out as "latte, large, oat milk, vanilla, extra shot, warm" with the upcharges attached, the agent understood the transaction. If it comes back as "large oat milk latte with extra shot and vanilla, not too hot" as free text, the bar will read it and guess at the price.
Milk is a price, not a preference
Ask any coffee operator which phone question they get wrong most often and the answer is usually the milk.
Whole, skim, oat, almond, soy, and whatever local alternative is on the shelf this month. Each has a cost, some are an upcharge, some are out, and the difference between oat and almond is the difference between a caller's afternoon and a wasted drink.
Kovo prices what the menu says. That means the milk list has to exist on the item as a choice group with its own prices, not as a line in a description that a cashier reads. A shop that adds oat milk to the menu but not to the modifier group has a phone agent that will take an oat latte and charge for whole.
Two details are worth writing down before launch:
- What to say when an alternative is out. A caller who hears "we're out of oat" will substitute; a caller who arrives to a drink made with something else will not come back.
- How the upcharge reads in the total. The read-back should price the swap, so the number at the bar matches the number on the phone.
The menu sync guide covers the import side; the summary is that a modifier group is only as good as the POS catalog behind it.
The morning line is the real cost
At 8:12 on a weekday, a coffee bar has six tickets on the rail, a line at the register, and one person on bar. The phone rings in the middle of that, and the person who answers it is the same person steaming milk.
That is the actual loss, and it is not measured in missed calls. It is measured in the ninety seconds the barista spends with a handset against their ear while the queue grows and the milk sits at temperature. One caller asking whether the pumpkin is back can cost the shop the patience of 3 walk-ins.
A phone agent's job in a coffee shop is therefore closer to triage than to sales. It answers the four questions that arrive all morning — hours, the seasonal drink, whether the beans are in stock, whether there is a table — and takes a drink order in the same call. The bar keeps both hands.
It also picks up the calls nobody answered before: the 6:40 a.m. question about opening time, the 4 p.m. question about whether the shop closed at 3. Kovo answers 24 hours a day, including after close, from the hours and holiday exceptions the shop set.
The arithmetic at a small ticket
Do this honestly, because the category numbers do not fit a coffee shop.
A phone ticket here is small. If your average order is $7 and you miss 10 calls a week, that is $70 a week, or about $280 a month. Industry estimates put a missed restaurant call at $35 to $85, but that range assumes a meal-sized ticket; a coffee call that is really "are you open" is worth zero, and pretending otherwise is arithmetic with a thumb on the scale.
So run the comparison differently. What does the plan buy? Starter is $99 a month for 750 connected minutes, roughly 300 calls. If those 300 calls are mostly the four questions above plus a drink, the return is the bar hours you keep, not the tickets you recover. Ten minutes a day of bar time is five hours a month, and labor runs 25 to 36% of revenue, so those five hours are a line worth counting.
The missed-call math is still worth reading, because it shows where the range comes from and when it does not apply.
Where the caps and the gaps land
Kovo's limits belong to the shop: defaults of 10 items in total and 8 of one item, both editable, an optional order-value limit, and a zero that removes any of them. For a coffee shop the per-item cap is the one that matters.
An office run of 10 lattes is past it. So is a dozen croissants, a box of cold brew for a meeting, or the Saturday morning order from the salon two doors down. None of those are calls the agent should take. Above a limit it rings the handset only if you turned on large-party routing, and otherwise it says the order is too large to send through and keeps taking what fits, so a shop with a standing office account needs that routing, a handset number, and a process on the counter.
The gaps matter as much. The agent keeps no order history, so "the usual" is a person call. Allergens are a fixed script: it repeats your own allergen words back with a disclaimer and tells the caller to confirm with the bar, and it never says a drink is safe, so the syrup question is only as good as the note you wrote. Wait time is your own number, a baseline you set with an optional per-ticket addition and a cap, and if you have set nothing the agent says nothing rather than inventing twenty minutes. All three are sentences to put in the knowledge the agent answers from, plus a handset for the rest.
What Kovo does on a coffee line
Kovo answers, takes the drink from the shop's own modifier groups, reads the order back, and sends the ticket to Toast, Square, Clover, Olo, or an Epson or Star printer. Menu questions, hours, and recommendations stay on the call, and the handset takes over when a caller presses 0 or asks for a person. A website chat runs on the same engine, which keeps the 6:40 a.m. opening-hours question off the bar phone entirely.
Payment is the part to decide before you buy. Coffee shops usually run the simplest path first: cards off, so the order confirms on the call and the caller pays at the bar. With cards on, a card the shop already holds for that number is charged off-session while the caller is still on the line, which for a $7 drink is the path that costs the bar nothing. A first-time caller gets a Stripe Checkout link by text with the line held open, and paying it vaults the card; for a $7 drink that link is friction, and it is worth testing with your own regulars before you turn it on. No card number is spoken, recorded, transcribed, or sent to a model. The Toast routing page shows the ticket on one of the four systems.
Who this is wrong for
If nobody ever calls your shop, do not buy a phone product. A mobile-order-first café with a full app and a quiet line has already solved this problem with software it pays for.
If your phone is mostly vendors, deliveries, and the occasional applicant, the agent answers menu and hours questions and nothing else worth $99. Test that honestly for a week: count calls and sort them.
And if your morning peak is so tight that any answered call is a problem, no phone agent fixes staffing. It shortens the interruption, it does not remove the need for a second person at 8 a.m. The burger restaurant guide is a better fit when the ticket is a meal, and the restaurant phone AI guides list the rest.
FAQ
Can an AI phone agent handle milk swaps and syrup upcharges?
It handles them as modifiers when the menu defines them that way. Kovo collects all 5 choices, size, milk, shot count, syrup, and temperature, from the choice groups on the item, then reads the whole drink back with the priced add-ons included. If oat milk is not a choice group on your menu, the agent cannot charge for it, so check the menu before the phone.
How does the agent handle an order for ten lattes?
The per-item default is 8, and 10 lattes is past it. So is an office run of a dozen pastries, which trips the same line. Both are the shop's numbers to set: defaults of 10 items in total and 8 of one item, an owner can change either, add an order-value limit, or set either to zero, and above a limit the call rings the handset if you turned on large-party routing. Otherwise the agent says the order is too large to send through and keeps taking what fits, so a shop with a standing office account should route those calls to a person.
Does the agent remember a regular's usual order?
No. Kovo keeps no order history, so there is nothing for 'the usual' to resolve against, and the agent answers from 4 first-party sources: your site, menu, FAQs, and hours. A regular who says 'the usual' has to reach a person, so set the handset number and say so on the line rather than letting the agent guess at a drink.
What happens to payment on a small coffee order?
Three paths. With cards off, the order confirms on the call and the caller pays at the bar, which is what most coffee shops start with because a $7 ticket is faster to tap than to read. With cards on and a card the shop already holds for that number, it is charged off-session while the caller is still on the line, which is the path a morning regular takes. A first-time caller gets a Stripe Checkout link by text with the line held open, and no card number is spoken, recorded, transcribed, or sent to a model.
How much does a coffee shop phone agent cost?
Kovo Starter is $99 a month for 750 connected minutes, roughly 300 calls, and Unlimited is $299 for 5,000 minutes, with no per-minute fee and no overage. At a $7 average ticket, the plan does not pay for itself through recovered orders alone, so buy it because the bar needs the interruption removed, not because the phone is a revenue channel.
Final Thoughts
For a coffee shop the decision is not "does a phone agent take orders". It is whether the agent can price the milk, hear the shot and the syrup, and answer the four questions that interrupt the bar every morning. The recovered tickets are small; the recovered bar time is the product.
Kovo takes the drink from your own menu, reads it back, and prints it to the POS you already run, from $99 a month with 14 days free and no card. Call in a large oat latte with an extra shot and a pump of vanilla, and read the ticket the bar receives. Take a test call with your own menu.