Sunday at 11:15. The room is full, three carts are working the floor, and the phone at the host stand has rung four times in ten minutes. Two of those callers wanted to know the wait. One wanted forty pieces of har gow for pickup at one. Nobody answered any of them, because the person who would have was seating a party of nine.
That is the whole problem in a single shift. Your busiest two hours of the week are also the two hours when nobody can pick up the phone.
The order is a list of counts, not a list of dishes
A dim sum takeout order is arithmetic. Six orders of siu mai, four har gow, two char siu bao, one cheung fun with shrimp, and the caller changes their mind twice while you write. Every line is a quantity attached to an item that has three common English spellings and at least two names.
That's structurally different from a burger order, where the modifiers hang off a single entree. Here the count is the order. Get it wrong and the ticket is wrong in a way the kitchen can't catch, because six siu mai and eight siu mai look equally reasonable coming off a printer.
A voice agent does well here for one specific reason. It reads the order back with numbers before it commits, every time, including at 12:40 on a Sunday. A host taking that call between two seatings skips the readback. That isn't a criticism of your staff. It's what happens when one person is doing two jobs and the party in front of them can see their face.
Steamed counts feed straight into kitchen capacity
Forty pieces of har gow at one o'clock is not a small request during peak service. It's four or five baskets that have to come out of the same steamers feeding your dining room, at the hour those steamers are already committed.
When the phone is answered by whoever is closest, that order gets accepted without anyone checking whether the kitchen can absorb it. The steamer team finds out when the ticket prints. A configured agent can cap how many large takeout orders land in a given fifteen-minute window and push the rest later, which is the same logic covered in order throttling and kitchen capacity.
That's a policy decision, not a technical one. You have to decide what your Sunday ceiling actually is. Most operators have never written that number down, and the exercise of writing it down is worth more than the software.
Cantonese, Mandarin, and English arrive on the same line
Plenty of your regulars order in Cantonese. Plenty order in English but say the dish names in Cantonese, which is the harder case for any speech system, because the sentence switches language mid-clause and then switches back.
Be clear about what is on offer, because we would rather lose the deal than have you find this out on a customer call. The agent does not converse in Cantonese. You choose English plus one other language, and for a dim sum room that choice is Mandarin. Cantonese dish names are a different problem and a solvable one: they go in as aliases against the POS item, so a caller saying har gow inside an English sentence lands on the right line.
Test this directly. During a demo, don't accept a general claim about language support from us or anyone else. Record five of your actual regulars placing their usual order, then play those recordings to the agent and listen. Our demo checklist covers how to run that, and the multilingual ordering guide covers what varies between languages. The languages page lists exactly what ships and what is only on the roadmap.
Here's the honest part. Accent and code-switching performance is uneven across vendors and across dishes, and no spec sheet will tell you where it breaks for your menu. Twenty minutes of testing with real recordings will.
Naming the menu is most of the setup work
The configuration effort for a dim sum restaurant is not the technology. It's deciding what each dish is called.
Take shrimp dumplings. A caller might say har gow, ha gow, hargow, "the shrimp ones," "the pink ones," or "number 12." All of those need to point at the same POS item. Same for shumai, siu mai, and "the pork ones with the orange top."
Before setup, sit down with whoever answers the phone most often and write out the alternate names people actually use for your top thirty items. Not the names on the menu. The names on the phone. This is the same exercise described in training a voice agent on your menu, and it is the hour that decides whether the whole thing works.
While you're at it, flag anything that only runs on weekends, anything the kitchen makes in limited quantity, and anything you stop selling after 3 p.m.
Numbered menus deserve special attention. If your printed menu numbers items, some callers will order entirely by number and some will mix numbers and names in the same sentence. Both have to work, and the number-to-item mapping has to match whatever is currently printed rather than the version from two menu revisions ago. If you've reprinted recently and the numbers shifted, that's the first thing to check.
One more thing worth writing down: which items come in a standard basket count. A caller who says "four har gow" might mean four baskets or four pieces, and your staff resolves that ambiguity by instinct. The agent needs an explicit rule, and a readback that states the unit out loud closes the gap.
What it won't fix
If your kitchen 86s items by shouting across the pass instead of marking them in the POS, the agent will keep selling a dish that ran out at noon. The agent reads the same availability data your registers read. It cannot see the steamer. That failure mode and its fixes are laid out in 86ing and menu sync.
It also won't help much if your phone rarely rings. A room doing eight calls a week doesn't have a phone problem worth spending on, and we'd rather say that than sell you something.
And it won't replace the judgment call on a party of twenty-five asking about a private room. Those calls should reach a person. Route them deliberately rather than letting the agent improvise.
The cart floor is the second, quieter win
Most of the attention goes to takeout orders, but for a cart-service room the bigger recovery is often the question volume. How long is the wait. Do you take reservations for six. Is there parking. Are you open on the holiday.
Every one of those calls currently pulls someone off the floor mid-service, which is the dynamic described in reducing host stand phone interruptions. None of them require judgment. Answering them without a human touching the phone is unglamorous and it's where the hours come back.
A test to run this Sunday
Put a phone on the host stand and have someone tally, for two hours, three things: calls that rang, calls that were answered, and what each answered call was about.
If more than a fifth of your calls go unanswered during your peak window, and if more than half of the ones you do answer are questions rather than orders, you have a phone problem with a clear shape. Run the arithmetic on what a missed forty-piece order is worth against the pricing page, where plans start at $250 a month, and decide from there. If the tally comes back small, do nothing. That's a legitimate outcome and it costs you one Sunday to find out.