Square's Kiosk Takes Orders by Voice. Can Your Business?

By Lazar Milicevic · Published October 17, 2026 · 8 min read

You spend hours a day turning messy customer messages into something structured: a quote, a booking, an order. Square just shipped a kiosk that does this by voice, but the interesting part isn't the hardware. It's the pattern underneath, and you can build it this week without buying a screen.

What Square actually shipped

Square officially unveiled the new Square Kiosk on October 14, 2026, which it calls its flagship intelligent self-service kiosk (Square press release). The headline feature is an agentic, conversational voice layer that understands the seller's menu and responds in real time. Customers can ask menu questions, request modifications, and work through substitutions.

Facts worth knowing before you form an opinion:

• Hardware: a 21.5-inch display, three directional microphones with AI noise filtering, and on-device payment by tap, dip, or PIN (Square product page).
• Input: voice, touch, or both mixed. Touch stays available throughout the order.
• Pricing: $1,999 including a counter stand, plus a $99 monthly software fee per device. It requires Square Plus ($49 per location per month) or Square Premium ($149 per location per month) (Kiosk Marketplace).
• Limits: US only, English only. Square itself says voice may not suit extreme noise environments like music venues and recommends touch there.
• Safety valves: if the AI can't answer, it can loop in an employee, and there's a button to turn the AI voice off (Restaurant Business).

What's missing matters as much. Square doesn't name the model or vendor behind the voice layer. It publishes no figures for order speed, accuracy, or order-value growth, and its FAQ says results vary by business (FourWeekMBA). So anyone quoting you a "20% ticket lift" is making it up. I haven't seen independent accuracy numbers, and I won't pretend otherwise.

The real product is the grounding, not the voice

The AI isn't a general chatbot guessing at what a coffee shop sells. It's grounded in one thing: the seller's own menu. Its job is to turn messy speech into a clean, structured order against a catalog it actually has.

Square can do this because it owns the stack. Menu, price, and inventory changes made in Square reach every kiosk automatically, and orders land on the kitchen display in real time (Unite.AI). The kiosk also draws on the seller's sales data (what sells best at that location, time of day, what's already in the cart) to suggest add-ons (Square product page).

Strip away the pill-shaped enclosure and you have three layers:

[ interface ]      voice / touch / email / web form   <- commodity
[ grounding ]      your catalog, prices, rules         <- your asset
[ plumbing ]       POS, calendar, billing, kitchen     <- where projects die

The interface layer is the part everyone photographs. The grounding layer is what makes answers correct. The plumbing layer is what makes an understood order become a real one. Square controls all three. You probably don't, which is why the order of operations matters (more on that below).

This pattern is bigger than restaurants

Square's launch is aimed at quick-service restaurants. Whether it fits other shops is my inference, not something Square claims, and its own page notes a packaged-goods scanning feature isn't available yet. But the underlying problem is everywhere.

A customer emails: do you do X, how much, when can you start? That's the same translation problem. Messy human language comes in. Your services, prices, and rules are the menu. Someone has to produce a structured result: a quote, a booking, an invoice. If you lose two to four hours a day to that translation, that's manual work a menu-grounded agent can handle.

Here's what the structured output looks like when it's a service business instead of a burger joint:

{
  "intent": "quote_request",
  "service": "monthly_bookkeeping",
  "options": { "transactions_per_month": "under_100", "start": "next_month" },
  "price_range": "per menu doc, tier 1",
  "needs_human": false,
  "open_questions": ["Which accounting software do you use?"],
  "draft_reply": "..."
}

Notice needs_human and open_questions. A good grounded agent knows what it doesn't know. That's the same behavior as Square's "loop in an employee" fallback, and it's the difference between a useful assistant and a confident liar.

Write your menu: the no-hardware move

Here's the exercise that does the work. Open a doc and list:

• Every service or product you sell
• The price or price range for each
• The options customers choose between
• The rules: minimum order, lead time, what you don't do, what you always require upfront
• The five questions you answer over and over, with your actual answers

That document is your grounding data. Most small businesses can't hand an AI their offer because the offer lives in the owner's head plus a messy spreadsheet. If your pricing and rules are in your head, no AI, Square's or anyone's, can sell for you accurately.

A starting skeleton, in YAML so it's both human-readable and machine-readable:

services:
  - id: bookkeeping_monthly
    name: Monthly bookkeeping
    price: "tiered, see below"
    tiers:
      - { name: starter, includes: "up to 100 transactions", price_usd: 0 }  # fill in yours
    lead_time: "2 weeks to onboard"
    requires: ["read-only bank access", "prior-year statements"]
    does_not_do: ["tax filing", "payroll"]
rules:
  - "No work starts before the deposit clears."
  - "Rush requests: ask a human."
faq:
  - q: "Do you work with QuickBooks?"
    a: "..."

The price_usd: 0 placeholder is deliberate: put your real numbers in. The agent should never improvise a price. If a field is empty, it escalates.

Once this exists in clean form, you can point an AI assistant at your inbox or website form and have it draft accurate replies, build a quote, or pre-fill an order for your approval. The pattern to aim for: the model drafts, the structured doc constrains it, and you approve before anything goes out.

Draft-and-approve first, then loosen the leash

Don't start autonomous. Start with the agent preparing the answer and you hitting send. Then measure. Real numbers, no estimates.

For two weeks, track two counters:

# a dead-simple log: one line per handled message
echo "$(date -Iseconds),handled,quote_request,corrected=no"  >> agent_log.csv
echo "$(date -Iseconds),handled,booking,corrected=yes"       >> agent_log.csv

# end of week: correction rate
awk -F, '{n++; if ($4=="corrected=yes") c++} END {printf "%d handled, %d corrected (%.0f%%)\n", n, c, 100*c/n}' agent_log.csv

How to read the result:

• Low correction rate: loosen the leash. Let it auto-send the narrow categories it nails (FAQ answers, standard quotes) and keep approval on the rest.
• High correction rate: your menu doc has gaps, and you just found them cheaply. Fix the doc, not the prompt.

This is the same discipline Square is implicitly buying with its human-escalation button. The failure mode of voice AI isn't that it's dumb, it's that it's wrong with full confidence. Survey data cited in Restaurant Business (Technomic) shows 46% of diners find voice AI in the drive-thru unappealing, and ordering with a human at the counter is still the most popular method at 46%. Customers aren't automatically thrilled with AI intake. A human fallback isn't a nice-to-have.

The plumbing is where this succeeds or dies

An order that's understood but never reaches your calendar, accounting, or inventory is a demo. Square gets integration for free because it owns the POS, payments, and menu data. You probably have Gmail over here, a CRM over there, billing somewhere else.

So build in this order:

1. Menu doc (an afternoon, $0)
2. Draft-and-approve on one channel, usually email or a web form
3. One write-back: the approved quote becomes an invoice or a calendar hold, not a copy-paste
4. Correction logging, then expand

In my experience, the AI part, reading a request and producing structured fields, is the easy afternoon. Getting the structured result into an actual invoice reliably takes far longer. If you have to pick where to spend effort, spend it on step 3.

A rough comparison of paths, with the caveat that I'm describing categories, not pricing any specific tool:

If you're a restaurant on Square, the kiosk may genuinely be worth pricing out. Square lists the individual prices (device, monthly software fee, and plan), but the total for the first year is something you add up yourself from those list prices, plus processing fees (2.5% + 15 cents per tap or dip on Square Plus per FourWeekMBA), and check current rates on Square's kiosk page. If you're a service business, the hardware is irrelevant and the grounding is everything.

Where this shows up in client work

bizflowai.io focuses on practical AI automation for solopreneurs and small teams, and this pattern is a good example of the approach: get the offer into a clean, structured form, connect an agent to the tools you already use, and keep a human approving until the numbers say otherwise. More at bizflowai.io.

The hot take

The kiosk isn't the story. In two years, a screen that takes orders by voice won't impress anyone. The businesses that win will be the ones that structured their own offer well enough that any AI, on any device, can sell it accurately. Your menu data is the asset. The hardware is a commodity.

Write the doc this week. You'll learn more from the gaps in it than from any demo.

Want more like this?

I write about practical AI automation for solopreneurs and small teams.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milićević, senior engineer, BizFlowAI.

Visit bizflowai.io for our services, case studies, and AI consulting.