BRENNEN

Industries / Retail, hospitality and personal services / Hospitality and tourism

Ordering and serving

This is the front of the restaurant: taking reservations, showing the menu, taking orders at the table, the counter, the drive-thru or through delivery apps, carrying food to guests and taking payment.

7

Milestones achieved, from 1765 to 2026.

2

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

In depth

The same ground is charted in more depth, as an industry of its own.

The road

AchievedWatch forOpen

Achieved

  1. 1765

    The first restaurants

    The word restaurant first meant a rich broth thought to restore health. Before restaurants, people in Paris ate from shared platters at inns, bought from street sellers or went to cook-caterers who each sold one trade, such as roast meat. By the account of Larousse Gastronomique, a cookery reference, a man named Boulanger opened a restaurant around 1765. The historian Rebecca Spang searched French archives and found no record of a Boulanger, and points instead to Mathurin Roze de Chantoiseau, whom an 18th-century gossip column called the "creator" of restaurants. The Grande Taverne de Londres, opened in 1782, is credited with a varied menu and separate tables.

    nationalgeographic.com

    Credit

    Boulanger or Roze de Chantoiseau; the sources disagree

  2. 1986

    Touchscreen tills

    A point-of-sale (POS) system is the till that records each order and payment. Gene Mosher, a restaurant owner, began writing POS software on an Apple II computer in July 1978. In 1986 Mosher paired an Atari ST computer with a MicroTouch touchscreen overlay to build what the ViewTouch site calls "the first commercial Point of Sale solution for restaurants and bars" with a graphical touchscreen, sold as ViewTouch. The "first" is Mosher's own claim.

    viewtouch.com

    Credit

    Gene Mosher

  3. 1998

    Reservations go online

    OpenTable, incorporated in October 1998, replaced the paper reservation book at the host stand with its Electronic Reservation Book, software that also took bookings made on the web. Restaurants paid a monthly fee and a fee for each guest seated through an online booking, and diners booked free. In its annual report to the US Securities and Exchange Commission for 2011, OpenTable listed about 25,000 restaurant customers and said it had seated more than 280 million diners since 1998.

    sec.gov

    Credit

    OpenTable

  4. 2014

    Allergens declared for meals

    From 13 December 2014 an EU regulation required businesses selling food without packaging, as restaurants do, to tell customers which of 14 listed allergens their food contains. The regulation's own reasoning is that most food allergy incidents can be traced to food that is not prepacked.

    eur-lex.europa.eu

    Credit

    The European Parliament and the Council of the EU, in Regulation (EU) 1169/2011

  5. 2020

    Caps on delivery fees

    Delivery apps such as DoorDash, Grubhub and Uber Eats charge restaurants a commission on each order. In May 2020 New York City capped those charges at 15% for delivery and 5% for advertising and other services, after fees as high as 30%, and made the caps permanent in August 2021. The three companies sued, and in June 2025 they settled in a deal that would let them sometimes charge more and that expects the City Council to change the law.

    insurancejournal.com

    Credit

    The New York City Council; DoorDash, Grubhub and Uber Eats

  6. 2023

    Robots carry the plates

    Wheeled robots that carry plates between kitchen and table spread as restaurants struggled to hire. The Associated Press reported in April 2023 that tens of thousands were in use worldwide, that Bear Robotics expected 10,000 of its Servi robots deployed by the end of 2023, and that Pudu Robotics had deployed more than 56,000. Chili's had halted its robot, Rita, in 2020 after expanding it to 61 US restaurants: it moved too slowly, got in servers' way, and 58% of surveyed guests said it did not improve their experience. In a 2024 University of Central Florida study of 42 restaurant workers, staff valued the physical relief and complained of slow speeds, glitches and missing training.

    LessonThe robots carry plates but cannot take orders or cope with steps and patios, and workers resented them when they arrived without training or explanation.

    fortune.comucf.edu

    Credit

    Bear Robotics, Pudu Robotics and Chili's; reported by the Associated Press and by Cynthia Mejia and colleagues at UCF

  7. 2026

    Voice AI at the drive-thru

    McDonald's ended its test of IBM's automated order taking in more than 100 restaurants by 26 July 2024. Sources told CNBC the system struggled with accents and dialects, and an analyst said franchisees reported accuracy in the low-to-mid 80% range, against the 95% or more that the analyst said was needed. In June and July 2026 mystery shoppers from Intouch Insight, a customer-research firm, made 120 visits to Burger King, Taco Bell and Wendy's restaurants reported to use voice AI. The AI took some or all of the order on 92 visits and completed 77 from start to finish. Orders were 90.8% accurate, up from 82.5% at AI locations in 2025; nearly three in ten shoppers had to repeat part of the order; and total time was 5:14 against 5:15 at the same brands in the main study.

    LessonIn 2024 an accuracy in the low 80s ended McDonald's test, and two years later outside shoppers found AI lanes about as accurate and as fast as ordinary ones, though still asking guests to repeat themselves.

    cnbc.comintouchinsight.com

    Credit

    IBM and McDonald's; Burger King, Taco Bell and Wendy's; measured by Intouch Insight

Watch for

  1. Taco Bell's voice AI spreads

    Taco Bell said in July 2026 that its automated voice ordering runs in nearly 900 US stores in 38 states. The only outside measure found, Intouch Insight's mystery shops, pools Taco Bell with Burger King and Wendy's over 120 visits.

    Achieved whenAn outside study reports order accuracy and hand-offs to staff for Taco Bell's AI lanes on their own, over enough visits to tell them apart from other brands.

    marketscale.comintouchinsight.com

    Claimed by

    Taco Bell, July 2026, as reported by MarketScale

  2. A robot that delivers

    DoorDash says its Dot robot, a four-wheeled vehicle small enough for doorways and driveways, travels at up to 20 mph on roads, bike lanes, sidewalks and driveways. It began working near Arizona State University's Tempe campus and in most of Mesa, Arizona, with a goal of covering 1.5 million people in the Phoenix area by the end of 2025. DoorDash would not say how many businesses use it, and said drivers would still make most deliveries.

    Achieved whenAn outside party, such as a city or a published study, reports routine Dot deliveries with delivery counts and incident rates.

    aztechcouncil.org

    Claimed by

    DoorDash, October 2025

Open

  1. 01

    Public drive-thru order test

    A public set of recorded drive-thru orders and a scorer. Volunteers in cars speak orders from a 30-item mock menu through a real speaker post and intercom in a private lot, with engines idling, wind, radios and passengers talking, and each recording is paired with the correct order ticket. The scorer compares any system's ticket with the correct one item by item, including sizes and changes such as "no onions". No real orders are placed and no payments are taken.

    The brief
    Credit · unclaimed
  2. 02

    Count a serving robot's steps

    A time-and-motion study of a serving robot. In a training restaurant at a culinary school, outside opening hours, staff and volunteers acting as guests run scripted services with the same set of orders, alternating services with and without a leased serving robot. Log each server's steps, trips to the kitchen, time from pass to table, spills, bumps, robot stops and each time a person has to help the robot. No public guests are served.

    The brief
    Credit · unclaimed
  3. 03

    Allergen test for order bots

    A public test of how AI ordering agents, the voice and chat systems that take orders, answer allergen questions. Write a mock menu of 40 dishes with full recipes and the labels of every bought-in ingredient, and an answer key for 500 questions about the EU's 14 listed allergens. Include hard cases: shared fryer oil, a sauce with a hidden ingredient, "may contain" labels, a recipe changed last week, and questions a bot should pass to a person. Score each answer as correct, wrong but safe, or wrong and unsafe, meaning it called a dish safe when it is not.

    The brief
    Credit · unclaimed

Open 01 · the brief

Public drive-thru order test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of recorded drive-thru orders and a scorer. Volunteers in cars speak orders from a 30-item mock menu through a real speaker post and intercom in a private lot, with engines idling, wind, radios and passengers talking, and each recording is paired with the correct order ticket. The scorer compares any system's ticket with the correct one item by item, including sizes and changes such as "no onions". No real orders are placed and no payments are taken.

Why it is still open

The one outside measure found, Intouch Insight's, rests on 120 mystery-shop visits and releases no audio. The ai-coustics drive-thru recordings, made in Berlin with an intercom, test only whether speech is present, not whether the order is right, and the post does not say they are public. McDonald's ended its 2024 test after reported accuracy in the low-to-mid 80s, and no public benchmark of drive-thru order accuracy turned up.

intouchinsight.comai-coustics.comcnbc.com

Built with what exists

A drive-thru speaker post or an intercom microphone and amplifier, as ai-coustics used; phone recorders; speech-to-text such as ElevenLabs Scribe v3, which accepts up to 1,000 listed terms such as menu item names, and Scribe v3 Live, which judges from the conversation as well as the silence when a speaker has finished (see the Note "Fewer word errors"); a language model to turn each transcript into a ticket; Python for scoring.

  1. First step

    Write the mock menu, record 50 orders from ten volunteers in a parked car through a low-cost intercom, and type the correct tickets.

  2. Done looks like

    At least 1,000 orders from at least 50 speakers at three noise levels, published with their tickets under an open license, and a table giving, for at least three systems, the share of whole orders right, the item error rate and the word error rate, with a person taking the same recordings as a baseline.

Open 02 · the brief

Count a serving robot's steps

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A time-and-motion study of a serving robot. In a training restaurant at a culinary school, outside opening hours, staff and volunteers acting as guests run scripted services with the same set of orders, alternating services with and without a leased serving robot. Log each server's steps, trips to the kitchen, time from pass to table, spills, bumps, robot stops and each time a person has to help the robot. No public guests are served.

Why it is still open

The one measured step count found comes from a 90-day pilot of Bear Robotics' Servi in two senior-living dining rooms in 2022, run by the Front Porch Center for Innovation and Wellbeing. Pedometers showed 5.3% and 8.8% fewer steps per server, measured before and after with no comparison group, and the authors wrote that the results "may not point to significant step savings". The University of Central Florida's 2024 study of 42 workers was built on interviews, and its authors call for quantitative validation of what workers reported. Robot makers publish shipment counts, such as Pudu's 100,000 robots in June 2025, not measured effects on staff or service.

fpciw.orgucf.edupudurobotics.com

Built with what exists

A leased serving robot such as Bear Robotics' Servi or Pudu's BellaBot; step counters or fitness bands for each server; ceiling cameras; a public spreadsheet; written definitions of a spill, a bump and an assist agreed before the first service.

  1. First step

    Ask a culinary school with a training restaurant to host it, and write the service script and the definitions of a spill and an assist.

  2. Done looks like

    At least ten scripted services of 40 covers each, five with the robot and five without, with public logs of steps per server per cover, pass-to-table times, spills, robot stops and human assists, and the script published so another school can repeat it.

Open 03 · the brief

Allergen test for order bots

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of how AI ordering agents, the voice and chat systems that take orders, answer allergen questions. Write a mock menu of 40 dishes with full recipes and the labels of every bought-in ingredient, and an answer key for 500 questions about the EU's 14 listed allergens. Include hard cases: shared fryer oil, a sauce with a hidden ingredient, "may contain" labels, a recipe changed last week, and questions a bot should pass to a person. Score each answer as correct, wrong but safe, or wrong and unsafe, meaning it called a dish safe when it is not.

Why it is still open

Vendors publish claims such as "94% accuracy on standard allergen queries" with no method, sample or source (Hostie AI, July 2025). The closest academic study tested ChatGPT's diet plans for people with food allergies, not answers about a restaurant's menu, and found it could produce harmful diets (Niszczota and Rybicka, Nutrition, 2023). No public benchmark of menu allergen answers turned up.

hostie.aipubmed.ncbi.nlm.nih.goveur-lex.europa.eu

Built with what exists

A mock menu and recipe sheets in a spreadsheet; the allergen list in Annex II of EU Regulation 1169/2011; general AI models given the menu, and commercial ordering agents where their terms allow testing; speech-to-text (see the Note "Fewer word errors") to run the same questions by voice; Python for scoring.

  1. First step

    Write 20 dishes with recipes and 100 questions, and have two people fill in the answer key separately and settle every disagreement.

  2. Done looks like

    500 questions with an answer key that two people agreed on independently, run on at least three systems by text and by voice, with a published table of correct answers, unsafe errors and correct hand-offs to a person for each system.