The restaurant industry ran the most expensive public AI experiment in retail, and it published the results by accident. Multi-year pilots, hundreds of locations, viral failures, an SEC enforcement action. If you run a single café or a three-site restaurant group, all of that research is now free.
The findings are consistent and slightly boring, which is why nobody sells them to you: AI for restaurants works well in quiet, structured channels and fails in noisy, unpredictable ones. Everything useful in this article follows from that one sentence.
What the drive-thru experiment proved
McDonald’s ran automated order-taking with IBM across roughly 100 test locations from 2021 and ended the partnership in June 2024, after accuracy sat around 85% against an internal bar reported nearer 90–95% — and after clips of the system mishearing orders spread widely online. The company didn’t abandon the idea; in December 2024 it announced a new voice-ordering partnership with Google Cloud (Restaurant Dive).
Wendy’s went the other way. FreshAI, built with Google Cloud, launched in Columbus, Ohio in May 2023 and expanded to 500-plus locations, with the company reporting a roughly 22-second improvement in order time and framing the system throughout as augmenting crew rather than replacing them. Taco Bell pushed voice AI to hundreds of sites, processed millions of orders, and then publicly moved into reassessment.
An honesty note on the numbers. Published accuracy figures for these deployments disagree substantially — you’ll see 83%, 85%, 86%, 90% and 92% quoted for the same systems, sometimes in the same month, and reports that roughly one in five AI orders still needs employee support. The figures come from company disclosures, trade press and vendor blogs with different definitions of “accurate” and different measurement windows. Treat the direction as real and any specific percentage as soft.
The pattern underneath them isn’t soft at all:
| Condition | AI performance |
|---|---|
| Quiet channel, one speaker, no time pressure | Strong |
| Constrained menu with predictable phrasing | Strong |
| Road noise, engine noise, multiple voices from a car | Weak |
| Regional accents and local shorthand | Weak until specifically trained |
| Modifications phrased three different ways | Weak |
The chains spent years discovering that the hard part wasn’t language, it was acoustics. That matters for you because it tells you which channel to automate first — and it isn’t the counter.
The “non-intervention rate”: read this before you believe any vendor number
On 14 January 2025 the SEC brought settled charges against Presto Automation over its Presto Voice drive-thru product. Presto consented to a cease-and-desist order without admitting or denying the findings (Cooley, DLA Piper).
Two findings are worth every restaurant operator’s attention.
The technology wasn’t what it was described as. The SEC found that Presto told investors its product eliminated the need for human order taking, when the original proprietary version required a human agent to enter the order in every instance, and the more advanced pilot version required human entry roughly 70% of the time during a measured period. Those humans were hired, trained and supervised staff working abroad, primarily in the Philippines and India. For an earlier period, units in the field ran on a third party’s technology while being described as Presto’s own.
And here’s the part that should change how you read every vendor deck. Presto reported “automated order completion” and “non-intervention” rates. Those rates measured orders completed without restaurant staff involvement. They did not mean without human involvement — the offshore agents didn’t count in the denominator. The metric was technically defensible and completely misleading, and the SEC found the company had identified the confusion internally and left it uncorrected until it learned of the investigation.
This is the same trick as deflection rate in support software, dissected in building an AI customer support system: a number that sounds like it measures success while quietly excluding the failure cases.
So when a vendor quotes you an automation rate, ask two questions before anything else. What exactly is in the denominator? And who or what completes an order the AI can’t? If the answer is “a person in a call centre,” that’s fine — it’s a perfectly good product. It just isn’t the product you were quoted.
Where AI actually pays for an independent
Ranked by return, for a single site or small group:
| Use | Why it works here |
|---|---|
| Answering the phone | Quiet line, one caller, no engine noise. The channel where the chains’ failures don’t apply — and where missed calls are lost covers |
| Being findable in AI answers | Costs nothing but structured data. See below |
| Drafting review replies | High volume, low stakes, owner approves each one |
| Menu descriptions, social posts, email | Real hours saved on work that was never billable |
| Demand forecasting for prep and rota | Genuine money, but needs a year of clean POS history first |
| Voice ordering at the counter | Last. This is precisely the noisy environment that defeated companies with far bigger budgets |
The phone deserves the top slot on plain arithmetic. A café that misses fifteen calls a day during service — bookings, large orders, “are you open on Monday” — is losing revenue that nobody records anywhere. It’s the one channel where the customer has already decided to buy, and it fails silently.
It’s also acoustically the opposite of a drive-thru speaker: one person, a handset, no road. Everything that broke the chains’ systems is absent. Set it up the way building an AI chatbot without code describes — write the fifteen or twenty questions that cover most of your call volume before you look at a single vendor.
Being found matters more than being fast
Here’s the shift most restaurant AI advice hasn’t caught up with. A growing share of “where should we eat” now gets answered by an assistant reading structured data rather than by a person scrolling a map.
Those systems pull the same things every time: opening hours, address, price band, cuisine, dietary options, whether you take bookings, whether you deliver. If that data is incomplete, stale or contradictory across your website, your Google Business Profile and the delivery platforms, you’re invisible to the query — not ranked lower, absent.
This costs nothing and almost nobody does it properly:
- Hours that are actually correct, including holidays and split service. This is the single most common error and the most expensive one.
- A machine-readable menu on your own site — real text, not a PDF and not an image. A photographed menu is invisible to everything that isn’t a human eye.
- Explicit dietary and allergen labelling as structured fields, not prose buried in a paragraph.
- Consistent name, address and phone across every platform, character for character.
- Booking and delivery links that resolve, on every listing.
This is the restaurant version of the product-feed argument in AI ecommerce tools that increase sales: the customer may never load your page, so your data is the storefront. It’s unglamorous, free, and worth more than any subscription on this list.
Reviews: automate the reply, never the review
Drafting replies to reviews is a legitimate, high-return use. You’re responding to dozens of comments a week in your own voice, most of them saying similar things, and a draft you edit in twenty seconds beats a reply you never wrote.
Three rules keep it useful. Every reply gets read and edited by a person before it posts. Every reply names something specific from that review, because generic gratitude reads worse than silence. And complaints about food safety, illness or allergens never get an automated reply of any kind — those go straight to a human, immediately.
What you must not do is generate reviews. Fabricated or incentivised reviews violate every major platform’s policies and sit squarely in consumer-protection territory. The line is simple: AI can help you answer customers, never impersonate them.
Three things to keep AI away from
Allergen and dietary answers. This is the hard stop. A generative model will confidently produce a plausible answer about whether a dish contains nuts, and a plausible answer is worthless when someone can be hospitalised. Allergen information must come from a maintained lookup that you control, or from a person with the recipe in front of them. If a phone or chat assistant is asked an allergen question, its only correct behaviour is to hand off to a human. Build that as a hard rule, not a prompt instruction. The general failure mode — fluent, confident, wrong — is covered in fact-checking AI-generated content; in a kitchen it stops being an editorial problem.
Automated price changes. Demand-based pricing is technically easy and reputationally expensive in hospitality. Customers read a price that moves with demand as being punished for arriving hungry, and the story travels faster than the margin.
Rota decisions without a manager. Forecasting demand is a good AI use. Converting a forecast directly into who works Saturday is not — it ignores everything the model can’t see, and in a sector already fighting for staff, it costs you people.
A sensible order for one site
- Fix your data first. Hours, menu as real text, allergens, address, links. A weekend of work, no subscription.
- Count your missed calls for two weeks. This is your baseline and often the whole business case.
- Add phone answering for the questions you wrote down, with a hard escalation rule for anything about allergens, complaints or large-party bookings.
- Draft review replies with AI, publish none unread.
- Move to content — menu copy, social, email. The list-building and deliverability side is in email marketing automation, and the posting cadence in AI tools for social media management.
- Only then look at forecasting, and only if your POS history is clean.
Steps one and two cost nothing and produce most of the return. That’s the honest shape of AI for restaurants in 2026, and it’s roughly the same shape as cutting business costs with AI generally. For the wider stack, the best AI tools for small business owners is the map.
Frequently asked questions
Does AI order-taking actually work in restaurants?
It depends almost entirely on acoustics. On a phone line with one caller it works well. At a drive-thru speaker or a busy counter, with road noise, multiple voices and modified orders, it has repeatedly underperformed — McDonald’s ended its IBM pilot in June 2024 over accuracy, while Wendy’s scaled a system it consistently described as supporting crew rather than replacing them.
What should I ask an AI ordering vendor?
What exactly is in the denominator of their automation rate, and who completes an order the AI can’t. The SEC’s 2025 action against Presto Automation found the company’s reported “non-intervention” rates measured orders completed without restaurant staff involvement, while offshore human agents were entering a large share of orders. Offshore support isn’t disqualifying; not being told about it is.
Can I use AI to answer allergen questions?
No. A generative model will produce confident, plausible allergen answers that may be wrong, and the consequences are medical rather than commercial. Allergen information must come from a maintained record you control or from a person with the recipe. Build a hard escalation rule so any allergen question routes to a human immediately.
Is it acceptable to reply to reviews with AI?
Yes, if a person edits and approves every reply, each response names something specific from the review, and anything involving illness, food safety or allergens goes to a human instead. Generating reviews themselves is a different matter entirely and breaches platform policy.
What’s the cheapest high-return AI move for a small café?
Fixing your structured data — accurate hours, a machine-readable menu in real text rather than a PDF or photo, labelled dietary information, and consistent details across every platform. Assistants answering “where should we eat” read those fields, and incomplete data makes you absent from the answer rather than merely lower down it.
Should a single-site restaurant buy AI forecasting software?
Not first. Forecasting needs a long run of clean point-of-sale history to be worth anything, and most of the available return sits earlier in the list — missed phone calls and incorrect listings. Get those fixed, then revisit forecasting once you have a year of reliable data.
