Wendy’s said its drive-thru AI hit nearly 99%. Read the definition before you quote that number. According to Restaurant Dive’s December 2023 report, an order counted as a success if the chatbot started it and it reached the point of sale, even if a human had to join the conversation to fix an inaccuracy.

I’d call it the most useful design detail in the whole story. It’s also the first thing dropped when the case circulates on LinkedIn.

Two drive-thrus, one lesson

Andrew Amann’s newsletter “McDonald’s AI McTrouble” put the two chains side by side this week. His read is that McDonald’s asked the model to hear the order, make sense of it and ring it up, with no quantity caps and no escalation, while Wendy’s wrapped its model in business logic. “AI was the ears,” he writes, and the rules engine was part of the brain.

I checked what I could. McDonald’s and IBM started testing automated voice ordering in October 2021 across roughly 100 locations, and McDonald’s said in June 2024 it was ending the test, per Nation’s Restaurant News. The AP story describes a 2023 TikTok where the system kept stacking nuggets on one car’s tab while the customers, laughing, begged it to stop. Franchisees were told to switch it off by July 26.

Whether McDonald’s had zero caps, I can’t prove from public documents. The failure mode in those videos fits Amann’s diagnosis, though. A model that hears “more” keeps adding, because nothing outside the model knows what a normal order looks like.

On Wendy’s, the public numbers say something a bit different from “rules engine.” They say person in the loop. Twenty-two seconds faster than the local market average at a Columbus test site. A success rate near 99% that includes human fixes. And a stated target for fully hands-off orders of 30% or more, climbing to 70% by the end of 2024. I couldn’t find whether they reached 70%, so I won’t guess.

Either way, Wendy’s measured the system, model plus staff. McDonald’s story reads like the model was expected to carry it alone.

The startup that bans its agents from arithmetic

Then there’s Hilbert. It raised a $28M Series A led by a16z in April. In a Startupeable interview (in Spanish), CEO Nazli Tan, who ran growth at Getir, argues that piling more context on an agent makes it worse. Context usually means someone typed out their gut definition of “good customer” or “churn risk,” and the agent runs with it. Ask twice, get two lists.

Hilbert trains deep learning models on at least three or four years of real customer history and lets those models decide. Agents work inside the bounds. They can’t even do math beyond rounding. Tan’s reason: her customers route hundreds of millions in marketing spend off those answers, and nobody trusts a budget to an answer that changes when you ask again.

An agent company that won’t let its agents add. That’s the most mature thing I’ve read about agents this month.

Where the rules go

Here’s the split we draw before automating anything that charges, pays or discounts. Language goes to the model: a rushed order, an accent, engine noise in the background, a customer changing their mind halfway. Arithmetic and policy go to code.

1. Caps on quantity and amount live in code, outside the prompt. You can talk a prompt out of a rule. You can’t talk a validation out of one.
2. Prices, discounts and taxes come from your system of record. Never from what the model remembers.
3. Escalation has an owner. Which cases go to a person, which person, and how long the customer waits.
4. Measurement separates the two. Cases the model closed alone, cases that needed someone. Wendy’s did this, and it’s why their 99% means something once you read the footnote.

Skipping that count has burned this industry before. Amazon’s Just Walk Out got in trouble because nobody outside knew how much human review sat behind it. And that measurement compounds when you keep it as your own eval set, rerun every time you switch models.

A cheap test if you already have an agent near pricing, credit or orders. Pull ten real cases from last week. Run each one twice. Any pair that disagrees marks where your first rule belongs.

Let’s map the rules your agent should sit inside