The New York Times Put AI in Its Recipes. What Every Business Adding AI Should Learn From It

NYT Cooking, The New York Times' recipe service, has added an AI assistant for subscribers. It sits beside a recipe and helps you change it: swap an ingredient you don't have, work with different equipment, or adjust the amounts for more people, according to JournalArta, which says it's on the iOS app and being introduced gradually.

The early reviews are worth reading closely. The Verge's reviewer said it rescued a weeknight chicken dinner but nearly ruined a cake. A WIRED writer, trying it on recipes, came away fairly impressed. Same tool, very different stakes. If you're thinking about putting AI into your own business, that gap is the whole lesson.

Why the chicken worked and the cake didn't

Cooking dinner is forgiving. JournalArta describes the assistant suggesting that baby spinach go in later than kale would, because it wilts faster. That's useful advice, and if it's slightly off, dinner is still dinner.

Baking is closer to chemistry. Small errors in a quantity change the result, and you don't find out until it comes out of the oven.

Businesses have chickens and cakes too. 'What time do you close?' or 'Which plan suits a small team?' are chicken questions. 'How much do I owe?', 'Is this drug safe with that one?' and 'What did my child score?' are cakes. Being a little wrong on those costs money or trust.

Start from something you've already checked

The reason this feature makes sense at all is what's underneath it. NYT Cooking has thousands of recipes, which JournalArta notes were tested by people. The AI isn't inventing dishes from nothing. It's adjusting something a person already got right, and the company describes it as customising recipes instead of replacing them.

That's the pattern to copy. An assistant that answers from your own price list, policies and records is far more useful than a general chatbot with your logo on it. If your information is scattered, out of date or stored in someone's head, sort that out first. AI will repeat your mess, only faster and more confidently.

Let systems do the sums

Here's a rule we'd apply to any business. AI can explain a number, but a number that matters should come from your system.

So a customer can ask a chatbot about their balance in plain English. The balance itself should be pulled straight from your records, not worked out or remembered by the AI. The same goes for prices on a quote, stock levels, dates on a contract and marks on an exam. Language is what AI is good at. Exact figures are what your database is for.

Roll it out like you expect mistakes

Two more details from the launch are worth borrowing. NYT Cooking is releasing the tool gradually and watching how it performs, which is far safer than switching it on for everyone. And customised versions stay private, so one person's experiment doesn't become another person's instructions.

There's a warning too. JournalArta reports that some subscribers want more control over how the feature reached them. People who already like your product may not want AI added to it. Make it easy to find, easy to ignore, and clear about when they're talking to software.

Before you add AI to anything

Ask three questions. What's the worst thing that happens if this answer is wrong? Where is the correct information kept? Who checks the answers in the first month?

If you have good answers, AI can save your team real time and give customers help at any hour. We built an AI tutor into EduBridge AI, and there's a working assistant on this site you can try with the 'Ask agent' button. If you're weighing up something similar, tell us what you want it to do and we'll tell you honestly which parts AI should handle and which it shouldn't.