In the first edition of my Focus on AI newsletter series, designer Derek Fridman and I discussed how AI could someday transform the shopping and commerce experience, from digital interfaces to in-store interaction.

Today: A look into an existing e-commerce business that’s already built on top of AI and machine learning — one whose customers rely on it to feed themselves.

First, a word from Work & Co, my partner and sponsor for this series. Check out their event this week:

Online grocery shopping is one of the pandemic-era consumer behaviors that actually stuck, moving tens of billions of dollars in annual spending to e-commerce. Some of that spending has shifted away from the traditional grocery retail industry (and intermediaries like Instacart) to digital-first players with new ideas and interfaces.

One of those winners so far has been Hungryroot, a nearly decade-old online grocery startup that takes a different approach to food shopping: The majority of the groceries it sells and ships across the country — 70% — have been proactively picked for its customers by algorithm, instead of relying on consumers to search or browse for specific items.

Specifically, Hungryroot is obsessively focused on customer retention, using machine learning AI to predict which assortment of meals and groceries will keep customers hooked for each subsequent shipment, its founder and CEO Ben McKean tells me.

Last year, the company grew 40% to $333 million in revenue, according to a release, generating $9 million in profit. That growth rate far outpaced the broader US online grocery market, which declined 1% last year (to $96 billion), according to the Brick Meets Click / Mercatus survey. And one other comparison: The much larger Instacart grew its marketplace gross transaction volume by 5% and total revenue by 19% last year.

Hungryroot customers start by filling out a detailed onboarding questionnaire, providing information about their health goals, dietary needs and restrictions, specific foods and cuisines they like and dislike (eggplants? Korean? Beyond Meat?), their spice tolerance, their budget, how much time they want to spend cooking, the kitchen tools and appliances they have, and how they eat, such as their breakfast and snacking routines.

It also asks them to provide feedback on specific photos of meal ideas and formats: What looks good for lunch and dinner?

The Hungryroot algorithms use this information to build an initial box of meals — simple recipes with three to four ingredients — and additional items.

On the back end, McKean explains, among other actions, Hungryroot is “clustering” its new customer with other users who have answered its onboarding survey similarly and have already been with the service for multiple years. “And so we can say, ‘okay, people who filled out that signup flow like you… they loved these top recipes with high probability, so we think you’re going to love these recipes with high probability’.”

I just went through the onboarding process, and Hungryroot built me a box with three recipes to make over the course of a week: A zesty beef and rainbow veggie rice bowl, a cheesy honey herb baked chicken, and juicy chicken and rosemary potatoes. I can say with certainty that I would not have made any of these Frommer-optimized meals on my own next week, so I’m curious to try them.

(The Hungryroot algo also picked me a bunch of other things for breakfast and snacks: Sous vide egg bites, strawberry fruit bites, a buffalo chicken jerky stick, chocolate ganache desserts, and more. I’ll try most of these, too, and the Hungryroot-developed almond chickpea cookie dough that is also coming — “literally the best cookie dough you’ll ever have,” McKean says. “Hugely popular.” But I’ll pass on the RTD mocha latté protein shake it suggested.)

One unique element — relative to a more traditional online grocery service — is that Hungryroot is really trying to use the meal or recipe as the atomic unit of grocery shopping, as opposed to an individual ingredient.

It isn’t what you’d think of as a “meal kit” service: Hungryroot’s produce and protein seems as you’d buy it in a store, not pre-cut or re-portioned (though it does offer heat-and-eat meals if you want them). And while a meal kit service may offer a few dozen selections per week, “we literally deliver 10,000 unique recipes” each week to customers, McKean says.

But this isn’t normal grocery shopping, either.

Even looking at your completed shopping cart, you have to click in to see what specific grocery items and brands are coming as part of each meal — they’re not fully itemized in the cart. That cheesy chicken recipe, for example, includes True Leaf Farms baby spinach, Andrew + Everett shredded mozzarella, Coleman pork bacon, Mission Driven boneless chicken breast, and a seasoning blend by Spicewalla.

You also can’t pick and choose brands — Hungryroot only has around 1,000 SKUs at a time, a tiny fraction of a grocery store’s — though you can remove items from the meal and manually add more to your cart.

And you don’t even see prices: Hungryroot uses “credits” to price a meal or item. This one happens to be 20 credits, the highest of my three.

One benefit here is that, by removing the customer from a big portion of the decision-making, in addition to saving someone time and mental overhead, and introducing people to new foods or brands, Hungryroot can adjust its recommendations based on what’s in stock or what it’s buying at any given point. An individual’s prescribed cart for the week could change several times before the customer confirms, McKean says.

“The algorithm is effectively our supply planning tool,” he says. That is, it can forecast what people will buy, and therefore how many of each item Hungryroot will sell over the course of a week.

It’s also optimizing to minimize food waste, which — in addition to being a good thing in general — also lowers costs. While a typical grocery store may have 10% spoilage, McKean says, Hungryroot has 2%.

“The reason for that,” he explains, “is there’s this concept of a ‘band of indifference’. So if you prefer broccoli over brussels sprouts, the algorithm is going to give you broccoli. But if you’re within the ‘band of indifference’, which a lot of people are, it’s going to give you the item that we have more inventory of.”

The highest goal of all this machine learning, again, is retention: What will, statistically, get someone to stick around for another week? (Hungryroot aims to be your primary grocery service — its average order is around $125.)

Sometimes, McKean explains, it’s by giving people want they say they want. Sometimes it’s by giving people something different.

“Let’s say that you love salmon — you’ve explicitly told us that, and you have ordered salmon in your last four deliveries,” he says. “As a human, I think we would probably say, ‘oh, let’s give him salmon’. But what the machine learning model picks up, with the highest probability: Even if you love salmon, if you’ve ordered salmon in your last four deliveries, you don’t want salmon in your fifth delivery.”

“It looks at all of the millions of data points of how our customers have ordered in the past, and it looks at people have told us that they love salmon. And people who have ordered salmon four times in a row, what do they do in their next delivery. And the ones that don’t order salmon retain at statistically higher rates than those that do.”

McKean says this is working — that the company’s customer retention has improved 50% over the past two years, as its ML talent and tech have improved. (It’s notably not yet using generative AI for its interface, unlike other companies’ experiments. But it can always add that later.)

Is this the future of grocery shopping?

It is definitely different, relative to how most people shop in stores and online — a combination of re-purchasing known staples and discovering new products and brands through random luck, word of mouth, and promotional placements.

I am going to try it, but I’m not sure I’ll love it. Grocery shopping is entertaining and meditative for me, in addition to being research for work.

But I believe this is a potentially great service for a lot of people who are either too busy to shop, indifferent about shopping, or really need to stick to a specific diet or plan. (In our latest Consumer Trends Survey, most Americans say they actually like shopping for groceries. But around 10% say they don’t, which is a lot of people and spending.)

So I would bet that AI-powered recommendation engines like Hungryroot’s will increasingly become a core part of online grocery shopping over time. Whether it’s a feature of most services — likely, I’d think — or unique to a few merchants, we’ll find out. Instacart, I’d note, has already started to integrate AI-powered recipe suggestions into its app, but it’s not nearly as core to the experience.

Where I’ll give Hungryroot credit, even before receiving my first shipment, is for getting me out of my comfort zone. Almost nothing it recommended me — with confidence! — is stuff I typically buy or cook.

“A typical American, when they go into the grocery store, 75% of what they buy are the exact same items is as their last purchase in that grocery store,” McKean says. (Guilty!) “So there’s very little variety in the typical American diet. At Hungryroot, less than 10% of what our customers buy are the exact same items as the week before.”

Whether that keeps me as a member or not, we’ll see.

Zooming out, this idea of eating by algorithm isn’t new — grocery store shelves and digital placements have always been decided by some sort of logic, of varying sophistication. On other online grocery services, many people purchase a significant portion items because of algorithmic recommendations, organic and paid; Instacart’s advertising revenue alone was nearly three times the size of Hungryroot’s entire business last year. But Hungryroot is really going for it as the main thing.

As someone who’s been covering the tech industry for nearly a couple of decades, and has seen what algorithmic consumption in other areas — news, for example — has done to society, this broadly raises flags. The last thing we need from the retail media boom is a more efficient way to funnel Cheetos and Oreos into Americans.

So one of the Consumer Trends Survey questions I’m most passionate about is one we’ve asked a few times: Whether consumers think online grocery services have a responsibility to recommend healthy foods. The majority of younger consumers — Gen. Z and Millennials — always say yes, online grocery services do have a responsibility. The majority of their parents vote no.

What seems useful about Hungryroot is that it’s specifically designed to help people achieve health and wellness goals, not just make carb consumption easy for the laziest.

And an internal survey — so, possibly biased, but still worth noting — suggests it’s working. According to the survey, 90% of those polled said they have seen progress in their health goals — things like feeling less stressed, more energized, and more easily managing their blood sugar.

AI on my radar

OpenAI is making some product updates and announcements today at 10am PT, streaming live on its homepage.

For the past few weeks, the top VC firm Union Square Ventures has been using an AI assistant to record its internal conversations, which then “synthesizes themes from them, and shares them with the world.” Here’s the summary for the week of April 22, April 29, and May 6. This is a neat experiment, but it mostly makes me wish I could watch the video of these chats on YouTube (or, better yet, be in the room).

The newly published Principles for using generative AI in The New York Times’ newsroom, by Zach Seward, my former Quartz colleague who recently joined the NYT as editorial director for AI initiatives.

Slop, the perfect new word for the AI-generated words that are starting to show up everywhere online.

Yelp is testing AI-generated restaurant-overview videos based on its user-submitted text, photo, and video content. (Via Expedite.)

Restaurants using Toast’s software will have access to an AI-powered writing assistant to help with email marketing campaigns; not clear if it has any special domain expertise or is just a part of the future where every text-entry field on the internet has an AI-powered writing assistant. Still, a potentially efficient tool to assist with demand generation, for an industry that could mostly use the help. (Also via Expedite.)