Having survived a similar frenzy in our industry (remember the quick commerce boom?), I wanted to take a few minutes to share some thoughts around AI applied specifically to food, since some consider that AI is going to make meal planning and groceries easy overnight, all with AI to thank. Know it alls! 😜

‍

Like many innovations, the current AI trend is at the same time both mind blowing and irrational. It’s mind blowing, for sure. In many of its current innovations (I’m looking at you, Dall-e) or potential applications, it is a game changing, paradigm shifting technology.

‍

In the same breath, we have to also realize that AI revolutionizing every industry overnight is irrational. This happens all of the time when we see big leaps forward in technology.

‍

In the past, many other innovations have found themselves at the top of the hype-cycle before (relatively quickly) teetering off. AI global applications will take a few more years to reach their plateau of productivity (see Gartner hype-cycle).

‍

Don’t get me wrong, AI will be one of those evolutions that happens every 20 years. It’s monumental, but the current AI-media-takeover should not lure us.

‍

The potential is big. It looks like magic; but once you have actually taken the time to project the capabilities, you still need to put it into fruition alongside real life constraints. A promising teaser does not necessarily guarantee a great movie.

‍

The question then becomes, what would it really take for AI to disrupt the food & grocery space?

‍

Meal planning and shoppable recipes: The future of groceries

‍

We believe that the future of groceries is shoppable recipes: A technology recommends a meal plan tailored to your needs and then automatically shops for all the ingredients you need. A dream right? Well, it already exists. It’s Jow; and we’ve already served 150 million meals with our service. With this history, I think it’s also fair to assume that we understand what the tech requirements are to make shoppable recipes work.

‍

Let's rephrase the question, and it’s an important one: Can AI enhance or replace the way we operate Jow?

‍

Lately we’ve seen many AI based use cases around shoppable recipes and meal planning. You ask AI’s grocer X for a meal plan for 3 people, containing 5 meals, and it answers with a menu idea, then a list of ingredients translated into products you will be able to shop in a click. But does it work, longterm?

‍

‍

How do you make a recipe shoppable and how could AI be useful along the way?

‍

It’s quite simple in theory. You need to translate a recipe into its ingredients, and then match the ingredients with the corresponding products. This sounds feasible from a tech standpoint, and even more so with AI’s help. And for the most part, yes, it is an easy equation for an AI engine. However, in the tech world (and also AI) there is a simple universal rule called SISO: Shit in, shit out. You can have an amazing sports car, but if you can’t drive, it is useless. The same is true for tech. If you don’t have the right data to “feed” your AI, the output just won’t be good. And the right input, in the case of shoppable recipes, does not exist on the market.

‍

‍

Product metadata: the shit in = shit out problem

‍

On everyday products, the metadata available isn’t currently standardized or normalized across  markets and grocers. Depending on the source, it isn’t formatted the same way, it isn’t exhaustive, it’s expressed in different units, etc… but more importantly: the existing metadata almost never contains information around real life usage.

For example, you might find the weight of a product, but not the cooking time. You might have the nutritional facts, but not the right portion size per person. Those are just a couple of examples about the many details we cannot see on a product package, and likely are not available in the metadata.

‍

To put this into perspective, say you want to cook two meals for a dinner party and both of the meals you’re thinking about contain chicken. One meal is a chicken filet, the other is chicken fajitas. In the first case, you need to make sure each guest gets an actual filet (with a variable weight depending on the brand/packaging/source…). You don’t want one guest to have 3/5 of a filet because the portion has been calculated by an algorithm. In the second case, you need 10 ounces of chicken for your fajitas, the packaging of the chicken doesn’t really matter much. It can come from 2 or 3 filets because you’re going to cut it up anyway.

‍

The real question here: can AI manage this problem? Well, it probably could, but only if it has access to all the info/use cases/edge cases available, and most of the time, that information isn’t available. As smart as AI is, it can’t compute the void.

‍

The same rings true for ingredients. You can absolutely scrap an ingredient list from a recipe, but the same issue exists: what kind of metadata do you have around it? Is this product seasonal? How do you know that a given ingredient could be replaced by another one in recipe A, but that it wouldn’t work as a relevant replacement in recipe B? You wouldn’t know this unless it’s been encoded by someone.

‍

Same with the recipe. If you just read/scrape a recipe with AI and see a cooking time of 15 minutes for 10 ounces of spinach, how does AI know that the spinach’s weight and cooking time do not translate the same way depending on whether the spinach is fresh, canned, or frozen?

‍

The tldr: if you do not have the right data model, the right data and metadata, and the right input, even the best AI won’t be enough. Sure, it will manage to create some generic suggestions and also the illusion of magic, but over the long run it won’t be a solution that stands the test of time and the consumer will find themselves hitting a wall.

‍

At Jow, not only did we build the best possible algorithm to make recipes shoppable, even sometimes using AI, but we also, first and foremost, encoded food.

‍

Each recipe you see on Jow has been designed by Jow, for Jow. It’s been ideated, researched, cooked, shot, described, and encoded in our database, manually. As we grow, we continue to encode food, which has become a proprietary homogeneous, standardized recipe/food and product dataset. From recipe video tutorial time codes to product barcodes - it’s an exhaustive dataset that is unique on the market and in the industry.

‍

‍

To be useful, AI first has to be usable: UX is the defining seat at the table

‍

Let’s add another layer to this. Say you have a crazy good AI-based tech solution to make recipes shoppable. Not only that, it might even be able to create meal plans. That’s wonderful. But it fails to address the most important need of the consumer: the actual customer experience. It’s as important, or maybe even more important than the underlying algorithm results and efficiency.

‍

If you look at the way we’re able to interact with AI, to date it is mostly text or voice.

‍

Sound familiar? Voice: remember those things called Siri, or Alexa? Text: remember this big hype around chatbot a few years ago? Not one of these technologies actually reached a crazy good level of adoption. Many could argue that the underlying tech or output just weren’t good enough. Now, for sure they are and were weaker than the new LLMs, but it’s also fair to say that the input and interaction modes (voice or text) were actually not the best possible UX.

‍

When developing a meal planning and grocery shopping service, experience is key. Grocery shopping takes hours, even online. You want to make it go faster, not add a new process on top of the already lengthy process of grocery shopping. Additionally, do you let the user pick each recipe one by one? Do you create a direct meal plan? Once created, how do you adjust it? How do you remove/replace recipes? Etc., etc.

‍

When you look at the variables, 4 to 10 customer prompts or voice commands don’t sound like the right approach. Not to mention the fact that pure text or audio solutions might not be what a consumer would expect to see when looking for tasty recipes (a photo or video may be more appealing ;-))

‍

In summary, even after solving the technical input/output equation, building an outstanding customer UX remains a critical factor, accounting for at least 50% of the job to achieve successful product adoption.

‍

‍

The future: how can grocers use jow and AI to reinvent their shopping experience?

‍

It took us five years at Jow to refine our technology and UX. It involved an internal process of actually cooking and encoding recipes one by one, A/B testing thousands of customer journeys, approaches, and interactions. And ultimately, it has resulted in 150 millions meals served and counting.

‍

AI is amazing and it will help us in the coming years to bring our service and company operations to the next level. Let’s grab this fantastic opportunity, but let’s stay wise: the FOMO is the fuel. The hard work and domain expertise remain mandatory to create the best product. Embrace the innovation, but do not be lured by its perceived magic.

‍

AI combined with Jow’s unique dataset, technology, and UX creates a one of a kind customer journey that has already been adopted by 6 million users in France and in the USA. We are happy to share our expertise and leverage AI with our partners to continue to reinvent groceries, not the wheel.

‍

‍

Key takeaways:

‍

• Shit in = Shit out: AI works best when it has access to amazing datasets, and if recipes are perfectly described from a content perspective, then they’re barely encoded with enough detail to make them truly shoppable… unless you use Jow’s unique catalog. We solved this issue. ;-)

• AI is an undeniably amazing tech enabler, but the key to adoption is UX: If an AI powered tech offers me an extremely relevant menu recommendation/shoppable cart, I’ll still need to interact with it in some capacity. Prompt based interactions and “conversational mode” are probably not the best UX you can offer. UX is as important (or may be even more important) than the efficiency of the underlying AI.