Do you want to know if the salmon, hake or any other fish you buy on the market meets the criteria for sustainable development? The Fish Me Well app will tell you. You take a picture of the fish and the artificial intelligence that triggers this application informs you if it is allowed to fish this season, of what size and warns if it is a pregnant female, whose capture is normally prohibited.
But don’t search for Fish Me Well on the App Store or Google Play. “This is one of the projects carried out by our students as part of their ‘Understanding AI’ courses at the Makers Lab, our lab dedicated to learning by doing,” explains Chloé Pelletier, Head of this course at EM Lyon Business School. particularity? The students who prototyped Fish Me Well aren’t data scientists or computer scientists: they didn’t write a single line of program. To develop their prediction model from fish photos, they used codeless AI, artificial intelligence software that requires no code.
A tangible economic reality
These platforms are constantly multiplying: the American Akkio (for data available in spreadsheet form), Lobe.ai (for image recognition, owned by Microsoft since 2018), Obviously.ai, the Chinese EasyDL (from Baidu), the French Prevision.io, the Swedish Peltarion… to name just a few. “No-code AI has been around for several years, but it only became a tangible economic reality two or three years ago,” observes Florian Douetteau, co-founder and CEO of Dataiku, a French company created in 2013, an expert in data science that, in turn, were thrown into the AI without code.
“The principle is very simple: in theory, with an AI-free tool, a business expert, marketing expert, sales, finance, logistics, production… , to apply some filters and adjustments and develop a predictive model in a few hours directly usable in your company: for example, to predict customers about to switch telephone operators”, explains Julien Laugel, chief data scientist (chief data officer) at MFG Labs, the entity dedicated to data processing and data science. , from the Ekino consulting group.
Lack of manpower
The current enthusiasm for no-code AI is obviously part of the more general trend towards low-code/no-code, which makes it possible to develop computer applications with minimal or no code at all, thanks to a drag-and-drop graphical interface. But it is also explained by the bottlenecks that companies are currently facing, more and more, wanting to get into AI. “It’s not a data availability issue,” emphasizes Florian Douetteau. Thanks to the falling cost of cloud platforms, companies can now centralize 90% of their information, at a reasonable price. »
What is lacking is the skilled labor to mine this material, which creates a vacuum for codeless AI. “Our typical client only has a small team of data scientists managing many projects,” says Jonathon Reilly, co-founder and COO of Akkio. So they don’t have time to build apps for salespeople, marketing or logistics managers, but with these tools these managers can manage on their own. “But there is a very large market for AI, Judge Khalil Alami, a data scientist who created a start-up, Quarks, which is expected to offer this summer a no-code AI application based on small, easily assembled components like Lego. We can imagine tomorrow using an AI to develop a complete marketing campaign, up to the production, in virtual images, of the advertising spot. Last advantage: all these solutions being online, it is not necessary to install additional computing power in the company.
Publishers of no-code AI solutions are, of course, praised for their algorithms, even if their handling is not always as simple as some claim (see sidebar). “We help companies and public sector organizations turn data into actionable insights much faster and more accurately than humans could alone,” swears Matthew Zeiler, CEO of intelligence platform Clarifai. to model image, video, text and audio data with little to no code.
Researchers are more cautious. “By allowing data to be manipulated without programming knowledge, AI without code has undeniable educational advantages, recognizes Jérémie Sublime, professor-researcher in data sciences and AI at Isep, a digital engineering school. But be careful, in business, not to play with fire: using this kind of tool without understanding its internal logic or its limits can make people believe that AI is a foolproof oracle and miss big societal issues such as possible biases in the data. »
What do users think? “We tried, without much success, to use the Ai no-code in complex problems of optimization of our metallurgical processes”, says Jean-Loup Loyer, data director of the French mining and metallurgy group Eramet. It is an educational tool that allows twice as many employees, non-AI experts, to adhere to the challenges of this technology, but in less ambitious projects. »
Another limit: the DSI (department of information systems) must give their consent before commissioning the final enterprise-wide application. Firstly, to avoid a “shadow IT” phenomenon, where we no longer know who does what, in terms of IT, within an organization. And validate an AI that, as a decision-making tool for certain executives, can have a formidable impact on the smooth running of the company. Finally, as with everything that happens in the cloud, the IT department often remains more competent to ensure that data is stored in France or Europe…
Advantages but also disadvantages
· A few hours of training are enough for a first introduction.
· The different stages of the data flow (availability in spreadsheet form, cleaning, choice and training of AI models, model verification, etc.) appear in a very educational way.
Also, it takes a few hours to choose a new dataset or new cleaning filters, a new AI model and see the resulting changes in the results.
· No explanation is given about the results obtained, which reserves these services for simple business cases and excludes any use for medical purposes, for example.
· Some platforms reserve the right to reuse their customer data to improve their performance, which could mean they can use it to retrain their AI models.
· All these platforms do not provide for a systematic control of their use; some malicious people could thus bypass image recognition to monitor their neighbors…
Tools not so easy to access
We tested the no-code AI platforms of the American Akkio and the French Dataiku, asking them to predict how often the short text that we will publish on LinkedIn to announce the publication of this article in “Les Echos” would be seen on this social network. network. Good news: the two AIs converge on the same estimate, i.e. around 1,100 “impressions” (go to my LinkedIn account to check this prediction…). Advantage: These tools are of operational interest, as a small interface allows you to enter a new version of the text and see if the forecasts increase or decrease. Disadvantage: The first use of these tools is anything but obvious and requires at least half a day of training and some basic knowledge about data filtering, error rates, etc. “These are tools that can be tamed, recognizes Eneric Lopez, director responsible for AI at Microsoft France. These pre-trained AI models are aimed at profiles that fall between business experts and data scientists, a bit like when Excel and Business Analytics software started to be democratized. »