Why agents are the future of AI

For 10 years we have been talking about AI as a revolution, we must recognize that the results are meager. The enthusiasm is there but operational successes are rare. There are many spectacular startup fundraisers, there are hundreds of articles in the press every day and all over the world. But for now, proof that AI will be able to produce value industrially and for everyone remains to be done.

Until now, AI is very often a matter of behemoths from ecosystems that are certainly well-informed by the media but difficult to access. Tools from Machine Learning are certainly numerous in our daily life but they remain the work of a handful of giants who, although having massively distributed these tools, have often failed to reassure as to their benevolence and their usefulness in daily life of ordinary mortals. So much so that some masters in the field fear a new AI winter: a long period (10 to 20 years) of disinterest during which we collectively doubt the relevance of this technology.

In our jargon, the AI-based software that we manufacture, that we code, to automate human tasks or more generally to solve operational problems (purchase predictions, fraud detection, price forecasts, automatic diagnostics, robots conversational…) have a name: they are the agents.

Since 2017, we have succeeded in scientifically demonstrating that agents were capable of solving real operational problems. For at least 7 years we have had clear examples of agents capable of producing a lot of value. This is true quantitatively when agents have demonstrated their ability to predict the presence of hydrocarbons in certain reservoirs rather than others and thus saved several million dollars in prospecting costs. This is also true qualitatively when agents helped doctors make the right diagnoses.

Until now, these agents, who were often laboratory rats, helped keep the fire of hope alive in AI but have never been able to escape the labs for the following reasons:

1. They were very expensive. The development of one of these software could range from 100,000 euros for the most basic to several million euros for the most complex.
2. They were very complicated to make and required bringing together teams of code superheroes.
3. Complexity very often and mechanically induced issues of reliability.
4. They were very difficult to explain. People were often asked to pay large sums of money for Blackboxes.
5. It was very complex to demonstrate its usefulness in daily life, especially for ordinary people.
6. It was difficult to share them within the same organization (companies, administrations, etc.) because in addition to the tool itself, it was necessary to develop aesthetic and ergonomic graphical interfaces. Which required a team and therefore additional complexities.
7. It was very difficult to guarantee data integration and confidentiality.

Recently, there have been automation platforms that provide solutions to these seven problems which until now prevented a massive appropriation of artificial intelligence by non-technical human profiles who have real skills. You can now get your data in lots of places, route it in a few clicks, build an agent by combining the colossal and available resources that exist in the “drag and drop” AI ecosystem. You put it all together in Lego mode, press a button and your agent is ready for your colleagues to use. Don’t like Lego? Never mind. Soon a robot will assemble them for you. The future is finally here. Will we be up to the task?