Succeeding in AI requires acting quickly: starting from a concrete use case, testing and iterating rather than theorizing everything, in order to quickly generate ROI and buy-in.
We have not launched an agentic AI project because we prefer to clearly define our needs and our specifications beforehand. “Bad idea to guarantee the successful adoption of AI! Writing a long description of needs, based on an existing one, more or less well known, and on a “magic wand” AI “fantasy”, disconnected from the daily reality of workflows as well as from technical feasibility, is a waste of time. Only “taking action”, on a case of concrete and impactful use, positions companies to win the battle for competitiveness.
Management and management generally do not know the details of employees’ activities, which constitutes a first obstacle to truly understanding automation and optimization opportunities. In a service company whose business consists of intermediating large volumes of transactions between its customers and suppliers, the “bottom up” analysis of processes as part of the implementation of AI agents highlighted a step that was particularly revealing of this reality… An external service provider billed millions of euros per year to reconcile the invoices received with the correct supplier account and resolve ambiguities. However, the very existence of these ambiguities regarding the identity of the supplier resulted from poor management of the database which contained numerous inactive, redundant or obsolete third parties. This task was simply eliminated after analysis!
In another vein, a service company wanted to optimize the deployment of its tens of thousands of employees on thousands of intervention sites. The creation of this AI scheduling and dispatch agent required nearly ten practical iterations, confronting management and management with the schedules produced by the AI, in order to make the “preconscious” business rules explicit. From these iterations emerged new “hard/mandatory” or “soft/preferential” constraints (such as the number of different employees who could intervene during the year on the same position, to ensure maximum quality of service to the client) or details on certain parameters (such as the actual costs of training and onboarding of an employee on a client site, much lower than the theoretical costs).
In conclusion, it is difficult to understand current operating modes “in the room” at a sufficiently granular level of detail as well as to formulate the expected optimizations in precise algorithmic terms.
Alongside this difficulty in “knowing what we want”, it is even more difficult to imagine what we can! The technologies available today are numerous and evolving. The generative AI of consumer chatbots that everyone uses is only the tip of the iceberg. Many other technologies are available and must be articulated to replace a human in carrying out tedious tasks or augment a human in higher value-added activities and complex decision-making.
We can use multi-modal models (analysis, improvement or generation of image, sound, voice, video), machine learning, different under-constrained optimization models, or even agentic autonomous reasoning technologies. Imagining what it is possible to do by stringing together the right technologies, with the appropriate instructions and contexts, is more or less easy ex ante. But above all, the “pipeline” of technologies thus constituted must be put to the test of the facts: the validity of the results obtained must be evaluated on concrete cases to understand what it is really possible to do, under what conditions, with what latency (speed / responsiveness) and for what costs.
Defining a scope (a persona in the company, a workflow, a problem) and taking action quickly, within a defined strategic and technical framework (regulatory constraints, technological choices, etc.) makes it possible to obtain tangible ROIs quickly, to build trust and gain buy-in, as well as to validate the strategic and technical bases, which opens the way to faster deployment of future use cases. The drafting of specifications (by definition theoretical) and the classification of use cases in spreadsheets must quickly give way to a more agile approach more adapted to the reality of AI technologies.
