The intelligent routing of application flows between, on the one hand, a fleet consisting of many equipments (fixed or mobile) equipped with a heavy client-side software or a light browser providing a service to end-users, and, on the other hand, a park of servers equipped with a server-side software contributing to this service, is a complex problem. A fleet may include, for example, users' computers accessing an e-commerce site, smartphones, payment terminals, ATMs and banking services, internet boxes, shopping terminals tickets, kiosks access public service, etc. The complexity of routing is inherent in the many constraints that this function should address. First, it must be adaptable to any client-side and server-side software via a single function point that interacts at the session and kinematic level. It must also take into account the compatibility matrix of the multiple deployed versions of the software. Second, it must be adaptable to any platform via a setting of topological rules incorporating the specifics of the software in the existing architecture, and allow a logical multi-criteria segmentation of the fleet and the park via a setting of rules of affinity and exclusion between any groups of equipments and servers (domains, markets, versions, service levels, etc.). Third, it must support an unlimited fleet with horizontal scaling capacity, and must dynamically enslave the load and load transfer with consideration of the overconsumption of resources generated by the software during rerouting, in order especially to avoid the cascaded spread of massive incidents (domino effect). Fourth, when necessary, it must be able to sacrifice certain flows, temporarily, to preserve a greater number of flows. Finally, the routing must offer a real-time vision of the state of the fleet and the park, and allow the automatic enrollment in the fleet of new discovered equipments. Because existing ADC (Application Delivery Controller) solutions insufficiently meet the requirements, this paper proposes a heuristic algorithm integrated into a reverse proxy, responding to the needs. This heuristic is presented in the context of automated banking machines, but can easily be transposed to any other context. Because the complexity of the problem does not allow to calculate routes in a reasonable time, the proposed heuristic does not seek the best mathematical solution but an effective and reactive routing that imitates social behaviour and maximizes availability, usability, quality of service. This inventive heuristic was implemented and tested successfully, jointly with a second invention playing the role of a distributed database (“A distributed, probabilistic, hysteretic, retroactive algorithm to locally weaken the CAP/Brewer’s theorem when reading data on a distributed computer system” ⟨hal-01966396⟩ https://hal.archives-ouvertes.fr/hal-01966396). Full understanding of this paper requires expertise in computer science, software engineering and algorithmics.