A Distributor Values Its AI Agents at 20 Times the Traditional Cost

A regional distributor that deployed an agent-based system for its business transactions estimates the cost to be about 20 times that of a traditional architecture, six months after going live. It notes a return on investment that is roughly equivalent. The system is functioning correctly, but its integration required more work than anticipated.
Agents Require Operational Resources
Access to the model is only part of the needs of an agent-based system. This also includes instruction management, tool coordination, and evaluation mechanisms. Security, action traceability, data governance, and human intervention capabilities also require appropriate operational resources and skills.
These systems can coordinate complex processes, reason about data, and mobilize tools for tasks that were previously difficult to automate. However, their behavior includes a non-deterministic element, even when the business process requires a deterministic response. Simple automation can thus become unnecessarily costly.
The Distributor's Project Required More Integration
The distributor, whose identity is not specified, needed a new transaction management system. The scope included part of the order processing, discount validation, exceptions, customer exchanges, and inventory control. The consulted suppliers presented agents that they claimed could decide, interpret customer intentions, and coordinate tasks based on the business context.
One of them compared their shared language model to hiring 10,000 experts without any salaries to pay. This argument received a favorable reception from the board of directors. The chosen architecture was to entrust the agents with decision-making, tool invocation, internal policy control, and interactions with the company's systems, with human intervention when necessary.
Integration exceeded expectations: the agents required clean data, governed APIs, business rules, protections, and observability. The team also built mechanisms for human recovery, protections surrounding decisions, and logs tracing the reasoning. Testing devices for unpredictable responses and new operational procedures completed this set. Expenses rose significantly, security checks took longer, and daily operations became more complex.
The economic assessment includes all support mechanisms, not just access to the models.
The Choice Depends on the Process to Automate
Interpreting a request, adapting to context, and partially autonomous action are the use cases for which agent-based systems present the most interest. A stable process governed by rules, heavily regulated, or easily translatable into classical logic offers a less convincing justification. A successful demonstration at a conference can precede a costly, fragile, or difficult-to-govern deployment.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Evaluation can focus on saved labor, customer relationship quality, exception handling, or a capability that traditional software could not provide at an acceptable cost. These benefits must be measurable: a general promise of automation does not constitute a measure. A process engine, a rules platform, API integration, or a conventional application provide points of comparison.
In some cases, a classic solution could achieve the same result for a tenth of the cost. The value of agents might justify their cost in certain situations, without justifying it in many others.
Services and Microservices: Benefits and Costs
Service-oriented architecture relied on capabilities distributed among services and exposed through contracts. It aimed to reuse business functions, limit duplication, and facilitate evolution. It also encouraged a design based on business capabilities, distancing itself from heavily coupled monolithic applications and point-to-point connections.
However, this approach experienced service buses that became congestion points and heavy governance. Discussions about contracts, standards, registries, and middleware consumed time that could have been spent delivering business functions. The implementation cost could then exceed the expected benefit of reuse. Many of its principles persist in APIs, integration platforms, and service-oriented design, while the AOS label has declined after overly large, slow, and costly projects.
Microservices offered another form of autonomy: small services, independent teams, separate deployments, and targeted scaling. They performed well in organizations with strong DevOps practices, mature automation, well-defined domains, and significant volume needs. Part of the system could evolve without necessitating an evolution of the whole.
Conversely, breaking ordinary applications into dozens or hundreds of services shifted complexity to the network. It introduced distributed failures, data consistency issues, and difficulties in monitoring and testing. Insufficiently prepared companies faced heavier operations, increased cloud spending, more meetings, and slower troubleshooting.
Kubernetes Promises Elastic Capacity
Containers and Kubernetes have brought the promise of cloud-native applications: portable workloads, declarative deployments, elastic capacity, and automated recovery. These models can enhance resilience, portability, and deployment speed. Their relevance particularly concerns large digital enterprises, platform teams, and applications with highly variable operational needs.
However, insufficiently justified migrations have added clusters, service meshes, observability tools, and security layers. They have also required platform teams and specialized skills. For systems that do not require this abstraction, operating costs have, in many cases, nearly doubled.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.