AI agents in production: access, quality and contracts
In brief
At the Datadog Summit in Paris, leaders from Amadeus, Cegid and Datadog described the constraints of AI in production. According to them, an agent must stay within its intended scope, with authentication and authorization handled by a deterministic system. Quality depends on a use-case dataset, tracking of actual actions, and feeding monitoring lessons into the delivery cycle.
Source : Silicon.fr - « Les regards croisés d’Amadeus, Cegid et Datadog sur l’IA en production »
What it changes for an SME
For an SME of 10 to 250 people, an agent in production must be limited to what you designed it for, with deterministic identity and access controls, as Amadeus and Cegid stress. According to Cegid, a reply can look correct while a database write or a tool call failed, so you need to measure actions, not only text. You cannot reuse classic software service commitments as they are, and roles between product, engineering, security and the DPO become less clear. Cegid says an internal platform launched in 2023 was too early, and that services such as Bedrock, Vertex AI and Azure AI Foundry are now becoming more standardized.
What you can do
- Check that authentication and authorization for your agents are handled by a deterministic system, with least privilege.
- Build a test set from your real use cases and verify completed actions, not only generated text.
- Clarify who owns reliability, customer value and data-leak risk, and adapt your service contracts.
Let's talk about the time you could save.
In 20 minutes, we find where AI could save you time or money. Free, no commitment.