We Deployed an AI Agent to Production. The Hard Part Wasn't the AI.
The first AI agent we put into production worked well in staging. In production, it made decisions that were technically correct and operationally disastrous because no one had defined what it could do on its own. Since then, we design permissions, boundaries and accountability first. The model comes later.
What Broke in Production
The team had spent weeks on prompt engineering and almost no time on logging. When the agent started making mistakes, nobody could reconstruct its inputs, decisions or actions. Since then, every critical step leaves a human-readable trace and every irreversible action requires human approval.
Before discussing models, we ask who is accountable when the agent takes an unexpected action. If nobody can answer immediately, the system is not ready for production. Every critical step needs an owner and an explicit human checkpoint.
A client approached us to 'add an AI agent to their workflow'. After analysis, we told them their workflow wasn't structured enough to be automated by a human, let alone an agent. We started from there.
Rules for Interrupted Work
The production failure was an agent stopping halfway through a critical process, with no rules for resuming the workflow or reversing completed actions. Before release, we define those paths and who needs to intervene.
Each workflow has an owner, a list of permitted actions and a precise point where a person takes over.
Related Services
If you're evaluating a similar project, explore our approach to AI development in Turin. When the agent needs to use business tools, read data or trigger controlled operations, the work also involves API, MCP and CLI integrations. For the broader architecture, see how we work as a software company in Turin.