A system in production is an operational commitment, not a closed project.
Release is not the end of the method: it is the point where the less visible part begins — observability, periodic review, adversarial testing, human oversight, incident management.
We don't publish a generic SLA or a guaranteed uptime that's the same for every project: they depend on context and on the system's criticality, and are defined with each client. What we can describe is the model we use to operate a system after release — the same mechanisms already applied on the production projects cited in our case studies.
Continuous observability
Accuracy and latency tracked over time, not only measured at handover: every system in production has a KPI dashboard comparing current results against the baseline measured during the Assessment.
Periodic review and knowledge updates
The typical cadence on production projects is quarterly: an accuracy review with operators and knowledge-base updates on every regulatory or procedural change. The exact cadence is agreed per project, based on the system's criticality.
Adversarial testing and red teaming
Before release, and on every significant architecture revision, systems go through adversarial tests aimed at compromising them — including prompt-injection attempts on content retrieved from documents, emails or external sources.
Human oversight on high-impact actions
Confidence thresholds below which a case is routed to an operator are calibrated per project on real data, not estimated on paper. High-impact actions remain subject to explicit human confirmation, regardless of how reliable the system has proven in testing.
Incident management and continuity
Escalation procedures and response times are defined for each project together with the client, consistent with their existing incident-management processes. Operational continuity depends on the deployment chosen: in the cloud it follows the provider's SLAs, on-premise it stays under the organization's direct control, in a hybrid configuration redundancy is spread across both layers.
Cost governance
For on-premise systems, infrastructure capacity and cost are sized to the real load measured in production, not to a consumption subscription estimated up front: it's one of the indicators tracked in the periodic review, alongside accuracy and latency.
Want to know how we would operate your system?
The operating model adapts to the process, the constraints and the criticality of your specific case: we define it together during the Assessment, not after the first incident.
Request an AI Assessment