ServicesOperate
AI systems drift. Source documents change, models get deprecated, usage moves in directions nobody predicted, and accuracy measured at launch quietly decays. Operate is a defined monthly service that keeps a production system accurate, adopted, and improving.
Start a conversation →An AI system is not a fixed asset. The policies it cites get revised. The product catalogue changes. A model version is retired and its replacement behaves differently on the same prompt. New staff use it in ways the original design never anticipated.
None of that announces itself. Quality degrades quietly, users lose confidence, and the system falls out of use months before anyone raises it as a problem. Operate exists so that decay is measured and corrected rather than discovered.
The evaluation set built during the build is re-run on a defined cadence, so accuracy is tracked over time instead of assumed.
Tuning retrieval, prompts, and configuration as usage patterns emerge, and assessing new model versions before adopting them.
Who is using the specific system, where sessions are abandoned, and where output is overridden or ignored, as evidence of system health.
Keeping access control, retention posture, and review paths current as policy, staffing, and vendor terms change.
An agreed monthly allowance for changes: refreshing sources, correcting failure patterns, and small refinements.
A regular working session on what the system is doing, what it is costing, what has degraded, and what to do next.
It is a defined scope at a fixed monthly fee, typically $5,000–$15,000 depending on system complexity, evaluation load, and the change allowance, with support during business hours. There is no uptime guarantee and no on-call obligation unless one is specifically contracted and priced. Improvement work is capped by an agreed monthly allowance; new capability is a new Build phase, not an extension of the retainer. Celadon would rather write that down now than discover the mismatch in month three.
Operate is available on systems Celadon built, or on an existing system that has been through a Sprint and is in a documented, understood state. Taking operational responsibility for an undocumented system nobody can explain is not a service Celadon offers, because it cannot be done honestly at a fixed fee.
Operate does not cover end-user helpdesk, infrastructure management for systems outside the scope, or unlimited development. It is not a substitute for internal ownership: someone in the business still owns the workflow and the decisions the system supports.
Operate may measure usage of the system it supports, but it does not run a role-based behavior-change program across teams or tools. That is Adoption. A live system can need both services, but neither silently includes the other.
If what you want is senior guidance across a broader AI program rather than care of a specific system, that is Advisory.
A focused conversation about what is live today, how it is performing, and what it would take to keep it accurate and adopted.
Tell us what you are evaluating, building, operating, or trying to get adopted. Email is the first step; if a call would be useful after we review the context, we will suggest one.
