Operate & Adapt

Deployment is not the end. The human–AI system must be operated, monitored, maintained, and improved — continuously, not once.

AI systems change over time: models drift, data shifts, users adapt, requirements evolve, edge cases appear. A system that worked at launch may not work six months later. Operating means staying honest about whether the system is still doing what it is supposed to do.

The system that launched is not the system that runs. Operating means watching for the difference.

• Outcomes — is the system still producing the results it was designed to produce?
• Behavior — is the system behaving as specified, or has something drifted?
• Inputs — has the data the system sees changed in ways that affect performance?
• Usage — are people using the system as designed, or have workarounds emerged?
• Exceptions — what edge cases and failures are occurring, and how are they handled?
• Feedback — what are users, reviewers, and downstream systems reporting?

AI systems drift: model performance degrades, user behavior adapts, data distributions shift, requirements change. Monitoring must detect drift before it becomes failure.

Workarounds are signals, not problems. When users invent ways around the system, that often means the system does not fit the real work. The question is not "how do we force compliance?" but "what is the workaround telling us about the gap between the designed system and the actual work?"

When something goes wrong:

• Detect — how was the problem discovered?
• Respond — what immediate action was taken?
• Investigate — what caused the problem, and why did it reach this point?
• Remediate — what changes prevent recurrence?
• Learn— what does this incident teach about the system's design?

Incidents are not failures of the operating team — they are feedback about the system. The goal is not zero incidents; the goal is a system that learns from incidents and improves.

Requirements change. The world changes. The system must adapt. Adaptation is not a failure of the original design — it is evidence that the system is being operated honestly.

Build adaptation into the operating model: regular reviews, clear triggers for redesign, authority to make changes, and feedback loops that surface what needs to change.