Most writing about enterprise AI ends at go-live. The pilot succeeded, the system is in production, the project is declared done. Yet the months after go-live decide whether a system keeps improving, holds steady or quietly drifts until people stop trusting it.
MIT's 2025 research found that most generative AI pilots never produce measurable business impact. A companion piece covers how to scope a pilot so it makes it to production. This one covers what happens next.
Production is a different environment
A pilot runs on chosen cases, with an engaged team watching closely. Production brings everything else: unusual inputs, busy periods, new products, changed prices, staff who were not part of the pilot, and the cases nobody thought to test.
Google engineers captured this in their ML Test Score, a rubric for production readiness. It checks not only the model, but data quality, infrastructure, monitoring and the ability to roll back. Their finding was that teams consistently underinvest in monitoring, the part that matters most after launch.
Systems drift, even when nothing breaks
Researchers call it concept drift: the world a system was built for changes, so its performance changes too, without any error message. Gama and colleagues' survey of the field describes several forms. Inputs can shift, as when a new supplier sends documents in a different format. Meanings can shift, as when a regulator changes a rule. Patterns can change suddenly, gradually or seasonally.
In operational AI, drift is ordinary. Prices change. Product ranges change. A customer segment grows. A new form version appears. None of these breaks the system. Each can quietly reduce its accuracy if no one is watching.
The first 90 days
IT service management has long recognised that new systems need extra attention after launch, a period ITIL calls early life support. For agentic AI, the first 90 days follow a recognisable pattern.
Weeks 1 to 2: close watch. Every output is reviewed. Volumes are often limited. The aim is to confirm that production behaviour matches what testing showed, and to catch integration problems early. Expect a burst of small configuration changes as real inputs reveal cases the pilot missed.
Weeks 3 to 6: patterns emerge. Review moves from every case to samples and exceptions, where the evidence supports it. The exception log becomes the most valuable document in the deployment: which cases go to people, why, and how they were resolved. Recurring exceptions point to configuration that needs adjusting or rules that need writing down.
Weeks 7 to 12: steady state and the first widening. Quality is measured by case type. Where the evidence shows reliability, autonomy can widen for those cases, while others stay under closer review. The team settles into a routine of weekly review, and the system's knowledge grows with every case it handles.
What to monitor
Quality by case type: acceptance, correction and rejection rates, not one overall figure.
Exceptions: volume, mix and resolution time, and whether the mix is shifting.
Overrides: where people disagree with the system, and why.
Inputs: new formats, sources or values the system has not seen before.
Cost and speed: per case, compared with expectations.
User experience: whether the people working with the system trust it, and where they work around it.
Feedback is how systems improve
W. Edwards Deming's plan-do-check-act cycle is the oldest model of continuous improvement, and it fits production AI exactly. Plan a change. Make it. Check the effect. Act on what you learn. The essential ingredient is feedback that reaches the system. A correction made in the review queue should not disappear. It should add to the knowledge the system draws on, so it handles the next similar case better.
When that loop works, a system improves month by month. When it does not, people correct the same errors repeatedly, lose patience and stop using it.
Regulation expects monitoring after launch
For high-risk AI systems, the EU AI Act requires providers to run post-market monitoring: collecting and reviewing performance data throughout the system's life. After the Digital Omnibus, those duties apply from 2 December 2027. Organisations that build monitoring into every deployment now will meet the requirement without rebuilding later, and will run better systems in the meantime.
Signs a production system is healthy
Exceptions are falling for case types that have been reviewed and configured.
Overrides are rare and explained.
New input types are spotted within days, not months.
People trust the system enough to stop double-checking routine work.
The team can say, with evidence, which cases are ready for more autonomy.
Questions to settle before go-live
Who owns the system day-to-day in production?
What will be reviewed in the first two weeks, and by whom?
How will corrections reach the system's knowledge?
What would trigger a rollback, and how fast can it happen?
When will autonomy be reviewed, and on what evidence?
How Mitochondria's agentic AI systems run in production
The approach. Synthesis, the third deployment stage, is the early-life period: the system runs on live work under close watch, and we investigate and resolve every anomaly before scope widens. Energy, the fourth stage, is steady operation within agreed limits, with monitoring and review built in.
What it enables. Quality is tracked by case type, exceptions and overrides are logged, and new input types are flagged. Every correction a person makes, and every document processed successfully, adds to the knowledge the systems draw on, so they improve with use.
The impact. Systems that get better after go-live rather than drifting, and a clear, evidence-based path to more autonomy where it is earned.
Read more about Mitochondria's approach.
Sources
Breck, E., Cai, S., Nielsen, E., Salib, M. and Sculley, D. (2017). The ML test score: a rubric for ML production readiness and technical debt reduction. IEEE International Conference on Big Data.
Gama, J., Žliobaitė, I., Bifet, A., Pechenizkiy, M. and Bouchachia, A. (2014). A survey on concept drift adaptation. ACM Computing Surveys, 46(4).
Deming, W. E. (1986). Out of the Crisis.
AXELOS, ITIL 4 (early life support).
Future of Privacy Forum, AI Act timeline after the AI Omnibus
—
Mitochondria builds agentic AI for evidence-led core operations. Research-led and model-agnostic. ISO/IEC 27001:2022 certified, with controls aligned to the GDPR, UK GDPR, India's DPDP Act and Australia's Privacy Act. Amsterdam and Pune. Request a working session →