Every AI project has a launch day. There is a demo, a round of congratulations, perhaps a slide in the quarterly update. Then the project team moves on to the next initiative, and the model keeps running. It scores applications, routes support tickets or predicts demand, day after day, with far less attention than it received during its first few months. Ask a year later who is responsible for it and the answer is often a pause, followed by the name of someone who left the company.
This is the orphaned model problem, and it is more common than most boards would like to admit. Traditional software ages too, but it tends to fail loudly. A broken integration throws errors, a crashed server sets off alarms. Machine learning models decay quietly. The data they see in production slowly drifts away from the data they were trained on, customer behaviour changes, a supplier alters a feed format, and accuracy erodes a fraction of a percent at a time. Nothing breaks. The decisions simply get a little worse, and nobody is watching closely enough to notice.
Much of what now gets called AI governance is really an answer to this question of ownership. It is less about ethics statements on the website and more about unglamorous structure: a named owner for every model in production, a documented purpose that says what the system is and is not meant to do, agreed thresholds for performance and fairness, and a clear route for someone to raise a concern. Organisations that get this right usually treat models like any other business asset with a lifecycle, from approval through monitoring to eventual retirement.
Retirement deserves particular attention, because it is the stage most often skipped. Models are rarely switched off deliberately. They are replaced by newer versions while the old ones continue to run in a forgotten corner, or they outlive the business process they were built for. A sensible governance framework includes criteria for decommissioning, as well as for deployment. It also keeps a record of past versions and the data they used, so that if a decision made two years ago is challenged, the company can reconstruct what the system knew at the time.
None of this has to become bureaucratic. Smaller organisations can start with a simple register and a monthly review of the handful of models that matter most, focusing on the ones that affect customers, employees or money. Larger ones may need a formal committee and tooling that tracks models automatically across teams. In both cases, the point is proportionality. A model recommending blog posts to readers does not need the same scrutiny as one assessing creditworthiness, and treating them identically only wastes effort that should go to the risky systems.
The regulatory pressure in Europe and elsewhere has certainly pushed the topic up the agenda. Yet the strongest argument for governance is not the fear of fines. It is the simple fact that a model nobody owns is a decision-maker nobody supervises, and few businesses would accept that arrangement from a human employee for very long.
Read also Christmas Falls On A Friday In 2026 So Rethink The Host Gift
