A model that has been making slightly worse decisions for months, and nobody noticed. Someone built it, it worked, and then the person who set it up changed teams or left. It kept running. No one was checking whether the numbers it produced still held, because on the surface it looked like a working system. It was drifting. And the first people to spot that may well be a buyer’s diligence team, three weeks into picking your business apart.
That system has a name in deal circles now. It is an AI orphan, and it is worth understanding before you have one, or before the one you already have shows up at the worst possible moment.
What is an AI orphan?
An AI orphan is an AI system that has no clear owner, no useful documentation, and no monitoring. It typically began as a sensible experiment that delivered something useful, so the business kept using it. Over time a real process came to depend on it. Nobody ever formally took responsibility for keeping it healthy, so it now runs on its own, unsupervised, with no one answerable for whether it still works.
The label matters because orphans hide in plain sight. They look like assets. A forecasting tool, a lead-scoring model, a document classifier, a pricing helper, all of these can sit in the operation for a year doing their job, while the person who understood them has moved on. The Reed Smith red-flag guide for buyers and boards describes the classic case as a system multiple teams rely on while no one can name who is accountable for it. That gap is the orphan.
Why do unowned systems degrade?
They degrade because a model is shaped entirely by the data it was trained on, so when the data flowing through it in production drifts away from those original conditions, its accuracy slips. A study from researchers at MIT, Harvard, and the University of Monterrey found that 91 per cent of machine learning models degrade in performance over time, an effect they call AI aging. Decline is the normal case.
The same study documented two things getting worse together, the error rate and the error variability. So an ageing model is not only wrong more often, it is wrong more unpredictably. The gap between its best day and its worst day widens. NannyML, which surfaced the research, calls this model performance deterioration, and it happens silently unless someone is watching for it.
Watching for it is the whole job. Effective model monitoring works across three layers, as WitnessAI sets out. Data monitoring tracks whether the inputs still look like the inputs the model was built for. Performance monitoring compares what the model predicted against what actually happened. Operational monitoring watches the plumbing, latency and throughput. An orphan has none of these, which is exactly why its decline goes unseen.
Where will you actually meet an orphan?
You meet it in the consumer world first, because there it is visible. The Humane AI pin is the clearest example. It launched at 699 dollars and was rendered close to obsolete within about a year of US availability, after HP bought Humane’s operating system and patents for 116 million dollars through an asset sale. That structure can let an acquirer step away from the previous company’s support obligations.
Buyers in the first 90 days got refunds. Everyone else was left holding several hundred dollars of e-waste. Inside a business the same dynamic plays out, just without the press coverage. The orphaned system is not a gadget on your wrist, it is a quiet piece of the operation. Maybe it scores inbound enquiries, maybe it flags which invoices look risky, maybe it drafts the first version of a quote. It still produces an output every day, so it feels alive. Whether that output is still any good is a separate question that, in an orphan, nobody is asking.
When does this become an exit problem?
It becomes an exit problem the moment a buyer starts diligence, because diligence is built to find exactly this. Acquirers cross-reference what your business claims to do against what is actually documented and owned. They interview your operational staff and listen for references to automated reports or prediction tools with no paperwork behind them. A process leaning on an AI capability that has no owner is a finding.
The model card is the simple thing a buyer expects and an orphan lacks. As 2B Advice describes it, a model card records a system’s purpose, the data it was built on, its versions, its performance, and its known limits. Its absence tells a diligence team that nobody has been minding the system. The American Academy of Actuaries sets a comparable bar for a model inventory, every significant model carrying a description, a risk assessment, a validation status, and a named owner.
Diligence has also moved towards demanding evidence rather than assertions. Kiteworks describes acquirers now looking for audit trails and monitoring logs, not a policy binder that says monitoring happens. An orphan cannot produce that evidence by definition. The research is candid that the M&A community is still settling on exactly how much an orphan knocks off a valuation, and the precise percentages floating around are not well grounded. What is not in dispute is the direction. An unowned, undocumented, unmonitored system reads as risk, and risk gets priced in or argued over. Neither is where you want to be in the final weeks of a sale.
What to do about it before you launch
The preventable fix is to decide ownership up front, before the system goes live, rather than finding the gap years later. For each AI system you run, name the person accountable for it, write down what it depends on, and set a regular slot to check whether it still behaves. Charter Global makes the point that AI delegated to an operator while the board looks away is how the accountability gap forms.
This does not need a heavyweight governance function, and for an owner-managed business it should not have one. Arkeo AI’s framework points to a simple RACI map, who is responsible, who is accountable, who is consulted, who is informed, written once and kept current. A one-page list of what runs, who owns each thing, and when it was last reviewed will carry you a long way. The MIT Sloan view of technical debt applies cleanly here too, you are not aiming to eliminate every risk, you are managing it deliberately so the highest-value systems get the attention.
The mindset shift is the real point. Switching an AI system on is the start of its working life, and working things need an owner who keeps an eye on them. Get that owner named at launch and you never grow an orphan in the first place. If you are weighing how exposed your business is, or want a clear-eyed look at the systems already running without supervision, book a conversation.
