AI, explained
What is model drift?
Model drift is the reason an AI system that passed every test in March quietly starts producing worse answers in July, with no deployment in between.
The four causes, and they are not the same problem
- Data drift. The inputs change. A supplier redesigns their invoice, a new product line appears, a form gains a field. Extraction accuracy falls on the new shape while the old cases still pass.
- Provider model updates. The endpoint stays the same, the model behind it does not. Behaviour shifts, often subtly, and prompts that were tuned to the old version become slightly wrong.
- Concept drift. The right answer itself changes. A policy is amended, a rate changes, a regulation comes into force. The system is still accurate against the corpus and wrong in the world.
- Usage drift. Users find the system useful and start asking harder questions than the ones it was built for. Nothing broke. The job got bigger.
Detecting it before a customer does
Three instruments, all cheap. A scheduled evaluation run against a labelled set, executed on a fixed cadence rather than only at release, so a provider side change shows up as a number. Distribution monitoring on the inputs, which catches a new document format faster than any human review. And tracking on refusal rate, low confidence rate and human override rate, since a rising override rate is drift being absorbed by staff before anyone reports it.
What to do about each
Data drift is fixed by adding the new shapes to the evaluation set and adjusting extraction, not by raising the temperature and hoping. Provider updates are handled by pinning model versions where the platform allows it and running the regression suite before adopting a new one. Concept drift is a content problem: reindex the corpus and remove superseded documents, because retrieval will keep finding the old policy for as long as it exists in the index. Usage drift is a scope decision, and often good news.
The organisational part
Drift is why an AI system is a running service rather than a delivery. Someone has to own the evaluation cadence, read the results, and have authority to pause the automation. Systems that are built and handed over without that role degrade until people stop trusting them, which is a slower and more expensive failure than an outage.
Frequently asked questions
What is model drift?
Degradation in an AI system quality over time with no code change. Four causes: the input data changes shape, the provider updates the underlying model, the correct answer itself changes because a policy or rate changed, or users begin asking harder questions than the system was designed to handle.
How do you detect model drift?
Run a labelled evaluation set on a schedule rather than only at release, monitor the distribution of inputs so new document formats surface quickly, and track refusal, low confidence and human override rates. A rising override rate usually means staff are quietly absorbing drift before anybody files a report.
How often should you re-evaluate an AI system?
Weekly for anything customer facing or high volume, and always before adopting a new model version. The run should be automated so it happens whether or not someone remembers. Pin model versions where your platform allows it, so provider updates become a decision rather than a surprise.
Related
Ready to put AI to work?
Book a discovery audit and we will map the highest-ROI AI agents and automations for your business.
Book a discovery audit →