Fernside Studio
Background
Archive
Journal Entry

Model Drift: Why AI Gets Worse Over Time

Documented
Capacity
6 MIN READ
Domain
Fernside CMS & Support

Your AI extraction tool was 95% accurate when you deployed it. Three months later, it’s dropping invoices and misclassifying emails. Nobody touched the config. Welcome to model drift: it happens to every AI system eventually, and most teams don’t notice until a customer complains.

What Model Drift Actually Is

Model drift is the slow decline in an AI system’s accuracy after deployment, caused by a mismatch between the world it was trained on and the world it’s now operating in. Nothing broke. The inputs changed, or the model behind your tool changed, and your carefully tuned setup didn’t move with it.

Think of it like a member of staff trained on last year’s price list and product range. They’re not incompetent, they’re working from outdated assumptions. Left unchecked, that gap widens every week.

The Three Types of Drift You’ll Actually Encounter

Most explanations of this get buried in ML research language. Here’s the plain-English version, because ML research on drift rarely translates into what a CTO or ops manager needs to do about it.

Data drift: the inputs your AI sees have changed shape. If you built an email triage tool around last year’s customer queries and your product line has shifted, the model is now classifying a different world than the one it learned. IBM’s overview of model drift describes this as a mismatch between training and live-environment data distributions.

Concept drift: the relationship between inputs and correct outputs has changed, even if the inputs look similar. A “high-priority” support ticket six months ago might mean something different today because your SLAs, pricing, or product changed. The pattern the model learned no longer maps to the right answer.

Model degradation (provider-caused drift): this one is uniquely modern. If you’re calling OpenAI, Anthropic, or another provider’s API, they can silently update the underlying model. Your prompts, which were carefully tuned against one version, now hit a different one. Researchers at Stanford and UC Berkeley documented this directly: comparing GPT-4’s behaviour in March 2023 versus June 2023, they found its accuracy on identifying prime numbers dropped from 84% to 51%, and its ability to follow formatting instructions fell sharply, with no warning to developers building on top of it.

How to Detect Drift Before Your Customers Do

Drift is quiet by design. Outputs still look plausible, confidently worded, correctly formatted, just wrong more often. Detection has to be deliberate.

  1. Set a baseline. Before anything goes live, measure accuracy against a fixed test set. Without this number, “it feels less accurate” is the only signal you’ll have.
  2. Re-run evaluations on a schedule. Monthly or quarterly, run the same test set through the live system and compare against baseline. A 5-point drop is worth investigating; a 15-point drop is an incident.
  3. Monitor input distributions. Track whether the shape of what’s coming in, ticket categories, invoice formats, email topics, is shifting from what the system was built for. This is where structured monitoring for AI systems earns its place, catching the trend before a customer does.
  4. Track user feedback and overrides. A rising rate of manual corrections or “this is wrong” flags is often the earliest real-world signal, well before a formal evaluation catches it.

Mitigation: What to Actually Do About It

Detection tells you drift is happening. These are the moves that keep it from becoming a customer-facing problem:

  • Version your prompts. Treat prompt changes like code changes: logged, dated, and reversible, so you can tell whether a dip in accuracy tracks a provider update or your own change.
  • Keep a fallback model or ruleset. If your primary model’s behaviour shifts, having a secondary path (a different model, or a simpler rules-based check) prevents a single upstream change from taking down your whole process.
  • Build in human-in-the-loop for edge cases. Route low-confidence outputs to a person rather than letting the system guess. This is cheaper than it sounds and catches the cases where drift does the most damage.
  • Schedule retraining or re-tuning. If you’re running your own model or a fine-tune, put a retraining cadence on the calendar rather than waiting for a visible failure.
  • Treat evaluation as ongoing work, not a launch task. This is the mindset shift most teams miss, accuracy checks aren’t a one-off QA step, they’re a recurring operational habit.

Why This Catches Non-ML Teams Off Guard

Most guidance on drift is written by and for machine learning engineers, full of statistical distribution tests and academic notation. If you’re a founder or ops manager running an AI tool bolted onto your invoicing, CRM, or support inbox, you don’t need a PhD-level explanation, you need a monthly check and a plan for when it fails. That gap between academic drift literature and practical business monitoring is exactly where most SMB AI tools quietly degrade.

This is also why we build AI systems with managed hosting and ticket-based support rather than a “set it and forget it” handoff. A tool that was accurate at launch needs the same ongoing attention as your website’s page speed or your conversion rate, checked, not assumed.

FAQ

How often should I check my AI tool for drift? Monthly for high-stakes processes like invoice extraction or financial data; quarterly at minimum for lower-risk tools like content drafting.

Can drift happen even if I never touch my prompts or code? Yes. Provider-caused drift means the model behind your API calls can change without any action from you.

Is model drift the same as a bug? No. A bug is a fixed, reproducible error. Drift is a gradual decline in accuracy with no code change, it requires monitoring to catch, not just testing.

Next Steps

If your AI tool’s accuracy feels like it’s slipping, don’t wait for a customer to catch the next mistake. We help teams set up baseline evaluations, monitoring, and fallback paths so drift gets caught in a monthly check rather than a support ticket. Talk to us about diagnosing your AI performance, or read more about how we approach managed AI systems for growing teams.

Sources