In 2024, Air Canada lost a court case because their chatbot gave a customer incorrect information about bereavement travel discounts. The chatbot was confidently wrong and still running, still responding, still appearing to work. Nothing in their monitoring stack flagged the problem. Air Canada was held legally liable for what their AI said.
This is AI drift in its most expensive form.
Unlike traditional software, which fails loudly with error messages, system crashes, alerts – AI models degrade quietly. A model performing well in January can be generating hallucinations or biased outputs by June, and nothing will tell you. No alarms fire. No tickets are raised. Your system just gradually becomes less reliable, and you find out when a customer complains or a court rules against you.
This is the central challenge of AI in production: not building it but keeping it trustworthy over time.
What AI Drift Actually Is
Drift is the umbrella term for any change that causes a model’s performance to degrade after deployment. It takes three main forms.
Data drift occurs when the inputs your model receives in production no longer resemble its training data. Customer demographics shift. Upstream data pipelines change. New product categories appear. The model was built for a snapshot of your world, not for the world as it is now.
Concept drift is more insidious. Here, the relationship between inputs and outputs changes even when the data itself looks similar. A spam filter trained on 2022 phishing patterns will increasingly miss 2026 phishing tactics. The inputs still look like emails. The model’s understanding of what “spam” means is just no longer accurate.
Feature drift involves structural changes in your data pipelines: renamed columns, altered categorical encodings, new data sources. This one is often detectable through validation but catastrophic when missed.
For large language models, drift manifests differently again. Decreased response relevance, inconsistent tone, incomplete task execution. Critically, the non-deterministic nature of LLM outputs means a successful API call can still deliver hallucinated or harmful content. The model “works”, it just doesn’t work correctly.
Why Silent Failure Is the Real Risk
Most AI monitoring in organisations focuses on infrastructure: uptime, latency, error rates. These metrics tell you whether your model is responding. They don’t tell you whether it’s right.
This gap creates the conditions for silent failure. Performance erodes gradually. Users notice before engineers do. Users stop trusting the system, work around it, or act on incorrect outputs. By the time the problem surfaces, it may have been compounding for weeks.
The consequences follow a predictable sequence: eroded customer trust, rising operational costs as human intervention fills the gap, and legal exposure of the kind Air Canada experienced. The harder problem is that hallucination risk compounds with drift. As a model’s semantic consistency weakens, it becomes more likely to produce plausible-sounding outputs that are factually wrong, and harder to distinguish from correct ones.
You don’t need better incident response. You need earlier detection.
The AI Vital Signs Dashboard
The solution is to track AI vital signs; measurable indicators that reveal what’s happening beneath the surface before it becomes a crisis. CTOs managing AI in production should track a number of metrics weekly, each functioning as a vital sign for model health. Some examples are below, and can vary depending on what systems you’re monitoring.
1. Hallucination Rate
A hallucination is when your model generates a confident, plausible-sounding response that is factually incorrect or unsupported by any source it has access to. For LLMs, this is the most visible form of drift, and the most legally exposed.
Monitoring hallucination rate requires more than infrastructure tooling. Effective approaches combine automated scoring systems, external knowledge-base verification, and human review of sampled outputs. LLM-as-judge frameworks where one model evaluates another’s outputs against a ground truth, are increasingly used in production environments. What matters is that you have a weekly number, established trends, and a threshold that triggers investigation.
2. Latent Drift in Conversations
In conversational AI, drift can occur within individual interactions as well as across them. As a context window fills, or as a conversation extends across multiple turns, the model’s internal reasoning can subtly shift resulting in losing track of earlier context, becoming less coherent, or drifting from the intended purpose of the interaction.
The signal for this is often indirect: rising conversation abandonment rates, increased follow-up questions, users repeating themselves. More precisely, tracking how the semantic distribution of model outputs changes over time by using a similarity measure to compare outputs from one period to the next can surface drift before it becomes visible in user behaviour.
3. Drift Across Conversations
Where latent drift measures what happens within a conversation, this metric tracks change across your entire user base over time. Are response lengths shifting? Is sentiment changing? Are the same prompts producing meaningfully different outputs week to week?
This is the macro view of model health. Monitoring output variance such as response length distributions, topic patterns across outputs, how similar prompts perform over time, reveals whether your model’s core behaviour is stable or drifting. Version-controlling your prompts and system instructions is a prerequisite for this analysis. Without it, you cannot isolate model changes from prompt changes.
4. Token Cost-to-Value
This is the economic vital sign that can be used to capture many other unmonitored aspects. LLMs charge by token, and costs can escalate quickly and quietly. But cost is relative to the value the model generates.
Track token usage per request and per feature. Then correlate those costs against business outcomes: conversion rates, support ticket resolution times, task completion rates, revenue attributed to AI-powered features. When cost-to-value deteriorates, shown through rising costs while outcomes remain flat or fall, this is often an early signal that model performance is degrading in ways other metrics have not yet caught.
What to Do with an AI Vitals Dashboard
The dashboard described here is not a one-time audit. It is an ongoing operating rhythm.
Set thresholds for each metric. Define what acceptable looks like for hallucination rate, conversation drift signals, output variance, and cost-to-value. When a metric breaches that threshold, that is your trigger for systematic diagnosis through structured investigation rather than panic. Is the model drifting, or has something changed in the data pipeline? Are prompts being updated without version control? Has an upstream data source changed?
The organisations that manage AI reliably do not necessarily have better models. They have better systems around their models. Monitoring is what separates AI that erodes quietly from AI that stays accountable.
Your models will drift. The question is whether you find out first.