A company deployed AI for hiring in January 2025.
Before launch, they tested everything. Bias testing showed a neutral demographic impact. Performance metrics were excellent, and the legal department approved it.
They went live.
It’s now February 2026. The AI has been running for over a year.
Last week, a rejected candidate filed an EEOC complaint alleging that the AI discriminates based on age.
During the investigation, they discovered the AI is now rejecting older candidates at significantly higher rates than it did at launch.
What changed? Nobody touched the AI. The model is the same.
But the AI’s decisions are completely different.
What Model Drift Actually Is
AI models change over time even when nobody modifies them.
Data drift: The data the AI receives changes. New types of resumes. Different job markets. Changed application patterns. The AI trained on 2024 data is now processing 2026 data that looks different.
Concept drift: The relationship between inputs and outputs changes. What made someone a “good candidate” in 2024 isn’t the same in 2026. The AI’s learned patterns are outdated.
Label drift: The ground truth changes. Job requirements evolve. Success criteria shift. The AI is optimizing for old definitions of success.
The AI keeps adapting to new data while nobody monitors whether the adaptations are appropriate or legal.
Why This Is a Legal Problem
Companies think testing AI before deployment protects them.
It doesn’t.
Testing proves the AI worked correctly at one point in time. It doesn’t prove the AI still works correctly months or years later.
When AI drifts and produces biased outcomes, “we tested it before launch” isn’t a defense.
Courts ask: “When did you last test for bias?”
If the answer is “at deployment 18 months ago,” that’s not adequate monitoring.
Regulators ask: “How do you know the AI isn’t discriminating now?”
If the answer is “we assume it’s fine because it was fine at launch,” that’s not governance.
EEOC asks: “Show us evidence the AI currently produces fair outcomes.”
If companies can’t produce recent testing results, they can’t defend the AI.
Why Companies Don’t Monitor for Drift
They assume AI is static: Companies think that once AI is deployed and working, it stays that way. They don’t understand that AI continues adapting to new data.
Monitoring is expensive: Regular bias testing costs money. Testing every few months feels wasteful when “the AI is working fine.”
Nobody’s responsible: IT deployed it. Legal approved it. HR uses it. Nobody owns ongoing monitoring.
Vendors don’t tell them: AI vendors rarely warn customers about drift or provide monitoring tools. They deliver the AI and move on.
There’s no requirement: Unlike financial systems with SOX controls or medical devices with post-market surveillance, AI has no mandatory ongoing monitoring in most jurisdictions yet.
So companies deploy AI, test it once, and hope it keeps working correctly forever.
What Monitoring Should Look Like
If companies want AI that stays compliant over time, they need continuous monitoring.
Monthly bias testing: Run demographic impact analysis monthly. Compare current outcomes to the deployment baseline. Flag significant changes.
Performance tracking: Monitor whether AI accuracy and reliability remain consistent. Declining performance often precedes bias issues.
Data distribution monitoring: Track whether input data is changing. If the AI receives data that looks different from training data, outcomes will drift.
Decision pattern analysis: Look for changes in decision patterns. If the AI suddenly rejects more of a certain category, investigate why.
Complaint tracking: Monitor discrimination complaints, customer complaints, and employee concerns about AI decisions. Patterns indicate problems.
Quarterly re-validation: Every quarter, formally re-test the AI against the same standards used at deployment. Document results.
Most companies do none of this until regulators or lawsuits force it.
When Drift Triggers Re-Deployment
Some drift is minor and acceptable. Some drift requires immediate action.
Immediate shutdown triggers: Bias testing shows a significant disparate impact that didn’t exist at deployment. Performance degradation affects decision accuracy. Data analysis shows the AI is processing fundamentally different inputs from the training data. Complaints indicate systematic problems with AI decisions.
Re-training triggers: Drift is significant but not severe enough for shutdown. The AI needs updated training data reflecting current conditions. Re-train the model, re-test for bias, and re-deploy with proper validation.
Enhanced monitoring triggers: Drift is minimal but trending in the wrong direction. Increase monitoring frequency and prepare for potential re-training.
Most companies have no defined thresholds for when drift requires action, so they ignore it until a crisis forces a response.
The Documentation Problem
When a discrimination lawsuit or regulatory investigation happens, companies need to prove they monitored AI over time.
Documentation should show when bias testing was conducted, what the results were, what changes in AI performance were detected, how drift was investigated, what corrective actions were taken, and when the AI was re-trained or re-validated.
Most companies have zero documentation of ongoing monitoring because they didn’t do any ongoing monitoring.
They can show they tested the AI at deployment. They can’t show what happened in the year since then.
That documentation gap becomes evidence of negligence.
What Companies Should Be Doing Now
Implement monitoring systems: Set up automated monitoring for AI performance, bias metrics, and data distribution changes.
Assign ownership: Make someone responsible for reviewing, monitoring results, and escalating issues.
Define action thresholds: Document when drift requires investigation, re-training, or shutdown. Don’t wait for a crisis to decide.
Schedule regular testing: Put recurring bias testing on the calendar. Treat it like financial audits, not optional maintenance.
Document everything: Keep records of all monitoring, testing, and corrective actions. This documentation protects the company in litigation and investigations.
Budget for it: Ongoing AI monitoring costs money. Include it in the AI operating budget from the start.
Companies that skip these steps are operating AI that’s drifting away from the tested baseline with no visibility into what it’s actually doing.
Ask Yourself This
For any AI system making decisions about people, when was the last time bias testing was conducted?
If the answer is “at deployment” and that was more than a few months ago, the company has no idea whether the AI is still compliant.
If the answer is “we’ve never tested it,” the company is operating AI blindly.
Either way, there’s a monitoring gap that creates massive liability.
Schedule a 30-minute AI Drift Assessment where we’ll:
- Review what AI systems need ongoing monitoring
- Show what drift monitoring actually requires
- Identify current monitoring gaps
- Help implement monitoring systems and testing schedules
- Create documentation practices that protect against liability
Because testing AI once at deployment proves it worked on day one, while ongoing monitoring proves it’s still working correctly today.
Brian Alexander
Founding Partner, VisionAI+ Consulting Group
Former General Counsel, AI Engineering Firm
Certified ISO 42001 Lead Auditor