AI-Driven Termination and Discipline: The Employment Decision Nobody Audits
Bias-testing budgets went almost entirely to the front door. Resume screeners get audited, interview scorers get audited, candidate rankers get audited. Meanwhile the systems that decide who gets written up, put on a plan, and let go run unexamined — with the same disparate-impact exposure and considerably worse documentation.
"We Don't Use AI to Fire People"
Almost every employer says this, and almost every employer is describing the wrong thing. The question is not whether a model signs the termination letter. It is whether an automated output materially drove the decision. If a manager opened a dashboard, saw an employee in the bottom decile of an algorithmic productivity score, and started a performance plan on that basis, the algorithm drove the decision even though a human executed it.
This is the same "substantially replace" analysis regulators apply elsewhere. A human who reviews the score, investigates independently, and can and does reach different conclusions is meaningful review. A human who acts on the flag because acting on flags is what the workflow expects is a rubber stamp with a signature block.
The Systems That Count
Productivity and Performance Scoring
ScoringComposite scores from ticket throughput, handle time, code volume, sales activity, or call quality. These become the evidentiary backbone of a performance-based termination, and they are rarely validated as job-related in the way a selection procedure would have to be.
Monitoring and Activity Analytics
MonitoringKeystroke cadence, idle detection, application usage, camera-based attention scoring. Deviation from a norm gets read as disengagement, when it frequently reflects disability, caregiving, medication schedules, or a different but equally effective working style.
Automated Scheduling and Attendance
AttendanceSystems that assign shifts and score adherence. Points-based attendance engines are a well-worn source of disability and religious-accommodation claims, and automation makes them apply the rule with perfect, undiscriminating consistency — which is exactly the problem.
Sentiment and Communication Analysis
SentimentTools scoring tone in support tickets, sales calls, or internal chat. Sentiment models carry documented performance gaps across dialect, accent, and non-native speech patterns — a direct line to national-origin exposure when the output feeds a coaching or discipline record.
Flight-Risk and Attrition Prediction
PredictionModels predicting who is likely to leave. Acting on a prediction — withholding a project, an assignment, or a raise from someone flagged as a flight risk — creates an adverse action based on a forecast, often built on tenure and demographic-correlated features.
The Accommodation Trap
This is the exposure most likely to surprise an employer, because it turns a good act into a liability. A company grants a reasonable accommodation — reduced hours, additional breaks, a modified workflow, leave for treatment — and then leaves the productivity system measuring the employee against the unmodified baseline. The score drops. The dashboard flags it. A manager who never saw the accommodation file starts a performance plan.
The employer granted the accommodation and then penalized the employee for using it, which is close to a textbook failure-to-accommodate and retaliation fact pattern. The same shape recurs with pregnancy accommodations, FMLA-protected leave, military leave, and religious scheduling. The fix is structural, not procedural: accommodations must be encoded into the measurement system itself, so the baseline adjusts rather than relying on a manager to remember to discount the number.
Exposure Triage
Does an algorithmic score appear in your termination documentation?
CriticalIf a score, ranking, or system-generated flag is cited in the PIP or termination memo, you have created the plaintiff's exhibit. Expect discovery into how the score is computed, whether it was validated as job-related, and how it distributes across protected groups. Be able to answer all three before you write the memo.
Have you compared discipline rates across protected groups?
CriticalMost employers run adverse-impact analysis on hiring and never on discipline. Run flag rates, PIP rates, and termination rates by group with the four-fifths rule as the initial screen. A system disciplining one group at well below four-fifths the rate of the highest group is a finding you want to make internally, under privilege, before someone else makes it.
Do accommodations flow into the measurement system?
CriticalCheck whether an approved accommodation actually adjusts the employee's baseline, targets, and alert thresholds — or whether it lives only in an HR file the dashboard never reads. This is a systems-integration question with direct ADA consequences.
Are employees told which automated systems assess them?
HighNotice obligations now attach in California, Colorado, and a growing list of jurisdictions, and CCPA's employee exemption has expired. Silent monitoring that drives adverse decisions is both a notice violation and, in litigation, evidence of concealment.
Can a manager actually override the system?
HighNot in theory — in the data. Look at how often flags are dismissed and how often overrides occur. If the override rate is near zero, you have a rubber stamp, and the 'human in the loop' defense will not survive the first deposition.
Did your vendor supply validation and impact data?
MediumAsk for bias-testing results, validation evidence, and the feature list. Many vendors will not provide them. That refusal is itself a finding: you are deploying an unexamined system into decisions with statutory exposure, and the liability sits with the employer, not the vendor.
The Proxy Problem
Removing protected characteristics from the feature set does not remove them from the model. Work location correlates with race in most metropolitan labor markets. Shift assignment correlates with caregiving status and therefore sex. Tenure correlates with age. Device type and connection quality correlate with income. Break patterns correlate with disability and pregnancy. A discipline model built on "neutral" operational features can reproduce protected distinctions faithfully, and disparate-impact doctrine does not require anyone to have intended it.
A Practical Audit Program
- Inventory every automated system whose output reaches a manager making discipline, promotion, scheduling, or termination decisions.
- Classify each as decision-driving or informational, and be honest — if it appears in documentation, it is decision-driving.
- Run adverse-impact analysis on outcomes: flag rates, discipline rates, PIP rates, termination rates by protected group.
- Review the feature list for proxies and test whether removing suspected proxies changes the distribution.
- Wire accommodations into the metrics so approved adjustments change the baseline automatically.
- Measure override rates to verify human review is real rather than nominal.
- Give employees notice of the systems assessing them and a genuine appeal path with a human decision-maker.
- Retain decision logs — inputs, outputs, and the human rationale — so you can reconstruct why a decision was made.
- Re-audit on a schedule. Workforce composition and model behavior both drift; a clean result in Q1 is not a defense in Q4.
The strategic point is simple. Regulators and plaintiffs' firms spent the last few years learning how to examine AI hiring tools, and that expertise transfers directly to the exit side of the employment relationship — where the decisions are more consequential to the individual, the documentation is thinner, and almost nobody has run the numbers. Run them first.
Check how your product presents to users and regulators
RatedWithAI scans your web properties for compliance and trust gaps — including the accessibility and disclosure surfaces that regulators look at first. Start with a free scan.
Scan Your Site for Free →Frequently Asked Questions
We only use AI to flag issues — a human always decides. Are we safe?
Only if the human decision is genuine and you can evidence it. Regulators and courts look at override rates, the time between flag and action, and whether the reviewer had independent information. A workflow where flags convert to discipline at near-100% is functionally automated decision-making with a human signature attached.
Does NYC Local Law 144 cover firing?
No — it is aimed at automated employment decision tools used in hiring and promotion screening, with a bias audit and candidate notice. Termination is outside its scope. But broader frameworks such as Colorado's AI Act and California's ADMT rules reach consequential and significant employment decisions more generally, so a hiring-only compliance posture leaves the exit side uncovered.
Can employees demand an explanation for an AI-influenced termination?
Increasingly yes, depending on jurisdiction. California's ADMT access right and Colorado's notice-and-explanation duties both contemplate telling an affected person that an automated system was used and giving a meaningful account of how. The practical constraint is that you cannot explain a decision you did not log — build the logging before someone asks.
Should we stop using productivity analytics entirely?
No, and that is rarely the right answer. Measurement is legitimate and often improves fairness relative to unstructured managerial judgment. The exposure comes from unvalidated metrics driving consequential decisions without impact analysis, accommodation integration, notice, or an appeal path. Fix those four and the tooling becomes an asset in litigation rather than the exhibit against you.
Who is liable — us or the vendor that built the system?
The employer, in the first instance. Anti-discrimination statutes attach to the employment decision, and 'the vendor's model did it' is not a defense. Vendor contracts should carry bias-testing commitments, documentation access, and indemnity, but those allocate cost between you and the vendor — they do not move the legal exposure away from the employer.