Employee Monitoring and AI Surveillance

Start with purpose, not capability

The vendor demo shows what the tool can capture — keystrokes, screenshots, message sentiment, faces, voice, location, an attrition score for every employee. The legal question is what is actually needed, and why. Most of the exposure on this page is bought when a company collects broadly because it can, then struggles to explain the purpose later. AI changes the workflow here; it does not change the duties. Notice, non-discrimination, and the limits of state privacy and labor law apply to an algorithm exactly as they apply to a human watching over a shoulder, and “the system flagged it” is not a defense to any of them.

• Write down the legitimate business purpose — before procurement. Security, safety, timekeeping accuracy, quality control. State it specifically. A purpose that can be articulated in one sentence is one that can be defended; “general productivity insight” is not.
• Collect the minimum that serves it. Map each data stream the tool offers to the stated purpose and switch off the rest. Continuous screen capture, ambient audio, and message-content analysis carry far more risk than aggregate activity metrics — turn them on only if the purpose genuinely requires them.
• Separate monitoring from scoring. Capturing data is one decision; generating a productivity, sentiment, or “flight-risk” score and acting on it is a second, riskier one. Treat them as distinct and govern the second more tightly.

Notice and consent — the failure that does the most damage

The single most common and most avoidable exposure is undisclosed monitoring. Many state laws restrict the interception or recording of communications, and several require the consent of all parties rather than just one; thresholds, definitions, and which channels count vary sharply by state and are worth confirming for every jurisdiction where employees sit. Separately, a handful of states impose their own affirmative written-notice requirements for electronic monitoring — obligations that apply on their own terms, independent of any recording or consent question, with the form and timing varying by state; confirm whether any apply where employees sit. Covert monitoring also reads badly to regulators, juries, and the workforce itself. Disclosure is cheap. Litigating its absence is not.

• Disclose what is collected, and how it is used. A clear written notice — what data, from which systems, for what purpose, and whether it factors into evaluation or discipline. Vague boilerplate about “company systems” is not notice of keystroke logging or sentiment analysis.
• Map the wiretap and eavesdropping rules state by state. Recording or analyzing calls, video with audio, or message content can trigger state interception and all-party-consent statutes. Where a workforce spans states, the strictest applicable rule tends to set the floor. Confirm current scope and consent requirements with counsel for each location.
• Get consent where consent is the mechanism that makes it lawful. Where a statute turns on consent, build a documented consent step rather than assuming an acceptable-use policy covers it. Confirm what form of consent the relevant law requires.
• Watch the personal-device and off-duty edges. BYOD, personal accounts, and conduct outside work hours raise distinct privacy issues, and several states protect lawful off-duty activity. Monitoring that reaches into an employee’s personal sphere is where notice and scope matter most.

Biometric capture is its own regime — clear it first

The moment a tool processes a faceprint, voiceprint, or similar identifier — facial-recognition video analytics, voice authentication, some “engagement” sensing — it may fall within a biometric privacy law. Illinois’s BIPA is the most developed example, and several other states have their own regimes with differing definitions and requirements; statutory damages and private rights of action under some of these laws can be significant, and the contours are actively litigated. Treat biometric features as off by default until compliance is confirmed.

• Identify whether the tool captures biometrics at all. Read the data sheet, not the marketing. “Facial analysis,” “voice verification,” and “identity confirmation” features often process exactly the identifiers these statutes cover.
• Do not collect biometrics without confirming compliance. Several biometric laws require specific notice, written consent or release, and a published retention-and-destruction policy before collection. Confirm which apply to your locations and what each currently requires before the feature goes live.
• Push the obligations into the vendor contract. Where a vendor captures or stores biometric data on the company’s behalf, allocate compliance, retention limits, deletion, and indemnity in writing. See AI Vendor Contracts.

Labor law: monitoring can chill protected activity

Surveillance of the workforce intersects with the National Labor Relations Act, which protects most employees — union and non-union alike — in their concerted activity over working conditions. New, expanded, or covert monitoring that could reasonably tend to discourage employees from discussing pay, organizing, or raising workplace concerns has drawn scrutiny, and the Board’s posture on AI-driven surveillance has been an area of evolving and shifting attention. Targeting monitoring at union or organizing activity is high-risk. Confirm the current state of the law before rolling out new surveillance.

• Never aim monitoring at protected activity. Tooling pointed at organizing conversations, concerted complaints, or who talks to whom about working conditions is the clearest danger zone. Do not configure it that way, and do not let a vendor’s “collaboration analytics” drift into it.
• Be deliberate about new or covert monitoring. Introducing surveillance — especially without notice, and especially during organizing — carries added risk. Disclosure and a documented business purpose reduce it.
• Mind bargaining obligations in a represented workforce. Introducing or changing electronic monitoring may be a subject for bargaining. Route it through labor counsel before deployment, not after.

Health inferences, disability, and uneven analytics

Tools that infer stress, fatigue, mood, attention, or wellness are inferring something close to health. Acting on those inferences can implicate the ADA and related state law, and screening that was not asked for can surface disability information that then has to be handled carefully. Separately, analytics applied unevenly — harsher scrutiny of one group, scoring models that perform worse for some populations — can produce discrimination exposure even where no one intended it. The relevant law here is the familiar law; the tool just creates new ways to run afoul of it.

• Treat health and disability inferences as hazardous. Stress, wellness, and “burnout-risk” scoring can generate disability-related information and ADA questions. Approach with counsel; do not let a productivity dashboard quietly become a medical-inference engine.
• Test for disparate impact, and apply monitoring evenly. Check whether scores or flags fall unevenly across protected groups, and ensure the tool is not deployed more aggressively against some teams than others. Document the review.
• Account for accommodation. An employee with a disability may legitimately work in ways a productivity model reads as anomalous. A score that ignores an approved accommodation is a discrimination problem waiting to surface.

The hardest decision: using monitoring data on people

This is where a defensible monitoring program becomes an indefensible termination. AI scores are probabilistic, not factual. They generate false positives, they reflect their training data, and they are frequently wrong about any single person. A score is a prompt to look, never the finding itself. The duty to make a fair, evidence-based employment decision stays with the company — a human, not a dashboard, has to own the call.

• Never treat a score as proof. A flight-risk number, a low productivity score, or a sentiment flag is a signal to investigate — not a basis to discipline. The evidence supporting an adverse action has to stand on its own without the score.
• Verify before you act. Confirm the underlying facts independently. AI tools hallucinate, miscount, and misattribute; the documented basis for a decision should be what a human verified, not what the model asserted. The same verification discipline applies in litigation work product.
• Keep a human owning every disciplinary decision. A named decision-maker reviews the actual evidence and makes the call. “The algorithm flagged it” in a deposition is the answer that loses the case.
• Train managers on the limits. The most expensive failures come from a supervisor who trusts the dashboard. Train them that scores are inputs, that false positives are routine, and that the obligation to verify is theirs.
• Remember the data is discoverable. Everything the tool collects — scores, flags, the criteria behind them — can surface in litigation. Inconsistent enforcement and unexplained flags read very differently a year later as exhibits.
• Put it in the inventory and govern it. Monitoring and scoring tools belong in the AI inventory with an owner, a documented purpose, and a review cadence. See AI Governance and Inventory.