DOL Guidance on AI in Hiring, Scheduling and Wage Decisions

The Department of Labor has not written an AI regulation. What it wrote is a field assistance bulletin telling its own investigators how existing wage-and-hour doctrine applies when the timekeeping, the scheduling or the screening is automated — which is more useful, because it tells you which specific doctrine your system is about to collide with.

What the bulletin actually is

Field Assistance Bulletin No. 2024-1, issued by the Wage and Hour Division on 29 April 2024, is titled Artificial Intelligence and Automated Systems in the Workplace under the Fair Labor Standards Act and Other Federal Labor Standards. A field assistance bulletin is guidance to WHD field staff. It is not a regulation, it went through no notice-and-comment process, and it creates no obligation that does not already exist in the FLSA or in 29 C.F.R. What it does is state the Division’s view of how the existing rules apply, which is what an investigator will bring to an audit.

The bulletin’s single organising claim is worth stating plainly, because everything else follows from it: responsibility for compliance stays with the employer, and the Division expects human oversight of automated systems that affect pay. An employer cannot answer an investigator by describing the vendor.

Hours worked: the doctrine automation breaks

The FLSA requires payment for all hours worked, and “hours worked” is a body of doctrine rather than a definition. The regulations at 29 C.F.R. Part 785 set most of it out, and three provisions are the ones automated systems repeatedly run into.

• Waiting time. Sections 785.14 to 785.17 distinguish an employee “engaged to wait” — which is compensable — from one “waiting to be engaged”, which is not. An algorithmic dispatch system that requires a worker to remain available for an assignment, keep an app open, or respond within a window is generating exactly the facts that decide which side of that line the time falls on. A system that logs a worker as inactive between assignments is recording an inference, not a fact about compensability.
• Short breaks. Section 785.18 provides that rest periods of short duration, customarily twenty minutes or less, are counted as hours worked. Automated systems that clock out on inactivity will systematically remove those minutes from pay unless configured otherwise, and the bulletin flags this directly.
• Rounding. The Division’s long-standing position at § 785.48(b) is that rounding practices are acceptable only where they average out over time so employees are fully compensated. Rounding implemented as a one-directional rule, or produced as a side effect of how a model buckets activity, does not satisfy that.

There is a subtler failure that follows from what these systems measure. Productivity monitoring infers work from keystrokes, location pings, or camera presence. Compensability does not depend on productivity. An employee reading a printed manual at their desk is working and is producing no telemetry at all, and a system that pays on telemetry will underpay them. This is a design problem before it is a legal one: the measurable proxy and the legal predicate are different quantities.

Who is responsible for the records

Section 11(c) of the FLSA and the regulations at 29 C.F.R. Part 516 require the employer to make, keep and preserve records of wages, hours and conditions of employment. Section 516.2 lists the specific items for non-exempt employees, including hours worked each workday and total hours each workweek, and § 516.5 sets a three-year retention period for payroll records with two years for the supplementary records from which they are computed — time cards, work schedules, records of additions to or deductions from wages.

Nothing in Part 516 contemplates the records living only inside a vendor’s system, and nothing in it excuses an employer whose vendor purges data on a ninety-day cycle. If a scheduling or timekeeping product is the system of record, the retention configuration is a compliance control, and the contract needs a termination-assistance clause that gets the data back in a readable form. The same reasoning appears in the general treatment of AI contract clauses.

Where a model, rather than a clock, produced the number, the supplementary record is not just the output. To reconstruct a disputed workweek you need the inputs and the version of the logic that ran. Systems that overwrite a schedule in place, or that silently update a model, destroy the ability to answer “why was this employee credited with 34.2 hours” two years later.

Break, leave and nursing-employee provisions

The bulletin extends past the FLSA’s pay provisions into statutes the Division also enforces, and each one has an automation failure mode.

• The PUMP Act amended the FLSA to require reasonable break time and a private, non-bathroom space for nursing employees for one year after a child’s birth. Break time here is not fixed in length or frequency — it is whatever is reasonable for that employee. A scheduling system that enforces uniform break windows, or a monitoring system that scores those breaks as idle time, produces a violation without anyone deciding to.
• The FMLA entitles eligible employees to job- protected leave. An automated attendance-scoring system that counts FMLA-protected absences toward a disciplinary threshold is the classic no-fault-attendance-policy problem in software form. The leave has to be excluded from the count, which means the system has to know which absences are protected.
• Retaliation. Section 15(a)(3) of the FLSA prohibits discharging or discriminating against an employee for filing a complaint. Where an automated system reduces hours or assignments for a worker who has complained, the sequence is what a plaintiff will point at, and “the algorithm did it” is not one of the statutory defences.

Note also what the FLSA does not do. It sets no maximum on scheduling volatility, requires no advance notice of a shift, and does not regulate on-call scheduling as such. Predictive-scheduling duties — advance-notice windows, predictability pay for late changes — are creatures of state and municipal law, and they are the rules an algorithmic scheduler is most likely to breach. Federal wage law is the floor, not the ceiling.

Federal contractors and the OFCCP layer

Employers holding covered federal contracts have historically carried a second set of duties administered by the Office of Federal Contract Compliance Programs, which in 2024 published promising-practices guidance on AI and equal employment opportunity for contractors. That layer is unusually unstable: Executive Order 11246, the source of much of OFCCP’s authority, was revoked by executive order in January 2025, while the obligations sourced to Section 503 of the Rehabilitation Act and to the Vietnam Era Veterans’ Readjustment Assistance Act have separate statutory grounding.

The practical reading is the same as on the EEOC page: distinguish what a statute requires from what an agency is currently choosing to enforce, and do not let the second be reported to you as the first. Section 503 requires reasonable accommodation, and an assessment tool that cannot be completed by a candidate using a screen reader is an accommodation problem independent of any AI-specific instrument. That is the kind of exposure that survives every change of administration, because it is written in a statute rather than in a bulletin.