Eighteen months after substantial completion on a $310M IIJA-funded transit project, a Department of Labor Wage and Hour investigator pulled the contractor's certified payrolls and found that a crew of pipefitters had been paid and reported under a laborer classification for nine weeks, that the fringe-benefit credit claimed on the WH-347 exceeded the actual contributions made to the plans, and that an electrician apprentice had worked an entire phase without the journeyworker on site that the registered apprenticeship ratio required. None of it was fraud. It was a payroll clerk reading a timecard, picking a classification from a dropdown, and trusting that the posted Davis-Bacon wage determination matched the work. The firm signed a consent finding, paid back wages plus liquidated damages, and spent a year on a monitoring agreement that slowed every future federal award. This lesson puts AI to work on exactly that machinery: ingesting weekly timecards, cross-checking classifications against the wage determination, generating the WH-347 batch and the statement of compliance, tracking DBE/MWBE participation against the contracting officer's goal, and assembling the audit-ready bundle. But certified payroll is a legal attestation signed under penalty, so the human verifies and signs, and the AI never does.

Why Federal Payroll Is Its Own Animal

Commercial private payroll asks one real question: did the worker get paid what the employment agreement says, on time, with the right taxes withheld. Federal-funded construction payroll under the Davis-Bacon Act and its related acts asks a harder set of questions, and the answers are submitted weekly to the contracting agency as a sworn document. The wage paid is not whatever the contractor and the worker agreed to; it is the prevailing wage for the worker's classification in the project's locality, as fixed in the posted wage determination, with two components that must both be satisfied: the basic hourly rate and the fringe-benefit amount, which the contractor either pays into bona fide benefit plans or pays out as additional cash wages. Misclassify the worker, pay below the determination rate, miscompute the fringe credit, or violate the apprenticeship ratio, and the firm has a finding waiting to surface.

The federal-funding wave of the late 2020s widened this exposure dramatically. The Infrastructure Investment and Jobs Act (IIJA), the CHIPS and Science Act, and the Inflation Reduction Act (IRA) pushed hundreds of billions of dollars into transit, broadband, semiconductor fabs, clean-energy, and grid work, and the overwhelming majority of that work carries Davis-Bacon prevailing-wage obligations. The IRA went further, conditioning the full value of major clean-energy tax credits on meeting prevailing-wage and registered-apprenticeship requirements, which means a payroll error is no longer just a back-wage liability; it can claw back a tax credit worth multiples of the labor cost. A firm that built its compliance habits on the occasional federal job now finds federal compliance is the main event, and the volume of certified payrolls, classifications, and apprenticeship-ratio checks has outrun the clerk-and-spreadsheet method that used to be good enough.

That volume is the case for AI, and it is a strong one. A 140-craftworker crew across multiple subcontractors generates thousands of timecard lines a week, each of which must be matched to a classification on the determination, priced at the correct basic rate, credited the correct fringe, and rolled into a WH-347 per contractor per week. Doing it by hand is slow and error-prone, and the errors are precisely the ones a Wage and Hour investigator is trained to find. AI can ingest the timecards, cross-check the classifications, do the rate-and-fringe math, and draft the forms at machine speed. What it cannot do is sign the statement of compliance, because that is a legal attestation, and that distinction is the spine of this lesson.

The WH-347 and the Statement of Compliance

The WH-347 is the standard certified-payroll form the Department of Labor publishes, and most federal agencies and their prime contractors require it weekly from every contractor and subcontractor who performed covered work that week. The form itself is two parts. The payroll page lists each worker by name and an identifying number, their work classification, the hours worked each day, the total hours, the rate of pay (basic plus fringe), the gross amount earned, the deductions, and the net paid. It is, in effect, the contractor swearing on a single page that every covered worker on the project was paid at or above the determination for the classification of the work they actually performed.

The second part is the statement of compliance, which turns a spreadsheet into a legal instrument. The signer certifies, under the penalties of the federal False Statements Act and related statutes, that the payroll is correct and complete, that each worker was paid not less than the applicable wage rates and fringe benefits for the classification of work performed, that no rebates or kickbacks were taken, and that the fringe-benefit representations are accurate. This is the structure the program has named repeatedly: the generative engine drafts fast, but the deliverable carries a signature that binds a real person under penalty, so the verification gate sits between the draft and the signature, not after it. A certified payroll is the federal-compliance instance of the dollars-and-contract-authority gate, and the AI lives entirely on the draft side of it.

AI can ingest the timecards, match classifications to the wage determination, and generate the WH-347 batch in minutes, but the statement of compliance is sworn under penalty by a human who has verified it, so the AI drafts and the person attests; the signature is the gate, and it is never delegable.

The Wage Determination Is the Source of Truth

Everything downstream depends on reading the wage determination correctly, so this is where the AI does its most valuable cross-checking and where the verification must be sharpest. A wage determination is the document, published per locality and per construction type (building, heavy, highway, residential), that lists every covered classification and the basic hourly rate and fringe rate that prevail for it in that area. The version that governs is the one incorporated into the contract at award, with any modifications that applied before the relevant lock-in date, and reading the wrong version or construction type is itself a classic finding. The determination is the source of truth against which every line on the WH-347 is judged.

The AI's job is to take each timecard line, read the description of the work actually performed, and propose the classification on the determination that matches it, then attach that classification's basic rate and fringe rate so the form math is correct. This is analytic work the AI does well: it holds the entire determination in context, parses hundreds of timecard descriptions, and surfaces the lines where the description does not cleanly map to any single classification. Those ambiguous lines are gold, because misclassification is the single most common and most expensive Davis-Bacon violation. A worker who spends part of a day on covered work in one classification and part in another may require split-rate reporting; a description like "helper" or "general labor" may mask work that the determination prices as a skilled trade. The AI flags these; it does not resolve them, because resolving a classification is a judgment with legal consequence.

The fringe-benefit component is the second half of the rate and the quieter trap. The determination states a fringe amount per hour, and the contractor satisfies it either by contributing to bona fide plans (health, pension, training, vacation) or by paying the unmet portion as cash. The WH-347 requires the contractor to represent how the fringe was satisfied, and a finding arises when the claimed fringe credit exceeds what was actually contributed, or when annualization of the fringe across all hours (covered and private) is computed incorrectly. AI can compute the fringe credit from the contribution data and flag where the claimed credit and the actual contributions diverge, but the representation of how fringes were satisfied is part of the sworn statement, so the contractor verifies the contribution records behind it before signing.

The Apprentice Ratio: The Quiet Killer

Apprentices are the place where a well-run payroll still generates findings, because the violation is structural rather than arithmetic. Davis-Bacon and the related acts allow an apprentice to be paid less than the full journeyworker rate for a classification only if the apprentice is enrolled in a registered apprenticeship program and only within the program's allowable apprentice-to-journeyworker ratio on the site. If the program permits one apprentice for every three journeyworkers and the contractor has two apprentices and four journeyworkers of that trade on site that day, one apprentice is out of ratio and must be paid the full journeyworker rate for the hours over the ratio. The payroll can be arithmetically perfect, every rate matched to the determination, and still be wrong because the crew composition broke the ratio.

This is the kind of cross-cutting check that defeats manual review and that AI does well as an analytic flag. The ratio depends on who was on site, in what trade, on what day, at what enrollment status, data scattered across timecards, apprenticeship registrations, and the daily crew composition. AI can join those sources, compute the ratio per trade per day, and flag every instance where the apprentice count exceeds the allowable ratio, including the partial-day cases that a clerk scanning a roster would never catch. In the applied problem below, this is where the three violations surface: not because anyone falsified anything, but because on three days across two weeks the crew composition put an apprentice out of ratio, and the AI's per-day per-trade ratio check caught what the eye would not.

The verification discipline mirrors the rest of the lesson. The AI flags the out-of-ratio instances; the compliance owner confirms the enrollment status against the registered program, confirms the journeyworker count on site that day, and decides the remedy: reclassify the over-ratio hours to the journeyworker rate and correct the affected WH-347, or document why the flag is a false positive (the journeyworker count was higher than the timecard data showed, for instance). The flag is analytic and the AI owns it; the remedy is a compliance judgment with back-wage and tax-credit consequences, and the human owns it.

DBE/MWBE Participation Is a Different Obligation

Running alongside the wage obligation, and often confused with it, is the DBE/MWBE participation requirement. Disadvantaged Business Enterprise and Minority/Women Business Enterprise goals are set by the contracting officer or the funding agency as a target percentage of contract value to be performed by certified DBE or MWBE firms, and the prime contractor has to track actual participation against that goal and report it, with good-faith-effort documentation if the goal is not met. This is a contract-performance obligation about dollars flowing to certified firms, not a wage-rate obligation about what individual workers are paid, and conflating the two is a common error that produces a memo that satisfies neither.

AI assembles the DBE/MWBE participation memo by aggregating the dollars committed and the dollars actually paid to certified firms (verified against the certification directories), computing the participation percentage against the contract value, comparing it to the contracting officer's goal, and drafting the narrative that explains the standing and, where there is a shortfall, organizes the good-faith-effort record. This is the same analytic-plus-generative pattern as the rest of the workflow: the AI does the aggregation and math and drafts the memo, and the human verifies the certification status of each firm (certifications expire and decertifications happen mid-project) and owns the representation to the agency. The participation memo is a representation about contract performance, so it sits behind the contract-authority gate just as the WH-347 does.

Responding to a DOL Wage and Hour Audit

When a DOL Wage and Hour investigation opens, the firm that survives it cleanly is the one whose records were assembled to be examined before anyone asked. An investigation typically requests the certified payrolls for a date range, the underlying timecards, the apprenticeship registrations and ratios, the fringe-benefit contribution records, and the documentation behind any classification or rate question. A firm scrambling to reconstruct this after the request arrives signals disorganization at best and creates the appearance of a problem at worst, while a firm that produces a clean, indexed, internally consistent bundle in days frames the investigation as a routine review.

This is where the AI's role shifts from drafting to assembling, a high-value role precisely because it is mechanical. The AI compiles the DOL-audit-ready records bundle: the WH-347 batch for the requested period, the source timecards tied line-by-line to each form, the wage-determination version that governed, the classification-mapping rationale for every line, the apprenticeship registrations and the per-day ratio computations, the fringe-contribution records reconciled to the claimed credits, and an index that lets an investigator trace any worker on any day back to the timecard and forward to the form. The bundle's value is that it is complete and traceable, the analytic strength of AI applied to records assembly. The compliance owner reviews the bundle for accuracy and completeness before it goes to the investigator, because the records support a position the firm is taking to a federal agency, and that position is a human's to own.

The Applied Problem: Two Weeks of Certified Payroll for 140 Craftworkers

Here is the exercise. You are the compliance owner for a general contractor on an IIJA-funded transit project with a 140-craftworker crew across your own forces and four subcontractors. You have two weeks of timecards, the posted Davis-Bacon wage determination for the locality and construction type incorporated at award, the apprenticeship registrations for the apprentices on site, and the fringe-benefit contribution records. Your deliverable is the WH-347 batch for both weeks, the three apprentice-ratio violations flagged and resolved, and the DOL-audit-ready records bundle, with your verification documented at every gate.

Run the workflow in order. First, ingest the timecards and have the AI map each line's described work to a classification on the wage determination, attaching the basic and fringe rates; review the flagged ambiguous lines yourself and resolve each classification, because that resolution is a judgment with legal consequence. Second, have the AI compute the fringe credit from the contribution records and flag any line where the claimed credit exceeds actual contributions; verify the contribution records behind the representation. Third, have the AI compute the apprentice-to-journeyworker ratio per trade per day and flag every out-of-ratio instance; here the three violations surface, on three days where the crew composition put an apprentice over the allowable ratio. Confirm the enrollment status and on-site journeyworker counts, reclassify the over-ratio hours to the journeyworker rate, and correct the affected forms.

Fourth, generate the WH-347 batch (one form per contractor per week, ten forms across the two weeks for five contractors) with the corrected rates, and draft the statement of compliance for each, which you and each subcontractor's authorized signer review and sign, because the statement is sworn under penalty and the AI never signs it. Fifth, have the AI assemble the DOL-audit-ready records bundle: the ten forms, the line-by-line timecard ties, the governing determination, the classification rationale, the apprenticeship registrations and per-day ratio computations showing the three corrections, and the fringe reconciliation, all indexed for traceability. The deliverable is the verified WH-347 batch, the three flagged-and-resolved apprentice-ratio violations, and the indexed audit-ready bundle, produced in a fraction of the manual time, with every legal attestation verified and signed by the human who owns it.

Key Takeaways

• Federal-funded construction payroll under Davis-Bacon asks a harder question than private payroll: not what the worker agreed to, but whether each worker was paid the prevailing basic rate and fringe for the classification of work actually performed, sworn weekly on the WH-347. The IIJA, CHIPS, and IRA funding wave made this the main event, with the IRA tying full clean-energy tax credits to prevailing-wage and apprenticeship compliance, so a payroll error can claw back a credit worth multiples of the labor.
• AI ingests the timecards, cross-checks each line's classification against the posted wage determination, computes the basic-plus-fringe rate math, and generates the WH-347 batch at machine speed, exactly the volume work the clerk-and-spreadsheet method cannot keep up with on a 140-craftworker, multi-sub crew.
• The statement of compliance is a legal attestation signed under penalty of the False Statements Act, so the verification gate sits between the AI's draft and the human signature, never after it: the dollars-and-contract-authority gate applied to federal compliance, where the AI drafts, the person attests, and the signature is never delegable.
• The wage determination is the source of truth, and misclassification is the most common and most expensive finding. The AI proposes the matching classification and flags ambiguous lines (split-rate days, "helper" descriptions that mask skilled work), but resolving a classification is a human judgment with back-wage and tax-credit consequence.
• The fringe-benefit component is the quiet trap: a finding arises when the claimed fringe credit exceeds actual contributions or annualization is computed wrong. The AI computes the credit and flags the divergence; the contractor verifies the contribution records behind the sworn representation before signing.
• The apprentice-to-journeyworker ratio is the structural killer: a payroll can be arithmetically perfect and still wrong because crew composition broke the ratio on a given day. The AI joins timecards, registrations, and daily crew composition to flag every out-of-ratio instance (including partial-day cases); the human confirms enrollment and journeyworker counts and decides the remedy.
• DBE/MWBE participation is a separate contract-performance obligation about dollars to certified firms, not a wage obligation, and conflating the two produces a memo that satisfies neither. The AI aggregates committed and paid dollars, computes participation against the contracting officer's goal, and drafts the memo; the human verifies certification status and owns the representation.
• The DOL-audit-ready records bundle is assembled to be examined before anyone asks: the WH-347 batch, line-by-line timecard ties, the governing determination, classification rationale, apprenticeship registrations and per-day ratio computations, and fringe reconciliation, all indexed for traceability. The AI's strength is complete, traceable assembly; the compliance owner reviews it before it goes to the investigator, because it supports a position taken to a federal agency.