AI-Trained Is a Supplier Claim. Put the Evidence in the Contract.

What does an agency mean when it says AI-trained?

AI-trained, on its own, says training happened, not that anyone measured what stuck. At least three different things can hide behind the phrase: a completion certificate, a vendor badge, and a self-rating a candidate typed into an intake form. All three are real records. None is a claim about the work that will arrive on your floor.

Google Cloud's Generative AI Leader certification says on its own page that it is for anyone in any job role, with or without hands-on technical experience: 90 minutes, 50 to 60 multiple-choice questions, $99, and no prerequisites ¹.

Passing it evidences vocabulary and one vendor's product line.

Credentials at the other end of the range ask for something else entirely. Microsoft's Azure AI Fundamentals exam page tells candidates they need knowledge of Python coding syntax and programming techniques, and familiarity with Azure resources ².

So ask which certificate before you read anything into the word.

The roster form of the claim is the weakest one. A sentence like every candidate here is AI-trained describes a purchase order, and it stays true long after the person who sat through the session has forgotten the session.

What you want is a claim about the individual being sent to you, which is a different sentence and a harder one for an account manager to write.

Ask what was assessed, not what was delivered

Three questions get you most of the way, and they fit in one email. What did the training cover, in tasks rather than topics? Who delivered it, and over how many hours? And what did each person have to do at the end to be counted as trained?

An attendance list answers none of it. An account manager who can answer the third question has something to sell you.

Then ask for one artifact per placed person. One page is enough: a task the person did with a tool, what the tool got wrong, and what they changed before the work went out. A supplier that assesses people on real tasks already holds the material for that page.

A November 2025 preprint points the same way, though not far enough to carry the ask on its own. Researchers building an AI literacy assessment for a US Navy robotics training programme found that, as a measure of applied AI literacy, a scenario task simulating AI use on the job outperformed the tests they had adopted from earlier work or written themselves ³.

Read the sample before leaning on it. Thirteen trainees in each training cohort, in one programme and one occupational setting.

The correlation behind that finding was computed on the ten participants who completed every measure, and the authors call the work limited in scope and sample size.

So the paper argues for a way to design a check and stops short of showing that the check works. The request for one artifact does not need it either way: a supplier either produces the page or does not.

One more question is worth the email: which tool. An agency that trained a cohort on one assistant and then sends them into a stack built on another is not wrong to call them trained.

But the judgment transfers across tools and the tool habits do not, so ask what a placed person will have to relearn in week one.

Which clauses turn the AI-trained claim into evidence?

Five clauses do it: named tasks, one artifact per person, one method for everyone presented, disclosure of automated screening, and a remedy you would use. None of them is a warranty that the people are good. A warranty buys an argument later. An evidence obligation gives you something to read now. Write the terms so a placement either arrives with the evidence or does not arrive, and so the standard is the same for everyone the agency puts forward.

• Name the tasks. The agreement lists what a placed person will do with an assistant in this role: draft a first-pass reconciliation, triage a ticket queue, summarise a record. The evidence obligation attaches to those tasks and survives the next model release.
• One artifact per person, before the start date. A single page from real or simulated work, naming what the tool produced and what the person changed.
• The same method for everyone presented. The supplier states its assessment method in writing and applies it to every candidate it puts forward for the role, not only to the ones you happen to ask about.
• Disclosure of any automated screening the supplier runs. Which tool, at which step, and what the supplier holds by way of an audit.
• A remedy you would actually use. A replacement at no fee inside thirty days where the evidence obligation was not met, which is a term agencies already understand.

The disclosure clause is not hypothetical in New York City. Since January 1, 2023, Local Law 144 has barred an employer or an employment agency from using an automated employment decision tool on a city candidate without a bias audit done in the prior year ⁴. It covers only a tool that substantially assists or replaces discretionary decision-making.

Notice has to reach the candidate at least 10 business days before the tool is used.

The law names the employment agency itself, not only the client.

Enforcement began July 5, 2023, and the statute sets penalties of $500 to $1,500 a day. Those figures are a ceiling written into the law. They say nothing about what anybody has been made to pay.

Disclosure puts a staffing agreement in front of the same problem as an assessment purchase. What comes back should look like the answer to what you ask a vendor who says they are bias tested: the audit, its date, the population it was computed on, and who ran it.

Written this way, the clauses change what the fee buys. Access to a roster and evidence about each person are two different products at the same price, which is why it is worth paying an agency for evidence rather than access.

Why the agency's screen can be partly your problem

The agency's screen can be partly your problem, because the federal definition of a selection procedure is written around the decision a step feeds and says nothing about who runs the step. The Uniform Guidelines, adopted by four federal agencies in 1978, define one as any measure, combination of measures, or procedure used as a basis for any employment decision ⁵. The same definition reaches everything from paper-and-pencil tests through informal or casual interviews and unscored application forms.

The EEOC said something close to this about vendors before it took the page down. Its Title VII technical assistance, issued in May 2023 and removed from the site in January 2025, took up whether an employer is responsible for algorithmic hiring tools designed or administered by somebody else.

The answer was "In many cases, yes" ⁶.

The same answer named agents directly: employers may be held responsible for the actions of their agents, which may include entities such as software vendors given authority to act on the employer's behalf ⁶. A staffing firm you have given authority to choose who reaches you could fit that description, though the document's own example was a software vendor.

The document was withdrawn and never had the force of law.

Staffing also has guidance of its own, and it predates the whole argument. In enforcement guidance issued December 3, 1997 under Notice 915.002, the EEOC set out when a staffing firm and its client are joint employers ⁷. Both are covered where both have the right to control the worker and each has the statutory minimum number of employees.

The same document says a client that rejects workers for discriminatory reasons is liable either as a joint employer or as a third-party interferer, where it has the requisite number of employees to be covered.

The day a charge arrives, what you need from an assessment vendor for a bias audit or an EEOC inquiry is the list you will want from the agency too. A records clause is cheaper to negotiate while the account is still being won.

Confirm the current position with employment counsel; none of this is legal advice.

Run your own check on the first five placements

Keep a twenty-minute exercise of your own and give it to the first five people the agency sends. The purpose is not to re-screen the supplier's work. It is to get a first signal on whether AI-trained on their roster tracks anything on your floor, which no clause can give you in advance and five placements can.

Build it from a task the job actually contains. Hand over a short piece of model output with one thing wrong in it that only checking catches: a total that does not reconcile, a policy line that does not exist, a delivery date the system contradicts.

Ask what they would send back, and why.

For frontline and hourly placements the exercise is smaller than it sounds. AI skill for an hourly hire is override and escalation, so the question is whether the person notices the app is wrong and knows who to tell.

Even a supplier with a genuinely assessed roster has other clients, and an evidence obligation is a reason to fill an easier requisition first. In a tight market, the clause you can enforce is the one you were willing to lose a week over.

The line in the agency's capability deck stays where it is. What changes is the sentence under it in the agreement, which now names what has to arrive with each person and what happens if nothing does.

Read the first artifact when it lands.

A claim about a roster was never checkable. One page about one person is.

See a sample report >