Your Answer to the EEOC Is Also a Letter to the Applicant Suing You
Most AI hiring compliance work is written for the period before anything happens. This is the other period: a charge has been filed, it names your screening tool, and the response you file can be handed to the person who filed it.
The first two weeks are about records, not arguments. Every strong position in this process rests on three artefacts: a reconstruction of the one decision, selection rates from your own data, and validity evidence for the cut score you applied. None of the three can be produced later if the vendor's retention timer runs first.
What the Clock Actually Says
Notice arrives fast, and it is not the deadline
An employer is normally notified of a charge within about ten days of filing. The response date on the cover letter is the agency's, not a statute's, and it is routinely extendable once for a stated reason. What is not extendable is the underlying preservation duty: from the moment you know of the charge, the records relevant to it have to stop moving. With a human screen that means files. With a model it means scores, thresholds, model versions and the vendor's retention timer.
The filing window tells you how far back the file reaches
Charges under the federal employment statutes must generally be filed within 180 days of the act complained of, extended to 300 days where a state or local fair employment agency has jurisdiction. Read this as scope, not as a defence: a charge filed on day 295 in a deferral state is asking about a screening run that is nearly ten months old, which is longer than many AI vendors keep per-candidate scores by default.
One applicant's charge can become a population question
An individual charge alleging that a tool screened one person out invites the obvious follow-up — what did the tool do to everyone else. Investigators can and do expand from an individual allegation into a pattern inquiry, and the agency has compulsory process available if a request for information is refused. An automated screen is the ideal subject for that expansion, because the same rule ran on every applicant.
Mediation is offered early and closes the file without findings
Voluntary mediation is usually offered before investigation and, if it resolves, produces no investigative record and no finding. It is worth a real decision rather than a reflex, and the decision depends on something you may not know yet: whether your own selection-rate data would help you or bury you. Find that out before you answer the offer, not after.
Five Things You Will Be Asked For
The narrative is the part employers prepare for. The data requests are the part that decides the file.
The position statement
Your narrative account of the decision and the reason for it. The important structural fact about it, which surprises employers every time: the charging party can obtain the position statement and its non-confidential attachments during the investigation. You are writing, simultaneously, to an investigator and to the person who may sue you — and to the lawyer reading it for an admission about how the tool works.
Confidential material, segregated or not protected
Sensitive attachments — commercially confidential information, trade secrets, medical or personnel data about other people — are treated as confidential only if they are separated out and labelled as such, with an explanation of why. Confidential text pasted into the body of the statement is simply part of the statement. This is the mechanism most often fumbled with vendor material, because the vendor's blanket confidentiality footer does nothing on its own.
Applicant flow data
Who applied, who advanced, who was rejected, and at which stage — with the demographic detail needed to compute selection rates. This is the request that decides most AI screening files, and it is the one an employer most often cannot fill, because the tool's stage sits between the applicant tracking system's records and nobody's.
The tool: what it scored, on what, and who set the cut
Expect questions about what the system measured, what data it was built on, what threshold was applied to this requisition, who chose that threshold, and what validity evidence exists for using it for this job. A vendor data sheet is not an answer to any of those five, and the last two belong to the employer alone.
The human step, described accurately
Who reviewed the model's output, what they saw, and what they were able to change. Overstating this is the trap. If the logs show recruiters approving a ranked list without ever moving anyone, an investigator reading that next to your statement has a conflict between your description and your records, and the conflict is now in writing.
Why the Vendor Holds Half of Your Answer
The scores may already be gone
Per-candidate scores are frequently retained on the vendor's schedule, not yours, and the schedule is frequently shorter than the charge window. There is no version of this conversation that improves after the data ages out, and 'our vendor deleted it' is not a neutral answer — the absence sits on the party with the recordkeeping duty.
The model that scored the applicant no longer exists
Vendors update models continuously. If the version that ran on the charging party was replaced twice since, rescoring them today answers a different question than the one asked. Ask the vendor, in writing, whether the specific version is retrievable and whether the scoring run is reproducible. The answer shapes everything downstream.
The bias-audit report is about the tool, not your requisition
A vendor's audit typically reports impact ratios over its own population at its own thresholds. Your file is about your applicant pool, your requisition and the cut score your team chose. The report can be useful context and it cannot substitute for the selection rates in your own data.
The contract may not let you produce what you need to produce
Check for the clause that lets you obtain and disclose per-candidate records for a legal proceeding, and for a support obligation with a named turnaround. If neither exists, the negotiation happens under a deadline, which is the worst possible time. This is a procurement fix, not an investigation fix.
Their confidentiality label does not bind the agency's process
Material is handled as confidential because it is segregated and justified under the agency's own process, not because a vendor stamped it. If real trade-secret material has to go in, say what it is and why — and tell the vendor in advance that a stamp alone will not hold.
The First Two Weeks, in Order
Freeze everything the tool touched, including the vendor's copy
Notice to your own systems is the easy half. Send the vendor a written preservation instruction naming the requisition, the date range and the charging party, and get an acknowledgement. Undirected deletion on a vendor's normal schedule after you knew about the charge is the avoidable version of this problem.
Compute your own selection rates before you write a word
Pass-through rates by stage and by group, in your data, for this requisition and for the tool's other recent uses. You are going to be asked. Knowing the answer changes whether you mediate, what you concede and what you say about the threshold — and it is far better to find a problem yourself than to have an investigator hand it to you.
Reconstruct the decision for this one person
Inputs received, output produced, model version, threshold applied, stage at which they were dropped, and the human actions after that. One page. If you cannot build it, that is the single most important finding of the first fortnight, and it needs to reach whoever owns the vendor relationship immediately.
Write the statement for two readers
Accurate to the records, specific enough to be credible, and free of any characterisation of the tool you cannot support with documents. Avoid describing the model as 'validated' or 'bias-tested' without naming the evidence — an unsupported adjective in a position statement is a gift to the person who gets to read it.
Notify the insurers, on time
Employment practices coverage is typically claims-made with a prompt-notice condition, and a charge is usually a claim. Late notice is one of the few own-goals in this process that no amount of good facts repairs. Notify, in writing, on the schedule the policy names.
Questions Employers Ask
Does the charging party really get to read our position statement?
Yes — the agency's practice is to provide the respondent's position statement and its non-confidential attachments to the charging party on request during the investigation, so that they can respond to it. Two consequences follow for an AI screening file. First, every description of how the tool works is a description given to a potential plaintiff, and it will be compared line by line with whatever discovery later produces; an approximation written for speed becomes an inconsistency. Second, material you genuinely need to protect has to be separated into labelled attachments with a stated reason, because confidential content placed in the body of the statement is part of the statement. Write the narrative so that it is complete and accurate without the confidential exhibits, then attach the exhibits properly.
Our vendor scored the applicant. Can we point the agency at them?
You can and should identify the vendor and the system, but it does not transfer the obligation. The charge is against the employer that made the employment decision, the recordkeeping duties attach to the employer, and an answer that amounts to 'the tool did it and we cannot see inside' reads as an admission that you deployed a selection procedure you cannot explain or defend. The practical move is to make the vendor a supplier of evidence rather than a shield: a written request naming the requisition and date range, asking specifically for per-candidate inputs and outputs, the model version that ran, the threshold in force, any validity evidence for this use, and their retention schedule. Keep the request and the reply; a vendor's refusal in writing is itself a document you will want.
We never collected demographic data, so we cannot compute selection rates. Is that a defence?
No, and it is closer to the opposite. Federal recordkeeping rules for selection procedures expect an employer to be able to show the impact of the procedures it uses, and an employer using a screening tool without any ability to measure its selection rates is an employer that cannot answer the central question about it. Not collecting the data does not make the impact question go away; it moves the question from your evidence to the agency's inferences and to whatever the investigator can reconstruct from names, locations and other proxies. The fix is structural and not retrospective: collect the demographic data properly, separated from the hiring decision-makers, so that impact can be measured by the people responsible for monitoring it.
Should we take the mediation offer?
Decide it on your own data rather than on principle. Mediation resolves the matter without an investigative record and without a finding, which is worth a great deal when your selection rates are unflattering, when the model that ran cannot be reproduced, or when the per-candidate records have aged out at the vendor. It is worth much less when you can produce a clean reconstruction of the decision, defensible selection rates and real validity evidence for the cut score — in that position, an investigation that closes without cause is a stronger asset against the next charge than a settlement. The mistake is answering the offer before anyone has computed the pass-through rates, because that is the fact that should drive the answer.
One applicant complained. Why is the investigator asking about every applicant?
Because an automated screen applies the same rule to everyone, which makes the population question both obvious and cheap to ask. An individual charge can be the entry point to a broader pattern inquiry, and the agency has compulsory process if a request for information is resisted. Resisting on scope alone rarely ends the question; it usually converts an information request into a subpoena and adds a fight about your cooperation to a file that was previously about a screening rule. The better posture is to know your own population numbers first, produce them in a form you have checked, and reserve the scope argument for requests that are genuinely untethered from the tool and the period in the charge.
The One-Page Test
Pick any candidate your screening tool rejected last quarter. Ask whoever owns the tool for one page: inputs, output, model version, threshold, stage, and what a human did afterwards. Give them a day.
If the page comes back, you can answer a charge. If it comes back with gaps, you have found them on your own schedule rather than on an investigator's — and the cheapest week to fix the vendor contract is a week when nobody has filed anything.
Related Reading
- AI hiring recordkeeping — what to retain, for how long, and the contract terms that make it possible.
- Validation evidence and job analysis — what "validated" has to mean before you write it down.
- Disparate impact and state law — where the exposure sits when federal priorities move.
- Employment practices coverage — the notice condition that decides whether the policy responds.