Policy

Vendor Indemnification Does Not Move Employer Liability. It Moves Who Pays.

Your AI hiring vendor's indemnity does not move discrimination liability off you. It is a promise about money between you and the vendor; a rejected candidate is not a party to it, so it neither decides who they can name nor shifts a duty a discrimination statute puts on the employer. California's FEHA regulations, effective October 2025, treat an agent operating the hiring system as an employer too, which adds a defendant rather than substituting one. The clause's worth turns on scope, caps, defense control, and the vendor's balance sheet.

The takeRead the indemnity as procurement, not as protection. A useful one is narrow and specific: it names third-party discrimination claims arising from the tool, it survives termination, it is not quietly capped at twelve months of fees, and it does not hand the vendor sole control of a defense you have to live with afterwards. Then turn to the data clause, which is the one that decides whether you can defend the decision at all. Most deals negotiate the first hard and accept the second as drafted.

Where Olive fits

Open a role and see what the work shows

A contract can move a cost, but it cannot move what you are able to show later. Every released Olive report exports with its rubric, scorer and item-bank versions attached, and the candidate is granted the identical report at no charge.

Rank your shortlist

What does an indemnity actually cover?

A defined set of third-party claims, and the money attached to them. A typical clause says the vendor will defend and pay claims brought by someone else arising from the tool, subject to exclusions and a cap. That is an allocation of cost between two contracting parties. Statutory duties are not property, so nothing in it transfers your obligations under an employment discrimination law.

Notice who the promise runs to. The vendor promises you, and the candidate is a stranger to the contract. A charge is filed against the employer named in the rejection, and the indemnity then becomes a second matter you pursue against the vendor, usually while the first one is still live.

The exclusions are where the clause narrows, and they tend to describe exactly what you did. Your configuration, your thresholds, your job requirements, your data, and any use outside the documented scope are all common carve-outs. So is a failure to tender the claim within a stated window, which is a live risk when a charge lands with HR and reaches procurement three weeks later.

Then check the ceiling and the source of funds. An indemnity capped at fees paid in the prior twelve months is a discount, not a defense budget, and it sits behind whatever the vendor can actually pay. Ask for the insurance behind the promise, the limits, and whether the policy names discrimination claims at all. A well-drafted clause from a company that folds is worth what the company is worth.

Who can a candidate name?

You, and sometimes the vendor as well. In the Title VII technical assistance it issued in May 2023, the EEOC answered whether an employer is responsible for a tool designed or administered by someone else with "in many cases, yes," and added that if the vendor is wrong about its own tool the employer could still be liable 1. That document came off the agency's site in January 2025, so read it as the agency's 2023 position, not current guidance.

State law has since said something firmer. California's FEHA regulations covering automated-decision systems took effect October 1, 2025, and an agent acting for an employer, including a vendor operating the system, is itself an employer under the Act 25. Illinois went to an effects standard: from January 1, 2026 it is a civil rights violation to use AI that has the effect of discriminating on a protected basis, with no requirement to prove intent 3.

None of that turns on what you knew. An effects standard asks what the tool did. An agency relationship attaches because the vendor acted for you, and good-faith reliance on a vendor's assurance does not work as a defense.

The litigation runs the same direction. In Mobley v. Workday, an order in July 2024 let disparate-impact claims proceed against the vendor on an agent theory, and in May 2025 the court granted preliminary certification of an ADEA collective 4. Nothing there is a finding that anyone discriminated, preliminary certification is a low bar that only permits notice, and it is an age case. What it demonstrates is the shape of the exposure: a claim that names both of you, with the employer still holding the decision.

Which clauses decide whether you can defend the decision?

The data clauses, not the indemnity. A discrimination defense runs on records you have to produce: applications and selections by stage, the tool's output at the time of the decision, the version that produced it, and the vendor's own testing. If the contract does not give you those on demand and after termination, the indemnity pays for a defense you cannot mount.

Four asks, in the order they get refused:

  • Your candidates' stage-level records, exported on request, in a documented format, during the term and for a defined window afterwards.
  • A version stamp on every output, so a score, list or recommendation can be tied to the ruleset that produced it.
  • Written notice before a model, ruleset or threshold change, because a change restarts the audit clock and invalidates the analysis you already ran.
  • The testing data behind the certificate. A summary states a conclusion. Ask for the counts underneath it.

That last one has a wrinkle worth raising with counsel early. A vendor's own bias testing may be protected work product in the vendor's hands, which means the number you would most want may be the number nobody produces. Plan the analysis you can run yourself, from data you hold, and make what you need from a vendor to survive an EEOC inquiry a procurement requirement, written into the contract. What you keep afterwards is a records question with its own rules.

Ask for these five things before signing

Bring the list to the vendor before the pricing conversation closes, because after signature every item on it is a change order. None of it is exotic, and all of it is refused often enough that a refusal is itself information about how the tool is built and how much the vendor knows about its own outputs.

1. Named scope. The indemnity covers discrimination and civil rights claims explicitly, not only intellectual property and confidentiality. 2. A separate cap. Discrimination claims sit outside the general liability cap, or above it. 3. Defense participation. You approve counsel, and no settlement characterizes your hiring practice without your consent. 4. Data rights. Export, retention and version stamping, with survival past termination. 5. Audit cooperation. The vendor supports your own analysis and any bias audit you owe, at a stated cost, within a stated time.

One caution about sequence. Diligence on the tool is a different exercise from diligence on the paper, and the paper cannot fix an instrument that does not hold up: work through what to ask a vendor claiming it is bias tested first, then negotiate. A strong indemnity attached to a weak tool buys a well-funded loss. Put the same five items on the renewal checklist, because the tool you signed for two years ago is not the tool running today.

Read the evidence

Common questions

Does an indemnity cover a state agency investigation?

Often not. Third-party claim usually means a suit or a charge brought by a person, and agency inquiries, subpoenas, audits and regulator investigations are commonly excluded or simply unaddressed. Those are the events most likely to arrive first. Ask for them by name, along with the cost of responding, because the expensive part of an early inquiry is producing records, long before anyone pays damages.

If the vendor can be sued directly, are we better off?

An added defendant is not a subtracted one. Two defendants can blame each other, discovery gets longer and more expensive, and the employer still made the employment decision that is being challenged. The practical benefit is narrower than the headline: a solvent co-defendant with its own counsel and its own documents, in a case you are still in.

Can we require the vendor to run the adverse impact analysis for us?

You can require it, and you should still think about who commissions it. A vendor-run analysis of the vendor's own tool answers a question the vendor chose. Ask for the underlying counts by stage as well as the summary, so your team or your counsel can compute selection rates on your candidates. Settle who directs the work and who receives the result before anyone runs a number.

Does the vendor's bias audit satisfy our obligations?

Only partly, and only in New York City, where the duty exists. A published audit describes the tool. The determination that your use of it is in scope, the public posting on your own site, and the candidate notice all stay with you. Nothing in a vendor audit answers whether your funnel produced a disparity, because it was not computed on your funnel.

Our insurance renewal now asks whether we use AI in hiring. What changes?

Answer it accurately, and treat the questionnaire as a document that will be read back to you. An inaccurate answer can become a coverage argument at the worst moment. Before the renewal, get an inventory of which tools touch selection decisions, including features switched on by default inside your applicant tracking system, so the answer describes what is actually running.

References

  1. 1. Select Issues: Assessing Adverse Impact in Software, Algorithms, and Artificial Intelligence Used in Employment Selection Procedures Under Title VII of the Civil Rights Act of 1964, Question 3 (archived capture, 2025-01-25) U.S. Equal Employment Opportunity Commission, via the Internet Archive Wayback Machine, 2023. web.archive.org Supports the EEOC's 2023 answer that an employer may be responsible for a vendor-built selection tool, cited as withdrawn guidance.
  2. 2. Final Unmodified Text of Proposed Employment Regulations Regarding Automated-Decision Systems (Attachment B), 2 CCR sections 11008, 11008.1 California Civil Rights Department, Civil Rights Council, 2025. calcivilrights.ca.gov Supports the October 1, 2025 effective date and the treatment of an agent operating an automated-decision system as an employer under FEHA.
  3. 3. HB3773 Enrolled (Public Act 103-0804), amending the Illinois Human Rights Act Illinois General Assembly, 2024. ilga.gov Supports the Illinois effects standard for AI in employment decisions, effective January 1, 2026.
  4. 4. Mobley v. Workday, Inc., 3:23-cv-00770 (N.D. Cal.) Civil Rights Litigation Clearinghouse, University of Michigan Law School, 2026. clearinghouse.net Supports the agent-theory order allowing claims against the vendor and the preliminary ADEA collective certification, with no merits ruling.
  5. 5. Rulemaking Actions - Civil Rights Council California Civil Rights Department, Civil Rights Council, 2025. calcivilrights.ca.gov The Council's own record of the automated-decision-system employment regulations: approved by OAL and filed with the Secretary of State, effective October 1, 2025.

5 sources, numbered by first appearance. How Olive sources claims

General guidance, not legal advice. Hiring rules differ by state and country and change often; check anything here against your own counsel before you act on it.

Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.

Back to answers

Open your first role Ten attempts a month against a live item bank, with a human-written report on every one.