Theo runs talent acquisition operations for a 1,200-person regional health system, leading a team of nine recruiters and two coordinators. Last spring his organization signed a one-year contract for an AI sourcing-and-screening platform that promised to cut time-to-fill on clinical roles. Eight months in, a rejected candidate in New York City asked whether an automated tool had screened her application and whether the employer had run a bias audit. Theo did not know the answer to either question, and neither did his vendor account manager. The contract said nothing about audits, nothing about data deletion, and nothing about who was liable if the tool discriminated. That gap is the subject of this lesson. Due diligence and contracts are not procurement paperwork you delegate and forget. They are the difference between a tool that protects your organization and one that quietly transfers the vendor's legal exposure onto your candidates and your employer.

Why the Employer, Not the Vendor, Carries the Risk

The single most expensive misconception in recruiting technology is that buying a tool offloads responsibility for what the tool does. It does not. Under nearly every relevant framework, the employer remains the accountable party. New York City's Local Law 144 governs automated employment decision tools, and it places the obligation to obtain a bias audit and to provide candidate notice on the employer that uses the tool, not solely on the vendor that built it. If a vendor never produced an audit, the employer is the one out of compliance. The same logic holds for federal anti-discrimination law. EEOC guidance and the Americans with Disabilities Act apply to your hiring outcomes regardless of whether a human or an algorithm produced them.

Be concrete about what that allocation means when something goes wrong. The candidate's question lands on the employer, the charge names the employer, and the regulator asking to see a published audit summary is asking the employer to produce it. The vendor enters afterwards and only as a party you may have a commercial claim against, assuming the contract created one. You cannot contract away Title VII or ADA liability. A vendor can promise its tool is fair, but if the tool screens out qualified candidates with disabilities, the employer answers for it.

This is why due diligence and contract terms matter so much. They are the only levers you have to pull the vendor's evidence, documentation, and cooperation toward you before something goes wrong. Once a candidate has asked the question your leverage has collapsed: the vendor has your signature, renewal is months away, and the evidence either exists in a form you can publish or it does not. Theo learned this in reverse order. The right sequence is to assume you are liable, then write the contract so the vendor must give you everything you need to discharge that liability.

Classify the Tool by What It Decides

Not every tool carries the same risk, so due diligence should scale to what the tool decides. Sourcing and matching tools surface or rank candidates from a database or the open web. Screening and assessment tools score resumes, administer skills or personality assessments, or rank applicants against a job profile. Interview tools transcribe, summarize, or in some cases analyze video and audio. Chat and scheduling tools handle candidate communication and logistics. Background and verification tools sit adjacent to recruiting but touch highly sensitive data. These categories are a sorting aid rather than a legal test; their value is that they stop a team applying identical scrutiny to a calendar bot and to a ranking engine.

The dividing line that matters legally is whether the tool substantially assists or replaces a human employment decision. A scheduling chatbot that books interviews is low-risk. A tool that scores and ranks applicants so recruiters review only the top of the list is, in NYC's terms, very likely an automated employment decision tool, and it triggers the audit and notice obligations. What makes the ranking case decisive is that the recruiter may retain formal authority over every reject, but if the ordering determines which applications a human ever opens, the tool has already made the consequential cut. Authority on paper and effect in practice are different things, and the second is what gets examined.

When you evaluate a vendor, classify the tool by what it decides, not by how the salesperson describes it. Many vendors market a ranking engine as mere "efficiency" to dodge the regulatory label, and the language is persuasive because "we surface the best matches faster" describes the same mechanism as "we decide which applications a recruiter sees" without sounding like it. Your classification, not theirs, is what a regulator will scrutinize. Write it down at evaluation time, with the reasoning, so the judgment is on record before anyone has an incentive to remember it differently.

A Due Diligence Scorecard You Can Reuse

Theo rebuilt his evaluation process around a written scorecard so that every vendor faces the same questions and no charismatic demo skips a category. He scores each vendor on six dimensions, zero to five, and sets a hard floor: any vendor scoring zero on bias-audit evidence or data security is disqualified regardless of total. The floor is the part that does the work, because a weighted total invites trading, and a strong showing on integration will otherwise quietly compensate for a missing audit. That is precisely the trade that cannot be made: integration convenience is recoverable later and a discrimination finding is not.

The six dimensions are bias-audit evidence, data security and hosting, data processing and deletion terms, transparency about how the model works, indemnification and liability, and integration and exit. For each, he documents the question asked, the evidence the vendor supplied, and the score. Recording evidence separately from score is deliberate: a score is an opinion that decays, while "supplied an independent audit dated within the last year, with selection rates broken out by sex and by race and ethnicity" is a fact that either was or was not true on the day you asked, and it is what you hand to counsel a year later.

Consider how two finalist sourcing vendors fared on the same rubric. Vendor A scored 4 on bias-audit evidence because it supplied a recent independent audit with selection rates broken out by sex and by race and ethnicity, plus the date and the auditor's name. It scored 5 on security with documented encryption at rest and in transit and a current SOC 2 Type II report. It scored 4 on data terms because it offered a signed Data Processing Agreement and a 30-day deletion guarantee after contract termination. It scored 3 on transparency, 3 on indemnification, and 4 on exit, for a total of 23 of 30. The two threes are worth noticing: A was not perfect, and the scorecard's job is to make an imperfect choice legible rather than to produce a flawless winner.

Vendor B demoed better and quoted 20 percent less. But it scored 1 on bias-audit evidence, offering only a marketing claim that its model was "designed to be fair" with no audit document. It scored 2 on data terms because it would not commit in writing to deleting candidate data after termination, and 1 on indemnification because its contract capped its total liability at the value of one month's fees and excluded discrimination claims entirely. Its total was 11 of 30, and Theo treated the bias-audit score as failing the audit floor outright.

Theo chose Vendor A. The 20 percent saving on Vendor B would have evaporated the first time a candidate filed a complaint and the contract left his health system holding the entire claim. Read B's three answers as a set and they tell one story: a vendor that will not produce an audit, will not put deletion in writing, and excludes discrimination claims from its indemnity has made the same judgment three times about how far it wants to stand behind its own tool.

Reading a Bias Audit Against the Four-Fifths Rule

A bias audit is only useful if you can read it, and many recruiters accept a green summary page without checking the math. The most common screen is the four-fifths rule, also called the 80 percent rule, drawn from the federal Uniform Guidelines on Employee Selection Procedures. It compares selection rates across groups: take the selection rate of each group, find the highest, and divide every other group's rate by it. Any ratio below 0.80 is a signal of potential adverse impact that demands investigation. The arithmetic is deliberately simple, which is the point: anyone on a recruiting team can run it without specialist help, on numbers the audit should already contain.

Suppose a vendor's audit reports that a screening tool advanced candidates at these rates: men 50 percent, women 42 percent, white candidates 50 percent, Black candidates 36 percent, Hispanic candidates 44 percent. The highest rate is 50 percent, so every other rate is measured against it.

The audit may still display an overall "pass," and understanding how that happens is what protects you from it. An aggregate summary can absorb one group's shortfall inside a larger population that clears the threshold comfortably, which is exactly what a group-level breakdown is for. The 0.72 impact ratio for Black candidates is the kind of finding that should stop a purchase or trigger a hard conversation about remediation, and it is invisible in any number that is not broken out. Two of the three ratios here are fine, which is what a real audit usually looks like; the discipline is checking every one rather than stopping at the first reassuring result.

If a vendor's audit does not break out rates by group, or reports only an aggregate score, it is not giving you what Local Law 144 expects you to publish, and it is not giving you what you need to defend the tool. Treat a missing breakdown as a failing grade, not a formatting quirk. The request is usually met with an offer to send a cleaner summary, which is the opposite of what you asked for. What you need is the underlying selection rates by group, the date, and the auditor's identity, and if those cannot be produced during a sales cycle, they will not appear on demand during an investigation.

The Contract Clauses That Actually Protect You

Due diligence tells you which vendor to pick. The contract is what holds the vendor to its promises once the demo glow fades. Everything a vendor said during evaluation was said by people whose job ends at signature, to people whose leverage ends at the same moment. Several clauses do the real work of converting those statements into obligations.

Bias-audit delivery. Because the employer carries the Local Law 144 obligation, the contract must require the vendor to supply audit results on a defined cadence, typically annually, in a format detailed enough to publish and to provide candidate notice. A vendor who will only say its tool "passed an audit" is asking you to publish a claim you cannot substantiate, which puts your organization's name on an assertion whose evidence you have never seen. Require the underlying selection-rate data and the auditor's identity, and specify the format, because "an audit will be provided" and "selection rates by group, auditor named, in a form suitable for publication" are different promises.

Data processing agreement. If any candidate data falls under the GDPR, Article 28 requires a written Data Processing Agreement, and the vendor is your processor while you remain the controller. That is the same allocation pattern as the liability above: the party deciding why and how data is processed is the party the obligations attach to. The DPA must specify the purpose of processing, the categories of data, security measures, sub-processor rules, breach notification timelines, and deletion or return of data at the end of the engagement. For a US-only employer this may not be legally mandatory, but the same structure is best practice everywhere because it pins down deletion and breach obligations in writing.

Data ownership and deletion. State plainly that candidate data belongs to the employer, that the vendor may not use it to train models for other clients without explicit consent, and that all data is deleted or returned within a fixed window, 30 days is a reasonable target, after termination. The training restriction is the clause most often absent and the hardest to remedy afterwards: once candidate data has contributed to a model serving other clients, there is no clean way to withdraw it. Theo's original contract was silent here, which meant his candidates' resumes could in principle have lived on the vendor's servers indefinitely.

Indemnification and liability. The vendor should indemnify the employer for claims arising from the vendor's own negligence, security failures, or a defect in its tool. Watch two traps: a liability cap set to a trivial amount such as one month's fees, and a carve-out that excludes the exact claims you care about, often discrimination or data-breach claims. They are frequently combined, and each looks reasonable alone in a document written to be skimmed. Remember the hard limit, though. Even an ironclad indemnity clause does not erase the employer's direct statutory liability to a candidate under Title VII, the ADA, or Local Law 144. Indemnification shifts cost between you and the vendor. It does not shift your duty to the candidate or the regulator.

Transparency and audit rights. Reserve the right to audit the vendor's compliance, to receive documentation of how the tool makes decisions, and to be notified of material changes to the model, since a model update can change selection rates and quietly invalidate last year's audit. That consequence is easy to underrate, because nothing visible happens when a model is updated. The interface is unchanged, the scores still arrive, and the audit on file still carries a date within the year, while the thing it measured no longer exists in the form it was measured in.

Theo's Revised Playbook

When the contested renewal came up, Theo did not simply re-sign. He ran the incumbent vendor through the same six-dimension scorecard he now applies to every tool, and he insisted on three contract changes before renewing: an annual bias audit delivered with group-level selection rates, a 30-day data-deletion guarantee backed by written confirmation, and an indemnification clause that no longer excluded discrimination claims and no longer capped liability at one month's fees. Renewal is the moment when leverage briefly returns, and it is the only scheduled opportunity most buyers get to reopen terms accepted in a hurry.

The vendor agreed to two of the three immediately and negotiated the third. The one it refused, broad indemnification for discrimination claims, told Theo something useful about how confident the vendor actually was in its own tool. A vendor that markets fairness and then declines to stand behind it financially has priced its own risk, and that refusal is a more reliable signal about the product than anything in the marketing. It does not automatically end the relationship, but it tells you the residual exposure sits with the employer, and it should change how much human review you build around the tool's outputs.

The deeper lesson for Theo's team was procedural. Due diligence is not a one-time gate at purchase. It is a recurring obligation, because models change, laws change, and a tool that was compliant last year may not be this year. He now schedules an annual review of every active vendor against the scorecard, ties renewal to delivery of a current audit, and keeps a one-page record per vendor so that when a candidate or a regulator asks the question that stumped him last spring, the answer is already on file. That record is the whole difference between the two versions of Theo: not more knowledge, but a place where the knowledge already lives when someone asks.

Anti-Patterns

Treating the purchase as a transfer of risk. This is buying a screening or ranking tool on the understanding that the vendor built it, so the vendor answers for it. It happens because that is how responsibility works for most purchased software, and because the vendor's materials encourage the reading. What goes wrong is that under Local Law 144 the audit and notice obligations attach to the employer that uses the tool, and EEOC and ADA duties attach to your hiring outcomes regardless of what produced them, so the first candidate question arrives at a team with no evidence and no contractual right to demand any. The counter is to invert the assumption before you evaluate anything: assume you are liable, then write the contract so the vendor must supply what you need to discharge that liability.

Accepting the vendor's classification of the tool. This is recording a scoring and ranking engine as an efficiency aid because that is how the salesperson described it and because a human still clicks the reject button. It happens because the description is accurate as far as it goes and because the less regulated label is more convenient. What goes wrong is that a tool determining which applications a recruiter ever opens has already made the consequential cut, and the audit and notice obligations that follow do not wait for anyone to relabel it. The counter is to classify by what the tool decides in practice, and to write the classification and its reasoning down at evaluation time.

Reading the audit summary instead of the audit. This is accepting a green summary page or an overall pass as evidence that a tool is fair. It happens because the summary is what vendors hand over, because it is designed to be reassuring, and because checking the math feels like distrust of a partner you are about to sign with. What goes wrong is that an aggregate figure can absorb one group's shortfall entirely, so a 0.72 impact ratio sits inside a document that says "pass" and nobody sees it until a regulator or a candidate does. The counter is to require group-level selection rates, run the four-fifths division yourself, and treat a missing breakdown as a failing grade.

Treating indemnification as compliance. This is negotiating a strong indemnity clause and concluding that the discrimination risk is now handled. It happens because the clause is genuinely valuable, because it took real effort to obtain, and because "the vendor indemnifies us" sounds like an answer to the question a regulator would ask. What goes wrong is that indemnification moves cost between two commercial parties and does nothing to your direct statutory duty to a candidate under Title VII, the ADA, or Local Law 144, so a team can be fully indemnified and fully out of compliance at once. The counter is to hold both facts together: the clause is the financial backstop, and the audit, the notice, and the human oversight are the compliance.

Diligence as a purchase gate rather than a cycle. This is running a thorough evaluation once, signing, and then renewing on the strength of the original decision. It happens because the evaluation was genuinely rigorous and because renewal arrives as an administrative task rather than a decision. What goes wrong is that models are updated and laws change, so last year's audit can describe a version of the tool that no longer exists, and the file that once proved compliance stops describing reality while continuing to look complete. The counter is an annual review of every active vendor against the same scorecard, renewal conditioned on a current audit, and a contractual right to notice of material model changes.

Practice

These exercises follow the order Theo worked in, and the first one is what makes the rest possible.

• Inventory and classify every tool that touches a candidate. For each AI-assisted tool in your process, write one sentence on what it decides, not what it is called. Mark the ones that score, rank, or filter applicants before a human sees them, and note whether you hold a bias audit dated within the last year.
• Build the six-dimension scorecard and set your floors. Write out the six dimensions, the question you will ask for each, and the evidence that would earn a high score. Then set your disqualifying floors in writing, before any vendor is in front of you, so they are not being set by someone who has already seen the demo.
• Run the four-fifths check on an audit you already have. Take the most recent audit for a tool in production, find the selection rates by group, divide each by the highest, and write down every ratio. If the document has no group-level rates, record that, because it is the finding.
• Read your live contract against the clauses above. Search your vendor agreement for audit delivery, a data processing agreement, data ownership and a deletion window, indemnification with its cap and carve-outs, and change notification. Note which are absent and which are present but written so weakly they promise nothing specific.
• Draft the one-page vendor record and the renewal ask. For one active vendor, produce the record you would want on file if a candidate asked Theo's two questions tomorrow. Then list the contract changes you would require at renewal, ordered by what you would refuse to sign without.

Reflection

• If a rejected candidate asked today whether an automated tool screened her application and whether you had run a bias audit, how long would it take you to answer, and where would you look?
• Which of your tools would you have classified as an automated employment decision tool if the vendor had not described it otherwise?
• When you last chose between vendors, what actually decided it, and would that factor survive being written down next to the audit evidence?
• What does your current contract oblige the vendor to give you that you have never asked for?
• If your vendor updated its model next month, how would you find out?
• Which vendor in your stack has declined to stand behind its own fairness claims financially, and what have you changed about how you use it as a result?

Glossary

• Automated employment decision tool. The category covered by New York City's Local Law 144, turning on whether the tool substantially assists or replaces a human employment decision rather than on how the vendor markets it.
• Local Law 144. The New York City law governing those tools, which places the obligation to obtain a bias audit and to provide candidate notice on the employer that uses the tool, not solely on the vendor that built it.
• Bias audit. An independent assessment reporting selection rates broken out by group, with a date and a named auditor. A summary asserting an overall pass is not a substitute for the underlying rates.
• Four-fifths rule. Also called the 80 percent rule, drawn from the federal Uniform Guidelines on Employee Selection Procedures. Divide each group's selection rate by the highest group's rate; any ratio below 0.80 is a signal of potential adverse impact that demands investigation.
• Due diligence scorecard and hard floor. A written rubric scoring every vendor on the same six dimensions, zero to five, plus a minimum on specific dimensions, in Theo's case bias-audit evidence and data security, below which a vendor is disqualified no matter how strong its total or its price.
• Data Processing Agreement. The written agreement GDPR Article 28 requires between controller and processor, specifying purpose, data categories, security measures, sub-processor rules, breach notification timelines, and deletion or return of data at the end of the engagement.
• Indemnification. A contractual promise by the vendor to cover claims arising from its own negligence, security failures, or tool defects. It shifts cost between the parties and does not shift statutory duty to a candidate or a regulator.
• Liability cap and carve-out. A cap is a ceiling on what a vendor can owe, sometimes set to a trivial amount such as one month's fees. A carve-out removes specific claim types from the indemnity, most damagingly discrimination or data-breach claims, which are the exact categories the clause was sought for.
• Change notification right. The contractual right to be told about material changes to the model, because an update can change selection rates and quietly invalidate an audit that still carries a recent date.

Related Lessons

• Vendor and Tool Selection -- Evaluating AI Solutions covers the earlier stage of the same decision, choosing which tools to shortlist at all, which is the pipeline this lesson's scorecard filters.
• Data Privacy Fundamentals: GDPR, CCPA, FCRA, and Regional Requirements develops the privacy frameworks the contract clauses here are written against, including the controller and processor allocation.
• Retention and Deletion: Reasonable Timelines and Clean Data Practices is what the 30-day deletion clause connects to, since a deletion promise from a vendor is only meaningful inside a retention schedule you actually run.
• Compliance Risks and Legal Exposure develops the exposure this lesson allocates, covering the categories of risk that make an indemnity carve-out expensive rather than technical.
• Documentation and Evidence: Building a Trail for Compliance is the discipline behind the one-page vendor record, since diligence you cannot produce on request is diligence you did not do.
• Legal and Compliance Partnerships: Ensuring AI Use Is Defensible takes the contract review beyond a recruiter reading clauses into an ongoing working relationship with counsel.

Closing

The question that stumped Theo was not a hard one. A candidate asked whether an automated tool had screened her and whether the employer had run a bias audit, and both answers were facts about a system his organization had chosen, paid for, and operated for eight months. He could not answer because nothing in the purchase had required anyone to find out, and by then the moment when the vendor would have answered easily was long past. That is the shape of this failure almost every time: not a bad decision made knowingly, but a decision made without asking, protected by a contract nobody read as an instrument of compliance.

The corrective is unglamorous and entirely learnable. Assume the liability is yours, because under Local Law 144, Title VII, and the ADA it substantially is. Classify each tool by what it decides rather than by what it is called. Score every vendor against the same six dimensions with floors you set before the demo. Read the audit rather than the summary, and do the division yourself. Insist on the clauses that turn sales assurances into obligations. Then treat the whole thing as a cycle rather than a gate, because the model, the law, and the selection rates all move while the file on your shelf stays still.

Key Takeaways

• The employer carries the liability, not just the vendor. Under NYC Local Law 144 the employer must obtain the bias audit and provide candidate notice, and EEOC and ADA obligations cannot be contracted away. Buying a tool does not buy you out of responsibility for what it decides.
• Classify the tool by what it decides. A ranking or scoring tool that substantially assists an employment decision is an automated employment decision tool and triggers audit and notice duties, no matter how the vendor markets it. Scale your diligence to the decision the tool makes, and record the classification and its reasoning at evaluation time.
• Use a written scorecard with hard floors. Score every vendor on bias-audit evidence, data security, data terms, transparency, indemnification, and exit, and disqualify any vendor that fails the audit or security floor regardless of price or demo quality. Record the evidence alongside the score, because the evidence is what you will hand to counsel later.
• Read the bias audit, do not just accept the summary. Apply the four-fifths rule to group-level selection rates. A ratio below 0.80 for any group is a signal of adverse impact that demands investigation, and a missing breakdown is a failing grade rather than a formatting quirk.
• Demand the right contract clauses. Bias-audit delivery in publishable detail, a GDPR Article 28 Data Processing Agreement where applicable, employer data ownership with a fixed deletion window and a restriction on training other clients' models, real indemnification without trivial caps or discrimination carve-outs, and audit and change-notification rights.
• Indemnification shifts cost, not duty. A strong indemnity clause protects you against the vendor, but your direct statutory duty to the candidate and the regulator remains yours. Treat the clause as a financial backstop, never as a compliance substitute.
• Diligence is recurring, not one-time. Models, laws, and selection rates change. Review every active vendor annually, tie renewal to a current audit, and keep a one-page record per vendor so you can answer the candidate's question before it is asked.

Frequently Asked Questions

The vendor says its tool is not an automated employment decision tool. Can I rely on that? Not as your own determination. The classification turns on whether the tool substantially assists or replaces a human employment decision, and the obligations that follow attach to the employer using it. A vendor describing a ranking engine as an efficiency feature is not lying about the mechanism, it is choosing the framing with the lighter regulatory consequence, and that consequence is not the vendor's to bear. Make your own classification, base it on what the tool decides in practice rather than on who formally clicks reject, and write the reasoning down while you still have no stake in the answer.

What if a vendor simply refuses to provide a bias audit? Treat that as the answer to your question rather than an obstacle to work around. The obligation to hold an audit sits with you, so a vendor who will not supply one has asked you to carry a duty you have no means of discharging. Theo's floor exists for this case: the disqualification is not a negotiating posture but a recognition that no discount compensates for operating a screening tool you cannot defend. If the tool is genuinely irreplaceable, the alternative is to commission the audit yourself and price that into the deal.

We have a strong indemnification clause. Does that cover us on discrimination claims? It covers cost between you and the vendor, assuming no cap that makes it nominal and no carve-out excluding discrimination specifically, which are the two traps worth checking first. What it does not touch is your direct statutory liability to a candidate or a regulator under Title VII, the ADA, or Local Law 144. Those duties are not commercial and cannot be reassigned by agreement. A team can therefore be fully indemnified and fully out of compliance at once, so the clause belongs in your financial planning and never in your compliance answer.

How often should we redo diligence on a tool that is already working fine? Annually at minimum, tied to renewal so it has a natural forcing point. Nothing in the tool announces the changes that matter: a model update can shift selection rates without altering anything a recruiter sees, and an audit on file can show a date within the year while describing a version that no longer exists. Contract for notice of material changes so the silent events become visible, and run the same scorecard you used at purchase so this year's answer is comparable to last year's rather than a fresh impression.