How Should Employers Run HR AI Vendor Due Diligence in 2026?

What HR AI Vendor Due Diligence Actually Requires

HR AI vendor due diligence is the process of testing whether a tool can perform its claimed HR function lawfully, accurately, securely, and fairly before employees or applicants use it. It is not a paper exercise focused on glossy product demonstrations, sales promises, or a list of security certifications. As of September 24, 2026, the review should cover hiring screening, workforce analytics, performance management, employee monitoring, generative assistants, automated compliance guidance, and any use of applicant-tracking technology that ranks, filters, or assists decisions about people. The central question is not simply, “Does the software work?” but “Can the employer explain what the software does, demonstrate control over its inputs and outputs, and document compliance with every law that applies to the use case?”

Also worth reading: How Can Employers Audit AI HR Compliance for Hiring, Monitoring, and Vendor Risk in 2026? · What should be included in an AI vendor due diligence checklist for labor law compliance and HR regulatory management in 2026? · What Laws Govern AI Hiring Decisions in 2026, and How Should Employers Manage Them?

That standard has become more demanding because employment AI rules are developing faster than federal legislation. Illinois employment AI regulations took effect on January 1, 2026, while New York City has required annual bias audits and advance notice for certain automated employment decision tools since July 2023. Colorado’s amended AI statute was scheduled to take effect on June 30, 2026, and the European Union AI Act originally scheduled its main application date for August 2, 2026, although specific timing and requirements have faced revision. The Delaware amendments discussed by Baker Donelson also raise privacy and AI-governance questions for employers operating across jurisdictions. Organizations should therefore evaluate the real deployment, not assume a vendor’s general compliance statement resolves every location-specific duty.

A defensible review brings together legal, HR, information-security, procurement, and business stakeholders. For smaller organizations, a qualified outside attorney or independent consultant may supply the missing specialist capacity. The conclusion should be documented, time-limited, and tied to intended use, because adding employee monitoring, an emotion-recognition claim, or an external data source can change the risk profile. No widely accepted “AI vendor seal” proves that a system is lawful, fair, and reliable in every setting. Due diligence builds evidence so that the employer can make and defend its own decision.

Map the Tool, Its Claims, and Its Legal Role

The first phase is to turn marketing language into a precise inventory of functions. Ask the vendor exactly what inputs it collects, what inferences it generates, whether it ranks or filters people, and whether a human can meaningfully change the result. “Assistive” is not automatically neutral: a tool that scores 20,000 applicants and leaves a recruiter with 50 automatically changes which applications receive attention. Likewise, an AI notetaker may create an accurate transcript yet still introduce privacy, consent, biometric, retention, and employee-relations concerns. Aon’s guidance on preparing for the EU AI Act and emerging employment rules supports mapping how systems operate before treating them as ordinary HR software.

The employer should classify the tool by function rather than by the vendor’s chosen category. Employment-related uses such as recruitment filtering, task allocation, promotion, termination monitoring, or performance evaluation receive greater scrutiny than tools with no employment effect, such as drafting a generic holiday newsletter. It is useful to record each decision point, the data subject, the geographic worksite, the vendor, the customer, and any subprocessors, including tools such as large language model providers that may receive copied text or audio. As of September 24, 2026, that inventory is more reliable than a one-time questionnaire because a platform may add features or clients without changing the product name.

This comparison is deliberately practical. A large suite may offer stronger centralized controls but still produce poor results if a customer selects the wrong hiring or monitoring settings. A point solution can be easier to test, but its narrow scope does not make it safer by default. External consultants cannot replace vendor cooperation, access to data, or technical documentation. Cost figures are planning ranges derived from typical enterprise purchasing patterns, not vendor quotes, and FORTUNE Business Insights’ broader HR software forecast should not be treated as a price for one compliance feature.

Test Security, Privacy, and Employee Data Controls

Security review must follow the actual data path, including the vendor’s infrastructure, support access, subprocessors, model providers, and any development or analytics environment. The employer should obtain current independent assurance rather than relying on a sales page that says “enterprise-grade.” Depending on the system, relevant evidence may include a SOC 2 Type II report, ISO 27001 certification, penetration-test results, incident history, disaster-recovery testing, and a recent letter or response to security questionnaires. A SOC 2 report is valuable, but it is not a guarantee that a specific AI output is accurate or non-discriminatory. The security team should examine exceptions, remediation dates, scope, and whether the service supporting the purchased module falls within the audited environment.

Privacy review should cover collection, purpose limitation, access, retention, deletion, and every cross-border transfer. The employer needs to know whether employee data is used to train vendor or third-party models, whether human support staff can view it, and whether the contract permits the types of secondary use the vendor actually performs. Privacy and AI governance analysis from Baker Donelson and Morgan Lewis reinforces the need to treat employment data as an active governance matter rather than a passive archive. International transfers may require an appropriate contract mechanism, and the vendor should provide current subprocessor and hosting information.

For monitoring, summarization, and conversation tools, the employer should test whether raw recordings are sent outside the organization, how long they remain available, and whether employees can correct the underlying notes. Biometric categorization, emotion inference, or health-related inference can trigger distinct restrictions in several jurisdictions and may create sensitivity even when the vendor labels the feature as optional. A data-protection impact assessment should be updated before rollout and revised when data categories, model versions, or retention rules change. Encryption in transit and at rest is a baseline expectation, not proof that an AI system is safe. The business also needs deletion instructions that work across backups, analytics stores, and subprocessors, because a vendor’s “delete” button may cover only part of the chain.

Demand Evidence About Accuracy, Fairness, and Human Oversight

Vendors frequently demonstrate a product using balanced, low-risk examples, but the employer should test the purchased configuration with employment data that reflects its workforce. Hiring tools should be assessed for rank-order reversals, drop-off differences, access to comparable opportunities, and the treatment of applicants who use accommodations or speech patterns outside the testing set. Performance and promotion tools require a different review because historical workplace ratings may already contain managerial bias. The purpose of testing is not to claim that a statistical metric proves fairness; it is to identify conditions under which the tool behaves poorly enough to justify safeguards or removal.

A useful validation plan establishes a baseline, defines unacceptable outcomes, and measures performance after a defined trial. For example, the employer could compare shortlist rates across relevant groups, test consistency by asking the same substantive application to be evaluated under limited formatting changes, and require stable outputs across repeated runs. If the system ranks résumés, swapping only a name should not materially alter the assessment unless a validated job-related reason supports that effect. If the tool summarizes performance feedback, reviewers should check whether invented facts, omitted criticism, or altered tone changes a manager’s understanding. These are examples, not universal statutory tests.

Human oversight is meaningful only when the reviewer has authority, competence, time, and enough information to contest an output. A recruiter who must approve 400 ranked applications per hour has little practical control, even if a policy says to ignore the ranking. The employer should document which decisions the AI may make, which remain prohibited, and how escalation works. Illustrative Performance Labs has published that the EU AI Act generally treats many employment-related AI as high-risk, with duties such as risk management, data governance, technical documentation, logging, transparency, and human oversight; the scheduled application date and related changes should be checked with EU counsel. Employers should not treat an AI-generated policy as automatically compliant with labor law.

Convert Compliance Claims into Contractual Rights

Contract language is where a persuasive demonstration becomes enforceable. Baker Donelson, Fisher Phillips, Aon, and other advisory publications consistently stress that the legal review must follow the function being purchased. The agreement should define the tool’s intended purpose, prohibited uses, data ownership, retention, deletion, incident notification, audit evidence, model-change notice, and cooperation with regulators or affected employees. A commitment to comply with “applicable law” is useful but not a substitute for specific service levels, permitted data uses, and remedies.

The contract should also allocate responsibility for consequential decisions. State exactly whether the vendor warrants that outputs comply with named laws, whether compliance depends on customer configuration, and what happens if an audit finds a material defect. Ask for indemnity or insurance information, but do not assume either eliminates the need to stop unlawful use. The employer should preserve the ability to suspend a model update, export records, transition data, and terminate the service without losing evidence needed for an investigation.

Common Due Diligence Mistakes

A frequent mistake is asking whether a product is “compliant” without specifying a country, workflow, decision, employee population, or data category. The answer is usually too broad to test. Another error is treating vendor-provided benchmarks as independent validation. A product can perform well on a curated benchmark but behave differently when recruitment data, job language, accents, disability-related accommodations, or employees’ ordinary work patterns are introduced. Security certifications, contractual promises, and fairness evidence are related parts of the review, but they are not interchangeable.

Organizations also overstate the value of a pilot. A small demonstration may include 20 test cases and 3 recruiters rather than 20,000 applications and 30 hiring managers using a shared queue. Some teams review only precision, ignoring the false-negative rate. Others ask for aggregate demographic information but fail to set a lawful collection purpose, consent basis, retention limit, or access threshold. The weakest approach is to deploy the tool before agreeing on controls, then ask the vendor to update a security document. That sequence places employee data at risk and weakens the employer’s position if a regulator, applicant, or worker later asks how the system was governed.

When to Act and How to Budget the Review

The clearest trigger for action is a planned deployment involving applicants, employees, contractors, or applicants’ personal data. A second trigger is an existing system whose vendor has changed models, ownership, subprocessors, or processing purposes. A third is an audit finding, employment claim, regulator inquiry, or incident involving the tool. Organizations should not wait for a federal AI employment law to supply a complete rulebook, because Illinois, Colorado, New York City, California, the European Union, privacy statutes, discrimination law, and sectoral duties already apply. The employer should aim to complete initial review before production data is ingested, not merely before the contract is signed.

A workable planning model allocates approximately 3% to 10% of a project’s first-year budget to legal, privacy, security, and fairness review, subject to the tool’s role and sensitivity. For an early-stage assessment, $10,000 to $50,000 may be reasonable when a limited pilot and a smaller vendor are involved. A high-impact hiring or monitoring deployment can justify $50,000 to $250,000 or more when independent testing, workforce consultation, and cross-border legal analysis are necessary. These are planning figures, not universal rates. FORTUNE Business Insights has projected the HR software market from approximately $38.96 billion in 2024 to $123.3 billion by 2032, using a roughly 15.7% compound annual growth rate for its stated forecast period. Growth creates more purchase options, but it does not reduce the employer’s responsibility.

The Recommended Due Diligence Decision

The strongest diligence process produces a written decision rather than a binary “approved” label. It should state what the system may do, which populations and jurisdictions are covered, what human review is required, which evidence supports the decision, and what will cause reconsideration. A conditional approval can be appropriate for a low-impact internal drafting tool with restricted data and no employment decision. A hiring-ranking system affecting thousands of applicants should ordinarily require more evidence, stronger controls, and a documented reason for continued use. An emotion-recognition or health-inference feature should receive a presumption of caution because its data and employment consequences are unusually intrusive. A contract should be signed only after the vendor has answered open technical questions in writing.

The final test is whether the employer can explain its decision to a regulator, an employee, an applicant, or a court without relying on phrases such as “the vendor certified it” or “the software was only a tool.” It should be able to identify the inputs, the decision workflow, the controls, the review process, and the person accountable for the outcome. This approach also avoids buying unnecessary AI: a spreadsheet plus accountable human review may be cheaper and more defensible for a small task. HR AI due diligence is not about finding one perfect vendor; it is about establishing which capabilities are justified, how risk is contained, and when the employer should stop relying on the system.