Soraya leads talent acquisition at Northwind Health, a 2,400-person regional hospital network, and last quarter her team processed roughly 6,000 applications across nursing, allied health, and back-office roles. Her CHRO had just signed off on an AI screening tool that promised to cut time-to-hire and surface more qualified candidates. Soraya was excited about the efficiency, but her DEI director, Aisha, asked one question that reframed the whole project: "When this tool ranks 6,000 people, who is it advancing, and can we prove the pattern is job-related?" That question is the heart of this lesson. AI does not have an opinion about inclusion. It executes whatever pattern you give it, consistently, at scale. The work is making sure that pattern advances inclusion rather than quietly undermining it.

DEI and AI in recruiting are deeply connected. AI can reduce unconscious bias and improve diversity, or it can amplify existing bias and worsen it, and the difference lies entirely in how the system is designed, how it is monitored, and how early the DEI team is brought in. Aisha is not another stakeholder for Soraya to manage. She is a partner in making sure the tool strengthens rather than weakens the network's diversity efforts.

What DEI Goals Look Like and Where AI Touches Them

Most organizations carry explicit DEI goals in recruiting: a target share of women in a technical function, a target share of underrepresented minorities moving into leadership, a target share of candidates with disabilities, increased recruiting from historically Black colleges and universities, or a commitment to candidates with non-traditional backgrounds. Northwind's version is a mix of all of these, tuned to clinical and non-clinical roles that draw from very different labor pools.

AI can help move any of these goals, but only when it is designed intentionally for that purpose. A tool selected purely on efficiency will not advance inclusion by accident, and the assumption that it might is one of the most persistent errors in this space. Soraya's practical test is whether she can name the specific mechanism by which a tool touches a specific goal, and whether that mechanism can be measured. A general expectation that a more objective process will produce better outcomes is an untested hypothesis, not a plan.

Where AI Genuinely Helps Inclusion

The strongest inclusion case for AI is not that it is unbiased. It is that it can make human evaluation more structured and more consistent, and inconsistency is where a great deal of bias lives. When five of Soraya's recruiters screen resumes by gut feel, the same candidate can be a strong yes for one and a pass for another, and the difference often tracks with how familiar a name, school, or career path feels. Human brains make automatic associations from exactly those signals. A well-designed AI step forces the same job-related criteria onto every applicant in the same order, removing much of the random variance that disadvantages people who do not fit a reviewer's mental template.

Consistency is neutral on its own, so what it produces depends entirely on the criteria. If the criteria are fair and job-relevant, consistency helps achieve diversity by ensuring every candidate is evaluated equally. If the criteria are biased, consistency delivers bias at scale. That conditional is why the rest of this lesson is about criteria, monitoring, and partnership rather than tool selection.

Blind screening is a second concrete mechanism. Removing names and schools from what the reviewer or the model sees reduces demographic bias directly, because it removes the signals that trigger the association. Structured assessment is a third: asking every candidate the same questions, in the same order, against the same rubric shrinks the unstructured judgment where bias thrives. Instead of "seems like a good fit," Soraya can require the tool to extract specific, role-relevant signals such as years of acute-care experience, named certifications, or demonstrated competencies.

Expanding reach is the fourth and in some ways the cleanest. AI can help Aisha and Soraya build Boolean searches optimized for diverse sources, run outreach to underrepresented talent communities, and do skill-based sourcing that finds candidates with relevant experience outside the traditional feeder companies. That surfaces candidates from community colleges, military transition programs, and employers Northwind had never recruited from. Expanding who gets considered improves diversity without touching how anyone is evaluated, which keeps it on firm legal ground.

Speed matters too, in a way that is easy to miss. A shorter cycle reduces the window in which a competitor can move first, and diverse candidates, women and candidates of color in particular, often hold multiple offers. A process that runs three weeks longer than a competitor's loses candidates it never rejected. Efficiency gains are not separate from the inclusion goal; they are one of the levers on it.

Where AI Quietly Undermines It

The same properties that make AI helpful make it dangerous when the inputs are wrong. The most common failure is training or tuning a tool on past hiring decisions. If Northwind historically hired a narrow profile, a model that learns from that history will reproduce it and present the result as objective. What you end up with is objective discrimination: biased patterns that feel more legitimate precisely because an algorithm produced them.

The subtler risk is proxy discrimination, where a criterion looks neutral while correlating tightly with a protected characteristic. Filtering for graduates of specific schools correlates with socioeconomic status, which correlates with race. Penalizing employment gaps disadvantages people who took caregiving leave, which skews by sex and disability status. Rewarding a continuous, linear career history screens out career changers and people returning to work. None of these criteria name a protected class, yet each can produce disparate impact, and the tool never has to discriminate explicitly for the outcome to be discriminatory.

A third risk is the transparency gap. Candidates rejected by an automated step frequently do not understand why, and an unexplained rejection feels unfair regardless of whether it was. That damages the employer brand generally and damages it most among candidates from underrepresented groups, who may already have well-founded concerns about bias in recruiting. The absence of an explanation is itself read as evidence.

The fourth risk is over-reliance. Human judgment is often what recognizes potential in a non-traditional candidate: unconventional experience, a non-linear career path, an underrepresented educational background. Removing it entirely can screen out exactly the candidates a thoughtful recruiter would have flagged as strong, the second-career nurse or the veteran medic. The goal is AI-assisted evaluation with a human in the loop, not full automation of the decision.

Reframing Culture Fit to Values and Behaviors Alignment

One phrase deserves special attention because it is where good intentions most often turn into bias: culture fit. When a hiring manager says a candidate is not a culture fit, they are frequently describing a feeling of dissimilarity, that the candidate did not share their humor, their background, or their communication style, rather than a job-related deficiency. Culture fit is one of the most reliable vectors for bias precisely because it sounds positive and is hard to challenge. If Soraya lets that language into her AI prompts or scoring rubrics, she encodes "people like us" into the system and amplifies it across every requisition.

The fix is to reframe culture fit as values and behaviors alignment, and to make it concrete. Instead of asking whether a candidate fits the team, Aisha and Soraya define the specific behaviors the role and organization require: gives direct feedback respectfully, collaborates across disciplines, raises patient-safety concerns even when it is uncomfortable. Those are observable, job-relevant behaviors that any candidate can demonstrate regardless of background. They also explicitly add culture add to the frame, asking what perspective or experience a candidate brings that the team currently lacks.

When you instruct an AI tool, the difference is stark. "Score for culture fit" invites the model to pattern-match on similarity, because similarity is the only thing that phrase can be operationalized as. "Assess evidence of these four behaviors, citing a specific example for each" forces it to evaluate job-relevant conduct against evidence in the application. The second prompt advances inclusion. The first quietly fights it, in language that sounds reasonable in every meeting where it comes up.

Measuring Impact: The Four-Fifths Rule Done Correctly

Soraya cannot manage what she does not measure, and the legal baseline for measuring selection fairness in the United States is the four-fifths rule from the EEOC's Uniform Guidelines on Employee Selection Procedures. The rule offers a rough screen for adverse impact: compare the selection rate of each group to the selection rate of the group with the highest rate. If any group's rate is less than four-fifths, that is 80 percent, of the highest group's rate, that is a flag warranting investigation. The rule is a trigger for scrutiny, not a verdict of illegality, and it does not establish discrimination by itself, but it is the standard first test.

Here is the math on Soraya's nursing pipeline. Of 500 male applicants, 200 advanced from screen to interview, a selection rate of 40 percent. Of 800 female applicants, 240 advanced, a rate of 30 percent. The highest rate is the men's 40 percent. The ratio for women is 30 divided by 40, which equals 0.75, or 75 percent. Because that is below the 80 percent threshold, this fails the four-fifths rule and is a clear flag for investigation.

Now run the same check by race. Of 700 White applicants, 280 advanced, a rate of 40 percent. Of 300 Black applicants, 96 advanced, a rate of 32 percent. The ratio is 32 divided by 40, which equals 0.80, exactly 80 percent, sitting right at the threshold and not flagging, though Soraya should still watch it because it is borderline. She also checks intersectionally, because a tool can pass for women overall and for Black applicants overall while still failing for Black women specifically. A single aggregate number can hide the disparity that matters most.

The monitoring questions that follow are straightforward to ask and easy to skip. Are women advancing through the AI screen at the same rate as men? Are candidates of color advancing at proportional rates? What disparities exist at each stage rather than only at offer? What does the picture look like when two characteristics are combined? Soraya runs these monthly rather than annually, because a year of data is a year of hires you cannot take back.

When a flag appears, the next question is diagnosis, not panic. A failed ratio can stem from the AI's criteria, from an applicant pool that was skewed before the AI ever saw it, or from a sourcing change made the same quarter. Soraya investigates root cause with Aisha before changing anything, because the right fix depends entirely on the cause.

Legal Guardrails and Bias Audits

The single most important legal point is also the most counterintuitive: you cannot fix a disparity by adjusting scores based on a protected characteristic. If Soraya discovered the four-fifths failure for women and fixed it by adding points to every woman's score to equalize advancement rates, she would be committing intentional discrimination based on sex. That is unlawful under Title VII even though the intent was to improve fairness, and it manufactures documented evidence of disparate treatment that is discoverable in litigation.

The legal levers all operate on the criteria and the pipeline, not on the people. Making criteria more strictly job-related, removing a proxy variable, expanding sourcing to reach underrepresented talent, and implementing blind screening that hides names and schools are all lawful. Quotas, score adjustments by demographic, and explicit preference for members of a protected group are not. Soraya runs any remediation past legal and DEI together, because the distinction between a lawful and unlawful remedy is frequently invisible to people thinking about fairness without also thinking about legality.

Jurisdiction matters as well. New York City's Local Law 144 requires employers using an automated employment decision tool for candidates in NYC to commission an independent bias audit within the prior year, publish a summary of the results, and notify candidates that the tool is being used. Because Northwind has clinics in several states, Soraya treats the NYC standard as a floor everywhere: an independent audit, published results, and candidate notice are good practice even where not yet legally required.

Where the network handles applicant data from candidates in the EU, GDPR adds rights around automated decision-making, including a candidate's right to obtain meaningful human review of a decision made solely by automated means. The through-line across all these regimes is the same: document your job-relatedness, audit for impact, and keep a human accountable for the decision.

Partnering with DEI Across the Whole Lifecycle

Aisha's early question mattered because DEI input is far more valuable before a tool is chosen than after it is deployed. DEI should be a partner in designing AI recruiting, not a gate at the end that reviews decisions already made. That partnership has five distinct moments, and skipping any of them costs something specific.

Before selecting a tool. DEI belongs at the table during evaluation, raising bias-risk concerns while the choice is still open. The questions that matter are concrete: does the vendor have disparate impact data from their other clients, have they tested the tool with diverse populations, what is their fairness approach, what transparency do they provide into how scores are produced, and do they support blind screening and structured assessment? Soraya brought Aisha into vendor evaluation for exactly this reason.

During validation. DEI helps set the validation criteria, which is where the genuinely strategic choices live. What counts as fair for this organization? What is the disparate impact tolerance? Should accuracy be weighted equally across groups, or should equal outcomes be prioritized? These are not technical questions with technical answers, and they should not be settled by whoever is configuring the tool.

During implementation. Any customization or configuration of the tool has diversity implications, so DEI weighs in when criteria are adjusted or weightings are changed. A weighting change made quietly during setup can undo everything the validation established.

During monitoring. DEI helps track diversity impact comprehensively rather than as a single headline number, covering advancement rates by group, disparities at each stage, and intersectional breakdowns. When problems are found. DEI helps investigate and respond quickly, starting with the root-cause question: is the disparity coming from the AI itself, from an underlying candidate pool that lacks diversity, or from something that changed upstream in sourcing?

What Shared Accountability Actually Looks Like

Partnership is not a set of approval gates; it is shared accountability for outcomes, and that shows up in four concrete practices. The first is defining success together. What does diverse hiring mean for this organization specifically? Proportional representation at each stage? A particular target share of women in a function? What about intersectionality, meaning whether women of color are represented at rates similar to white women? What about candidates with disabilities? Soraya and Aisha wrote these definitions down jointly rather than inheriting them from a slide.

The second is choosing tools together, using the vendor questions above. The third is setting monitoring thresholds in advance. What disparate impact ratio would concern you, 0.85, 0.80, 0.75? Different organizations hold different risk tolerance, and the point of fixing the number early is that you are not negotiating your risk appetite in the middle of a crisis.

The fourth is investigating together. When disparate impact surfaces, the failure mode is to investigate inside recruiting and then brief DEI on the findings. Pooling both kinds of expertise in the same investigation produces a better read on root cause, because the two disciplines notice different things in the same data. Soraya and Aisha also named a specific owner for monitoring diversity impact alongside efficiency metrics, and put representation by stage and demographic group on the same monthly dashboard as time-to-hire, so a divergence between the two is visible immediately rather than a year later.

Transparency and Candidate Trust

Candidates from underrepresented groups frequently arrive with a reasonable suspicion that recruiting is biased, so hiding AI use tends to confirm their fears while disclosing it can build trust. When candidates know an organization is taking fairness seriously, monitoring for bias, and working in partnership with its DEI team, they are more likely to engage even where AI is involved. When they suspect bias is being hidden or that nobody is checking, they disengage, and the candidates a diverse pipeline most depends on are often the first to go.

Soraya made Northwind's approach plain to applicants, covering four things: what tools are used and why, how fairness is monitored, what human review still happens, and how to appeal or request human review of a decision. Counterintuitively, telling candidates the organization audits itself makes them more willing to engage, not less. Transparency is not just an ethics line item. It is part of how Northwind keeps strong, diverse candidates in the pipeline through to offer.

Anti-Patterns

The DEI surprise. A company deploys an AI tool and later discovers it has disparate impact on a protected group. The DEI team was never involved and is, reasonably, upset. The company looks careless, and the DEI team loses trust in the recruiting function, which costs far more than the incident itself. It happens because teams move fast and treat DEI approval as nice-to-have rather than essential, or because they assume AI is inherently fair and bias-checking is therefore unnecessary. You end up implementing something in direct conflict with your own DEI goals and damaging the internal relationship you most need for the fix. Involve DEI before deployment, and get their input on bias risk and fairness criteria before you have chosen a tool and committed budget.

AI without accountability for diversity. A company implements an AI recruiting tool and is delighted because time-to-hire drops from 45 days to 25. They celebrate the efficiency win. Nobody monitors diversity impact. A year later they realize women's representation in hires has declined from 40 percent to 30 percent, representation of candidates of color has declined as well, and they cannot tell whether the cause is disparate impact in the AI, a change in sourcing strategy, or a combination.

It happens because nobody was accountable for monitoring diversity impact and the focus was entirely on efficiency metrics rather than fairness metrics. The damage is that you find the problem too late, after multiple cohorts have been hired from less diverse pools, and it is hard to recover from. The avoidance is to make someone accountable by name for monitoring diversity impact alongside efficiency, include diversity metrics in your quality audit, track representation at each stage by demographic group, and review monthly rather than annually.

Over-correction. A company discovers an AI tool has disparate impact on women, who advance from screen to interview at 25 percent while men advance at 40 percent. They implement a fix: weight women's scores higher to mechanically equalize advancement rates. This sounds fair and it is illegal. It is explicit preferential treatment based on a protected characteristic and a clear violation of EEOC regulations. It happens because the company is thinking about fairness without thinking about legality.

What goes wrong is that the fix creates a legal problem worse than the original one, because the company now has documented discrimination against men, and that documentation is discoverable in litigation. Work with legal and DEI together, and discuss any fix with legal before implementing it. There are lawful ways to improve diversity, including adjusting selection criteria to be more strictly job-related, expanding sourcing to reach underrepresented groups, and implementing blind screening. There are unlawful ways, including quotas, score adjustments by demographic, and explicit preference for protected groups.

Practice

Each of these produces something you could bring to a joint meeting with your DEI lead and your employment counsel, which is the right audience for all of it.

• Map your DEI goals to AI mechanisms. Write down your organization's DEI goals in recruiting. For each one, name specifically how AI could advance it and how AI could hinder it. Be concrete about the metric and the mechanism rather than the intention, and mark the ones where you cannot name a measurable mechanism.
• Design a validation process with diversity checks built in. Specify what you would need to know about a tool's disparate impact risk before deployment, and how you would test it against your own applicant data rather than the vendor's. Decide in advance what result would stop the deployment.
• Create a monitoring plan. Which metrics will you track: representation at each stage by demographic group, disparate impact ratios, intersectional breakdowns? At what frequency, weekly, monthly, or quarterly? And what specific result triggers action rather than discussion?
• Walk through a disparate impact response. If a tool showed disparate impact on a protected group, write out the steps you would take to understand root cause and decide on remediation, including who is in the room and what remedies are off the table for legal reasons.
• Interview your DEI leader. Ask what their biggest concerns about AI recruiting are, what support they need from you, and what their vision is for how AI could support diversity goals. Write down where their answers differ from what you assumed.

Reflection

• What is your organization's biggest DEI challenge in recruiting? Could AI help address it, and what would have to be true for that to work?
• What is your relationship with your DEI team like today? Are you involving them early enough in decisions, or bringing them finished ones?
• If you were deploying an AI tool tomorrow, what would your DEI team's concerns be? Have you actually asked them, or are you guessing?
• How would you explain to your CEO why DEI partnership in AI recruiting matters, covering both the fairness case and the business case?
• What would success look like: AI recruiting that advances your DEI goals? Define it specifically and measurably enough that someone else could check.

Glossary

• Disparate impact. A hiring practice that appears neutral but disproportionately affects protected groups. It does not require discriminatory intent to be legally significant.
• Protected characteristic. Demographic characteristics protected by employment law: race, color, religion, sex, national origin, age, disability, and in some jurisdictions sexual orientation and gender identity.
• Proxy discrimination. Using criteria that appear neutral but correlate with protected characteristics, such as school of graduation correlating with socioeconomic status and therefore with race.
• Four-fifths rule. The EEOC screen under the Uniform Guidelines: if any group's selection rate falls below 80 percent of the highest group's rate, that is a flag for investigation, not a finding of illegality.
• Automated employment decision tool. The category of tool regulated by NYC Local Law 144, which attaches independent bias audit, published results, and candidate notice obligations to its use.

Related Lessons

This lesson sits inside a cluster that supplies the measurement and legal machinery it depends on.

• Fairness Metrics: Defining and Measuring Bias in Outcomes goes deeper on the measurement discipline behind the four-fifths calculation, including how to handle small samples and intersectional cuts.
• Diversity and Bias in Sourcing: How AI Can Help and Harm develops the expanding-reach lever, which is the intervention with the cleanest legal position and often the largest effect.
• Fairness Interventions: Blind Reviews, Structured Processes, Diverse Panels covers the lawful remedies named here in practical detail, so a remediation plan has real mechanics behind it.
• Accessibility and Inclusion: Ensuring AI Processes Work for All handles the accommodation and accessibility obligations that sit alongside the demographic analysis in this lesson.
• Transparency and Disclosure: Telling Candidates About AI Use expands the candidate-facing disclosure that this lesson treats as a trust mechanism, including how much detail to give.
• Escalation Paths: When to Involve Legal, Compliance, DEI, or Leadership gives the escalation structure that makes the joint legal and DEI review described here happen reliably rather than by relationship.

Closing

AI recruiting can strengthen or weaken your DEI efforts, and the variable that decides which is partnership. When DEI teams are involved in tool selection, validation, monitoring, and the response to problems, AI tends to advance diversity. When they are bypassed, or consulted only after a problem has surfaced, AI tends to weaken it. That is not a claim about the technology. It is a claim about who is in the room when the criteria get chosen.

Concretely, the practice requires including DEI in procurement discussions, setting fairness and diversity metrics together, monitoring those metrics continuously, investigating disparate impact jointly, and acting on findings before diversity is damaged rather than after. Every one of those is a scheduling and ownership decision more than a technical one, which is what makes them achievable and also what makes them easy to let slide.

DEI partnership in AI recruiting is essential for both ethics and business. When you involve DEI teams early, take their concerns seriously, and work together to monitor and respond to problems, you end up with recruiting processes that are both fairer and more effective at attracting diverse talent. That is not a burden on recruiting. It is a competitive advantage, and it is the answer to Aisha's original question: you can say who the tool is advancing, and you can show that the pattern is job-related.

Key Takeaways

• There is no neutral AI; it either supports diversity or undermines it. The same consistency that strips out the random bias of inconsistent human review will faithfully reproduce biased criteria across every applicant. Your job is to make the pattern job-related, not to assume the tool is fair because it is automated.
• The real advantage is structured evaluation and expanded reach. AI helps most when it forces the same job-relevant criteria onto every candidate, supports blind screening and structured assessment, and sources beyond traditional pipelines. Expanding who gets considered improves diversity without altering how anyone is judged, which keeps it legally clean. Speed matters too, because strong candidates with multiple offers are lost to slow processes rather than rejected.
• Know the four ways AI worsens DEI. Amplifying historical bias when trained on past hiring data, proxy discrimination through neutral-looking criteria, an unexplained rejection that reads as hidden bias, and removing the human judgment that recognizes potential in non-traditional candidates.
• "Culture fit" is a bias vector; reframe it as values and behaviors alignment. Replace a feeling of similarity with specific, observable, job-relevant behaviors any candidate can demonstrate, and ask what a candidate adds rather than only whether they fit. Never let "culture fit" into an AI prompt or rubric.
• Use the four-fifths rule correctly: it is a trigger for investigation, not a verdict. Compare each group's selection rate to the highest group's rate; a ratio below 0.80 flags potential adverse impact. Check intersectionally, and diagnose root cause before changing anything.
• You cannot fix a disparity by adjusting scores by demographic. Demographic score adjustments, quotas, and explicit preference for protected groups are unlawful disparate treatment even when well-intentioned, and they create documented discrimination that is discoverable in litigation. Lawful remedies operate on criteria and pipeline: tighten job-relatedness, remove proxies, expand sourcing, add blind screening, and review any remediation with legal and DEI together.
• Know your jurisdictional guardrails and treat the strictest as a floor. NYC Local Law 144 requires an independent bias audit within the prior year, published results, and candidate notice for automated employment decision tools; GDPR grants a right to meaningful human review of solely automated decisions. Documenting job-relatedness, auditing for impact, and keeping a human accountable satisfy the common thread.
• Partner with DEI across the whole lifecycle, not at a final gate. Tool selection, validation criteria, implementation and configuration, ongoing monitoring, and joint investigation when problems appear. Define success and set the disparate impact threshold together in advance, name an owner for diversity monitoring, and put representation by stage on the same monthly dashboard as time-to-hire.
• Be transparent with candidates about AI use. Tell them what tools are used and why, how fairness is monitored, what human review remains, and how to request it. This builds trust, particularly with candidates from underrepresented groups who may already have concerns about bias in recruiting.

Frequently Asked Questions

Our four-fifths ratio came out at exactly 0.80. Are we fine? You do not flag at exactly 0.80, but treating that as a clean result misreads what the rule is for. It is a rough screen for adverse impact, not a certificate, and a ratio sitting precisely at the threshold is borderline and deserves continued attention. Soraya keeps watching that cut, checks it intersectionally in case a subgroup is failing inside an aggregate that passes, and looks at whether the ratio is trending toward or away from the line across batches.

We found disparate impact. Can we just weight the disadvantaged group's scores up to fix it? No. Adjusting scores based on a protected characteristic is explicit preferential treatment and unlawful disparate treatment under Title VII, even when the intent is to improve fairness, and it is a clear violation of EEOC regulations. It also creates documented evidence of discrimination that is discoverable in litigation, so the fix leaves you worse off than the original problem. The lawful paths all operate on criteria and pipeline: make the criteria more strictly job-related, remove the proxy variable driving the gap, expand sourcing to reach underrepresented talent, or implement blind screening. Discuss any remediation with legal and DEI together before you deploy it.

How do we tell whether the disparity is the tool's fault or the pipeline's? That is the root-cause question, and it has to be answered before you choose a remedy. A failed ratio can come from the AI's criteria, from an applicant pool already skewed before the tool saw it, or from a sourcing change in the same period. Investigate jointly with DEI rather than investigating alone and briefing them afterward. A remedy aimed at the wrong cause usually improves the measurement without improving the outcome.

What should we ask a vendor before we buy? Whether they have disparate impact data from their other clients and will share it, whether the tool has been tested with diverse populations, what their fairness approach actually is beyond a marketing claim, what transparency they provide into how scores are produced, and whether they support blind screening and structured assessment. Bring your DEI lead to that conversation rather than relaying answers afterward, and treat vendor fairness testing as an input to your own validation, not a substitute for it.

Is disclosing AI use to candidates risky? Practice runs the other way. Candidates from underrepresented groups often arrive already suspecting bias in recruiting, so silence about AI tends to confirm the suspicion rather than avoid the question. Disclosing what tools are used and why, how fairness is monitored, what human review remains, and how to request human review signals that someone is checking. The candidates who disengage are the ones who sense that bias is being hidden or that no one is monitoring.

Our time-to-hire improved a lot. Is that a DEI win? It can be, because a shorter cycle reduces the window in which a competitor takes a candidate holding multiple offers. But an efficiency gain proves nothing about fairness on its own, and celebrating it without fairness monitoring is exactly the second anti-pattern. Put representation at each stage by demographic group on the same monthly dashboard as time-to-hire, name someone accountable for watching it, and treat the two numbers as a pair.