Overview

Auditing AI decisions is how you catch bias, inaccuracy, and fairness problems early. You can't audit every decision -- with hundreds of screenings per month, that's impossible. But you can audit strategically using statistical sampling to detect problems in small samples and be confident your findings represent the whole population. This lesson teaches you the methods.

Audit Fundamentals

An audit is a systematic review of decisions to verify they're correct, fair, and compliant. In recruiting, audits typically answer three questions:

• Accuracy: Is the AI tool making correct decisions? (Are parsed resumes accurate? Are recommended candidates actually qualified?)
  - Fairness: Is the AI treating demographic groups equitably? (Is disparate impact present?)
  - Compliance: Can we defend these decisions if challenged? (Do we have documentation? Can we explain the reasoning?)

Why Sampling?

If you screen 500 candidates in a month, you can't manually review all 500 decisions. But if you randomly sample 30-50 of them and find no problems, you can be confident the whole batch is fine. That's the power of statistical sampling: you review a small sample and draw conclusions about the entire population with measurable confidence.

Key principle: Random sampling prevents selection bias. If you audit "every 10th candidate" or "candidates I'm worried about," you're not getting a representative picture. Random sampling (or stratified random sampling) is essential for valid audits.

Statistical Sampling Basics

Statistical sampling uses three concepts: population, sample, and confidence.

Population

The entire set of decisions you're auditing. Example: "All 500 candidates screened in March." The population is usually large; you can't review all of them.

Sample

A random subset of the population you'll actually review. You select this randomly to avoid bias. Example: "30 randomly selected candidates from the 500."

Confidence Level and Margin of Error

When you audit a sample, you're making an inference about the population. You're saying, "In my sample, 90% of decisions were accurate, so probably ~90% of all decisions are accurate." But how confident are you? Confidence level quantifies this.

• 95% confidence level (standard in recruiting audits): If you repeated this audit 100 times with different random samples, ~95 of them would give you results within your margin of error.
  - Margin of error: The range around your sample result. If your sample shows 90% accuracy with 5% margin of error, the true population accuracy is likely between 85-95%.

Example

You audit 40 randomly selected screened-in candidates. You verify each one against job requirements. You find 36 are qualified (90% accuracy). Your margin of error at 95% confidence is 5%. This means: "I'm 95% confident the true accuracy rate for all screened-in candidates is between 85-95%."

If your target is 95% accuracy, this sample result raises a concern: The true accuracy might be as low as 85%, which misses your target. You'd investigate further.

Sample Size Determination

Larger samples give you tighter confidence intervals (more precision). Smaller samples are faster to audit but give you wider ranges (less precision). Here's a practical guide:

Population Size
Sample Size (95% confidence, 5% margin)
Audit Time (min)
Use Case

50-99
40
40-60
Quick audit of small cohort

100-499
45
45-75
Weekly or bi-weekly audit

500-999
50
50-80
Monthly audit

1,000+
60
60-90
Quarterly audit of large population

Practical formula: For 95% confidence with 5% margin of error, use this rough estimate: Sample size 400 / (population size in hundreds). For a population of 500, that's 400 / 5 = 80, but 50-60 usually suffices. (This is simplified; statisticians use more precise formulas, but this gets you close.)

Recruiting tip: For weekly audits, use n=30-40. For monthly, use n=50-70. For quarterly, use n=100+. Larger samples give you more confidence but take longer. Choose based on the risk and the time you have.

Stratified Sampling for Fairness Audits

Sometimes simple random sampling isn't enough. For fairness audits, you want to ensure you're sampling across demographic groups, not just randomly. That's where stratified sampling comes in.

What is Stratified Sampling?

You divide the population into groups (strata) -- typically by demographic category or outcome -- and then randomly sample from each group. This ensures you have enough data from each group to compare fairly.

Example: Fairness Audit with Stratified Sampling

You screened 300 candidates in March. You want to audit for fairness across gender. Your population breaks down as:

• 100 men (screened in: 20)
  - 150 women (screened in: 22)
  - 50 non-binary/prefer not to say (screened in: 8)

For a stratified sample of 50, you'd sample proportionally:

• Men: 17 candidates (10 screened in, 7 screened out)
  - Women: 25 candidates (12 screened in, 13 screened out)
  - Non-binary/PNTS: 8 candidates (4 screened in, 4 screened out)

Now you manually verify fairness: Do the screened-in candidates from each gender group actually meet the job criteria? Are there biases? Is one group being filtered unfairly?

When to Use Stratified Sampling

• Fairness audits: Always stratify by demographic group so you have enough data per group
  - Accuracy audits for specific roles: If you have different screened-in rates by role, stratify by role
  - Detecting intersectional bias: Sample across combinations (e.g., Black women, Asian men) to check if some combinations are treated unfairly

Confidence Intervals and Action Thresholds

A confidence interval is the range you're confident the true value falls within. Understanding these helps you decide when to act.

Example: Disparate Impact Confidence Interval

You audit 50 candidates: 20 women, 30 men. 10 women screened in (50%), 12 men screened in (40%). Pass rate women / pass rate men = 50% / 40% = 1.25 disparate impact ratio.

But with small samples, there's uncertainty. Your 95% confidence interval might be 0.90-1.60 (meaning: I'm 95% sure the true ratio is somewhere in this range). Since the range includes 0.80 (the legal threshold), you can't be confident there's actually disparate impact. You'd need more data.

With a larger sample (200 candidates), same pass rates, your confidence interval might be 1.10-1.40. Now you're 95% confident there's disparate impact. You'd investigate and potentially pause the tool.

Decision Framework

• Confidence interval entirely above/below threshold: Strong signal. Act confidently.
  - Confidence interval crosses threshold: Weak signal. Collect more data before acting.
  - Multiple audits showing consistent pattern: Even if individual intervals cross threshold, a pattern is a signal. Investigate.

Legal note: The 80% rule (disparate impact ratio 0.80) is an EEOC guideline, not a hard legal threshold. Courts look at multiple factors: statistical significance (your confidence interval), business necessity, alternative practices, impact size, and more. Document your analysis carefully.

Disparate Impact Analysis in Depth

Disparate impact analysis is how you detect whether a neutral-seeming process is affecting groups differently.

The Four-Fifths Rule

The EEOC's 80% rule: If the selection rate for one group is less than 80% of the selection rate for the group with the highest rate, disparate impact is indicated.

Example

Group
Applicants
Screened In
Pass Rate
Disparate Impact Ratio

Men
100
25
25%
1.00 (baseline)

Women
100
18
18%
0.72 (below 0.80)

Black
50
10
20%
0.80 (acceptable)

Latinx
50
9
18%
0.72 (below 0.80)

Interpretation: Women and Latinx candidates are being screened at lower rates than men. Disparate impact is indicated. You must investigate: Is this intentional? Unintentional bias? Job-related? If you can't justify it as job-related, you're at legal risk.

Next Steps After Detecting Disparate Impact

• Investigate root cause: Why is this happening? Analyze the screening criteria, training data, or workflow design.
  - Test hypotheses: Is it a particular criterion? Does it happen for all roles or specific roles? Is it in the AI tool or the human part of the process?
  - Document findings: Keep detailed records of your investigation. Courts want to see evidence that you took it seriously.
  - Develop remediation: Remove or adjust the problematic criterion, retrain, pilot the fix.
  - Re-audit: After changes, re-audit to confirm disparate impact is resolved.
  - Communicate: To the DEI team, legal, and affected candidates if appropriate.

Audit Protocols by Frequency

Weekly Audit (n=30-40, ~60 min)

When: Every Monday morning

What: Randomly sample candidates screened the prior week. Verify accuracy of screening.

Questions:

• Did the AI extract resume info correctly?
  - Did it screen in candidates who actually meet criteria?
  - Did it screen out candidates who do meet criteria? (false negatives)

Output: Accuracy percentage. If
- What's the pass rate for each demographic group?
- Are disparate impact ratios 0.80?
- Is there a pattern of systematically screening out one group?

Output: Disparate impact analysis by group. If any group
- Overall accuracy trend: Is it stable, improving, or degrading?
- Fairness across all axes: Demographics, role, recruiter
- Quality of hire: Are screened-in candidates performing well?
- Candidate experience: Any feedback about fairness, transparency?

Output: Comprehensive audit report. Present to leadership with recommendations for next quarter.

Annual Comprehensive Audit (External Auditor, ~2000 min)

When: Once per year, ideally before public reporting or regulatory audit

What: Full-year data, internal and external audit, compliance assessment

Output: Audit report defensible in court, suitable for EEOC response

Documentation and Evidence Preservation

Every audit should be documented. This documentation is critical evidence if you're ever audited by regulators or challenged by candidates.

What to Document

• Audit date, auditor name, tool audited
  - Population definition: "All candidates screened for Software Engineer role, Jan 2026"
  - Sample method: "30 randomly selected from 150 screened-in candidates, stratified by gender"
  - Findings: "28 of 30 clearly met criteria; 2 were borderline"
  - Disparate impact analysis by group (if fairness audit)
  - Root cause analysis (if problems found): "Why did those 2 not clearly meet criteria?"
  - Actions taken: "Retrained tool on X criterion; flagged recruiter for coaching on Y"
  - Confidence level and margin of error

Sample Audit Report Template

>
WEEKLY ACCURACY AUDIT -- March 11, 2026

Tool: Resume Screening for SWE Roles

Auditor: Sarah Johnson, Recruiting Ops

Population: 142 candidates screened week of March 4-8

Sample: 35 randomly selected (25% of population)

Confidence: 95%, Margin of Error: 8%

Findings:

Accuracy (screened-in candidates meet criteria): 31/35 = 89% [CI: 81%-97%]

Accuracy (screened-out candidates don't meet criteria): 33/35 = 94%

Issues Found:

2 candidates screened in but didn't list required technical skill (though had it via prior role). Tool missed context clues. Recruiter caught both and moved them forward appropriately.

Action: Retrain tool on context clue recognition. Next week's audit will verify fix.

Pro tip: Keep a shared spreadsheet or database of all audits. Include links to full reports. This makes it easy to spot trends over time (e.g., "accuracy has trended down 2% per month") and shows regulators you're serious about monitoring.

Key Takeaway

Statistical auditing is how you verify AI quality at scale. Random sampling (or stratified sampling for fairness) lets you audit small samples with measurable confidence about the whole population. Weekly accuracy audits catch problems fast. Monthly fairness audits detect bias. Quarterly strategic audits assess overall health. Documenting every audit creates evidence you're being responsible and creates a defensible record if you're ever challenged.

Frequently Asked Questions

What if I don't have enough demographic data to do stratified sampling?

This is a real problem. Start collecting demographic data now (optional, voluntary, transparent). In the meantime, use simple random sampling for accuracy audits, and do proxy analysis (analyze by school, location, job history, which correlate imperfectly with demographics) for fairness checks. But prioritize collecting actual demographic data--it's the only way to audit fairly.

If my confidence interval crosses the 80% threshold, should I pause the tool?

Not automatically. One borderline month doesn't require a pause. But it's a signal to watch closely. Increase your audit frequency (weekly instead of monthly), increase sample size, or do a targeted investigation of the candidates just above/below the threshold. If the next month also shows a borderline or worse result, then consider pausing and investigating more deeply.

How do we audit AI tools from vendors we don't have direct access to?

You audit the outcomes, not the tool itself. You don't need to see the tool's code or weights. You screen candidates, then audit whether they were treated fairly and whether the screening was accurate. If the vendor won't give you access to verify outcomes, that's a red flag. Make transparency a requirement in your contracts.

What's the difference between accuracy and disparate impact?

Accuracy is "are the decisions correct?" Disparate impact is "are groups treated similarly?" A tool can be 95% accurate overall but have disparate impact (e.g., 95% accuracy for men, 70% for women, because it's calibrated for the dominant group). Audit both separately. High accuracy + no disparate impact = good. High accuracy + disparate impact = you have a fairness problem.

Do we need a statistician to do this auditing?

Not necessarily for weekly and monthly audits. Recruiting ops people can handle those with the frameworks in this lesson. For annual audits or when you detect complex patterns, bring in a statistician or external auditor. They'll have more sophisticated tools and can withstand legal scrutiny.