Overview

An AI tool recommends rejecting a candidate. Your hiring manager hires them anyway, and a year later, they sue for discrimination. The legal discovery process demands: "Show us every system, every score, every decision point in your hiring tool." You search your files and find spreadsheets, some chat logs, and a vague vendor manual. Your legal team goes pale.

This lesson covers what happens when documentation fails, and how to build an audit trail that survives scrutiny.

The Regulatory Landscape

Federal regulators don't yet have AI-specific employment rules, but they're writing them now. The EEOC has announced enforcement action on AI discrimination. The FTC is testing companies. The SEC is requiring disclosure of AI use in compensation decisions for executive compensation. State laws, California, New York, Illinois, are moving faster than federal regulators.

What they all want: proof that your AI system doesn't discriminate, and proof that your organization used it responsibly.

Courts have already set the bar. In *Obermeyer et al. v. University of Pennsylvania* (2019), researchers found that a widely-used algorithm systematized racial bias. When discrimination cases reach trial, judges now expect companies to produce:

• The algorithm's training data - what fed it? were protected classes represented?
  - Validation studies - did you test for bias across demographic groups?
  - How decisions were made - what weight did the AI recommendation carry?
  - Who reviewed it - did a human actually evaluate or rubber-stamp?
  - The outcome - what actually happened to the candidate or employee?

Without this documentation, you're arguing blind. With it, you have a defense, if the documentation shows responsible use.

What "Documentation" Actually Means in HR

Documentation isn't a single document. It's a system of records that tracks AI involvement from decision point through outcome. Think of it as three layers:

Layer 1: System Documentation
- What the tool does, its vendor-provided specs, your customizations
- Training data sources and size
- Validation results (bias testing, false positive rates, demographic parity analysis)
- Update history and versioning

Layer 2: Decision Documentation
- Input data (resume, performance metrics, salary history)
- System output (score, recommendation, explanation if the tool provides it)
- Human review notes (did the reviewer challenge the AI? on what grounds?)
- Final decision and justification
- Timestamp and decision-maker identity

Layer 3: Outcome Documentation
- What happened after the decision (hired/rejected, promoted/passed over, terminated/retained)
- Performance data if relevant (did the hired candidate perform well? was the rejected candidate later successful?)
- Complaint data (was there a complaint? on what grounds?)

Most HR teams document Layer 1 poorly and skip Layers 2 and 3 entirely. That's where you get exposed.

Building Your Documentation Standard

Start by asking: What would I need to defend this decision in court?

A candidate was rejected. They claim age discrimination. Your AI tool scored them 2.3/10. You need to show:

• The tool was validated for age bias (validation study showing no disparate impact across age groups)
  - The candidate's actual qualifications vs. the AI inputs (did the tool see what was actually in the resume?)
  - The threshold score (why was 2.3 below cut-off? was the cut-off set arbitrarily?)
  - Human review (did the hiring manager read the resume or just trust the AI?)
  - The alternative candidate who was hired (were they objectively stronger, across comparable metrics?)
  - Outcome data (how did the hired candidate perform vs. the rejected candidate's known later performance?)

Without this, the other side argues: "You used an unexplained black box that rated an age-protected applicant below a younger applicant." You lose.

Documentation Standard Template

DECISION DOCUMENTATION RECORD
======================================

DECISION CONTEXT
Tool Used: [Name, version, date installed]
Decision Type: [Hiring / Promotion / Termination / Compensation / Performance]
Date of Decision: [YYYY-MM-DD]
Decision-Maker: [Name, title]
Reviewer (if separate): [Name, title]

INPUT DATA
Data Sources: [List all inputs fed to AI system]
Protected Class Data Visible to Tool: [Yes/No - specify which: age, gender, race, etc.]
Data Quality: [Any missing fields? any corrections made?]
Last Updated: [When data was current]

SYSTEM OUTPUT
AI Score: [Numeric or categorical]
AI Recommendation: [Specific text]
Confidence Score (if available): [Percentage or range]
Explanation Generated by Tool: [If tool provides one, include verbatim]
Alternative Candidates Scored: [How many? what were their scores?]

HUMAN REVIEW
Reviewer Name: [Title, department]
Review Date: [Date]
Review Method: [Full file review / Score review only / [Other]]
Reviewer Comments: [Did reviewer accept, challenge, or override? Why?]
Override Decision (if any): [What did reviewer change? Why?]
Documented Concerns: [Any red flags noted about tool output?]

FINAL DECISION
Decision: [Hired / Rejected / Promoted / Other]
Justification: [What factors led to this decision? Acknowledgment of AI role vs. human judgment]
Alternative Considered: [Why wasn't the next-scored candidate chosen?]

OUTCOME TRACKING
Hired/Promoted Candidate Performance: [6-month, 12-month data if available]
Rejected/Not Promoted Candidate: [Known outcome - hired elsewhere? still seeking role?]
Complaint Filed: [Yes/No - if yes, grounds?]
Follow-up Actions: [Any systemic changes made based on this decision?]

Don't treat this as a box-checking exercise. A sloppy review note ("Looks good") is worse than no note. It suggests you didn't actually think. A strong review note ("Candidate flagged with high GPA score but hiring manager noted limited industry experience in required area; score downweighted accordingly") shows judgment.

AI-Specific Documentation Failures and How to Avoid Them

Failure Mode 1: The Disappearing Algorithm
Your vendor goes out of business or stops supporting the tool. You can't produce the validation study. You can't explain how it worked. Solution: Require vendors to provide full technical documentation as part of the contract. Store it. Require them to commit to a retention period for back-end access.

Failure Mode 2: The Unexplained Score
The tool says "Candidate score: 6.8/10" with no explanation. What drove it? You don't know. The candidate's attorney asks, and you can't answer. Solution: Require explainability. If a tool can't explain its output, don't use it. If it can, store the explanation with every score.

Failure Mode 3: The Rubber-Stamp Reviewer
Your hiring managers run candidates through the tool and hire the top 3 scores without reading anything else. Documentation shows "Reviewed by: [Manager name]" but no actual review. Solution: Make review mandatory and traceable. Require a narrative note. Audit review quality. Reviewers who consistently accept all recommendations should be asked why.

Failure Mode 4: The Lost Outcome
You documented the hiring decision but never tracked whether the hire succeeded. Three years later, they sue for discrimination, claiming the AI hired them as a diversity check but the company created a hostile environment. You have no data on performance, promotion, retention. Solution: Build outcome tracking into your HR system. Link original decision records to subsequent performance reviews, promotion decisions, termination reasons.

Failure Mode 5: The Undisclosed Tool
You used an AI tool in hiring, but your hiring manager didn't tell candidates. Candidates discover it later (maybe on a glassdoor.com review of your company's hiring process). They feel deceived and organize a complaint. Solution: Document disclosure practices. Who told candidates about AI involvement? When? Can you prove it?

Workflow: Integration into Daily HR Processes

Documentation can't be bolted on afterward. It has to be baked into the workflow.

Hiring Workflow with AI Documentation

STAGE 1: APPLICATION INTAKE
├─ Candidate submits resume
├─ System automatically flags protected class data visible in resume
│ (Age via graduation dates, Gender via pronouns, etc.)
├─ Documentation logged: Source data captured
└─ Flag: [If protected class data visible, document this]

STAGE 2: AI SCREENING
├─ Resume fed into AI tool
├─ System generates score + explanation (if available)
├─ AUTOMATED DOCUMENTATION:
│ ├─ Timestamp
│ ├─ Inputs sent to tool
│ ├─ Output score
│ └─ Explanation text
├─ System generates list of top-scored candidates
└─ Bottom-scored candidates flagged for human review (don't auto-reject)

STAGE 3: HUMAN REVIEW
├─ Hiring manager receives:
│ ├─ Candidate resume
│ ├─ AI score
│ ├─ AI explanation
│ └─ [Optional: Alternative candidates for comparison]
├─ Manager must document:
│ ├─ Did manager read full resume? [Y/N]
│ ├─ Manager assessment: [Text field - what stands out?]
│ ├─ Agreement with AI score? [Y/N] - If no, explain
│ └─ Move to interview? [Y/N]
├─ System prevents submission without narrative comments
└─ AUTOMATED DOCUMENTATION: Captures review entry

STAGE 4: INTERVIEW & OFFER
├─ Interview notes captured
├─ Reference check results captured
├─ Offer decision documented with justification
├─ DOCUMENTATION: Final hiring decision linked to initial AI score

STAGE 5: ONBOARDING & OUTCOME TRACKING
├─ New hire performance tracked at 30, 90, 180 days
├─ Performance review linked to hiring record
├─ DOCUMENTATION: Outcome data appended to original decision record

This workflow requires technology, an ATS that captures structured data, not just narrative comments. If your ATS can't do this, work with your vendor or implement a simple logging spreadsheet.

Sensitive Edge Cases: Documentation Under Pressure

Scenario 1: Overriding the AI to Meet Diversity Goals

Hiring manager wants to hire Candidate B (score 5.8) instead of Candidate A (score 8.2) to increase engineering team diversity. Legal risk: Candidate A could claim reverse discrimination.

Documentation approach:
- Record the override explicitly: "Manager overrode AI score to advance diversity strategy"
- Justify it: "Candidate B has foundational skills sufficient for the role (score above threshold of 5.0) and brings underrepresented perspective to engineering team"
- Make it defensible: Store your company's stated diversity rationale. If you have a legitimate business case for diversity (i.e., "we've had turnover in underrepresented groups and external research shows inclusive teams perform better"), that supports the override
- Track the outcome: Does Candidate B succeed in the role? This validates the decision

Scenario 2: Terminating an Employee Based Partly on AI Performance Data

An employee's AI-assisted performance score has declined. You're considering termination. Legal risk: Employee claims the AI was biased and you relied on biased data.

Documentation approach:
- Separate the AI output from your evaluation: "AI-generated performance score of 3.2/5.0 noted. In manager review, specific incidents observed: [list concrete behaviors/outcomes]"
- Show the human judgment: "AI score provided input, but primary factors in termination decision were [list manager-observed issues]"
- Prove you didn't auto-fire: Document conversations with the employee, improvement opportunities offered, coaching provided
- Question the data: "Was the AI model trained on historical data that might reflect past biases?" If yes, document that concern explicitly

Scenario 3: Using AI in Compensation Decisions

Some organizations are experimenting with AI to suggest salary adjustments. High risk: Women or minorities could claim the AI perpetuated pay gaps.

Documentation approach:
- Validate the tool against pay equity: Run analysis of what salaries the tool recommends across demographic groups. Document that analysis
- Require manual review: No AI-suggested salary should be used without a manager and HR review
- Document the gap analysis: "Tool recommended $X salary. Internal equity check noted that women in similar roles earn $Y on average. Salary adjusted to $Z based on [reason]"
- Assume litigation: Every compensation decision with AI input will be scrutinized. Document as if you're explaining it to a plaintiff's attorney

Templates You Can Use Now

Decision Log Template (Spreadsheet or ATS Field)

Date
Employee/Candidate
Decision Type
AI Tool
AI Recommendation
AI Score
Human Override?
Final Decision
Reviewer
Comments
30-Day Outcome

2026-04-01
Jane Smith
Hiring
TalentAI v2.3
Proceed to interview
7.8/10
No
Interviewed; offer extended
M. Johnson
Alignment with job req; strong technical score
Accepted offer

Review Narrative Template (Required Field)

HUMAN REVIEW NOTES
═══════════════════════════════════════════

Candidate: [Name]
Date Reviewed: [Date]
Tool Score: [Score]

REVIEW ASSESSMENT:
[Reviewer summarizes what they observed in the candidate file,
independently of the AI score. This forces human judgment.]

COMPARISON TO AI OUTPUT:
[Reviewer explicitly states: Do I agree with the score? Why or why not?
This creates the audit trail of human judgment.]

DECISION AND RATIONALE:
[What will happen to this candidate and why?
How much weight did the AI score have vs. other factors?]

CONCERNS NOTED:
[Any flags about the candidate, the process, or the tool output?]

Red Flags That Will Sink You in Litigation

• No validation study for the AI tool - You can't prove it's fair
  - Validation study shows bias - You used the tool anyway without adjusting
  - No human review documented - Looks like you automated the decision entirely
  - Review notes are one-liners - "Looks good" suggests no real analysis
  - Different documentation standards across hiring managers - Suggests inconsistent process
  - No outcome data - You didn't track whether the AI's recommendations actually predicted success
  - Undisclosed tool use - You hid that AI was involved
  - Tool vendor can't produce their validation study - Major red flag
  - Algorithm changes without re-validation - You updated the tool and didn't retest for bias
  - Demographic gaps in outcomes - You hired a much higher percentage of one group; can't explain it

What to Do Monday Morning

• Audit your current process: Pull 10 recent hiring decisions. Can you reconstruct the AI involvement from your records? If not, you have a problem.
  - Get your ATS vendor on the phone: Ask them what AI-related data they capture and store automatically. Ask about retention policies.
  - Create a decision log template: Start recording every AI-involved decision going forward, even if your ATS doesn't do it automatically.
  - Brief your legal team: Show them what you're planning to document. Get their sign-off. (They may want additional fields based on industry or litigation history.)
  - Implement mandatory review notes: Tell hiring managers they must document their thinking, not just accept/reject.
  - Run a bias audit on any AI tool you're currently using: If you can't get validation data, stop using the tool.

Key Takeaways

• Document three layers: System specs, decision inputs/outputs/reviews, and outcomes
  - Make human review traceable: Narrative comments required; one-liners don't count
  - Assume you'll defend it in court: If you wouldn't write it that way for litigation, rewrite it
  - Track outcomes: The best defense is data showing the AI recommendations actually worked
  - Validate for bias before deployment: Post-hoc validation doesn't save you if the study shows problems
  - Store everything: Vendor docs, validation studies, decision records, outcome data. Keep it for 7+ years

FAQ

Q: What if our AI tool vendor won't provide validation data?
A: That's disqualifying. A vendor who can't or won't prove their tool is fair is one litigation trigger away from costing you millions. Find a new vendor.

Q: Do we need to document every single hiring decision, or just the AI-involved ones?
A: Document every decision, but flag which ones involved AI. This prevents the argument that you used AI for adverse decisions but not for favorable ones.

Q: Our hiring managers are resistant to documentation burden. What's the pitch?
A: "Documentation protects us in lawsuits. It's now standard in sophisticated companies. And it makes you look like you thought carefully about hiring decisions, not like you trusted a black box."

Q: If an employee sues and demands discovery, how long do we have to produce this documentation?
A: Depends on the court and lawsuit timeline. Assume 30 days to produce. That means your documentation system has to be organized enough to pull a coherent file within a week.

Q: Does documenting AI use increase our legal liability?
A: Counterintuitively, no. Poor documentation is worse. If you used AI without documenting it and someone sues, they'll argue you hid it and ask for sanctions. Transparent documentation is a defense.

Q: What if the AI tool made a recommendation we disagreed with and we made a different decision? Do we document that?
A: Absolutely. Override documentation is valuable. It shows you didn't blindly trust the tool.

Q: Should we tell candidates that AI was involved in their rejection?
A: This depends on your jurisdiction and company policy. Document your disclosure practice regardless.

What's Next

In the next lesson, we'll cover the trust side: How to tell employees and candidates that AI is involved in HR decisions without triggering suspicion or pushback. Documentation proves you did it right. Communication explains why you did it. Together, they build defensibility and employee confidence.

The best documentation is transparent documentation. Secrecy kills trust. The companies winning the AI-in-HR game are the ones being direct: "Here's the tool we use. Here's how it works. Here's how humans review it. Here's what we monitor for bias."

That transparency is what we'll address next.