Appendix — AI in the Assessment Workflow

This is an appendix rather than a chapter for a reason worth stating. The boundary it operationalizes was drawn long before these pages, philosophically in the research corpus this book stands on and doctrinally throughout the book itself, and it fits in one sentence that readers of the firm's work will recognize: AI maps the field; consultants read the candidate. Fluency-class work (transcription, retrieval, summarization, mapping, comparison, drafting) is what these systems genuinely do, and this appendix welcomes it into the workflow under discipline. Judgment-class authority (scoring persons, ranking finalists, deciding) is never delegated, for reasons the book has now earned twenty-five chapters of standing to assert. What remains is the operational layer: exactly which uses sit where, what the law now requires, and what any client should demand of any firm.

One warning before the material, in the book's own tradition: these are its fastest-aging pages. The tools change quarterly and the law is mid-construction, and several obligations cited below take effect within weeks of this book's completion. The principles here are stable; the citations carry dates; verify both against the calendar in your hand.

The strongest case for the machines, stated first

The book's honesty pattern applies to its own boundary, so the pro-algorithm evidence comes first and at full strength. Chapter 23 already delivered the deepest result: mechanical combination of validated ratings beats expert heads by half, and that finding is algorithmic reasoning vindicated, in the instrument's chair. Add the field evidence: in a widely cited field experiment, an algorithmic screen selected candidates who outperformed those chosen by human screeners, and selected less traditional candidates while doing it, because the machine lacked the recruiters' taste for familiar pedigrees; algorithmic consistency is real (the model does not get tired at 4 p.m., does not glow at the last candidate, does not drift with mood); and for the drudgery layers of search (coverage, first-pass structuring, document synthesis) the productivity gains are genuine and this firm banks them daily.

Read carefully, though, what every one of those results is: consistency in executing validated criteria, inside the instrument role. The screen that beat human screeners was combining explicit signals against a defined outcome; the mechanical-combination result concerns aggregating evidence humans collected and validated. Nothing in the success record shows a system judging a person (reading identity, weighing a world, owning a claim), and the failure record shows, with expensive consistency, what happens when systems are promoted from the first role to the second.

The audit record, compactly

The canon, because clients cite it and candidates live it. Amazon's experimental recruiting engine, trained on its own hiring history, taught itself to penalize the word "women's" and was scrapped: proxy discrimination distilled from the past and automated forward. HireVue retired its facial-analysis component after sustained scientific and regulatory pressure: trait inference from faces never had the construct validity its marketing implied. The academic audit of algorithmic hiring vendors found bias-mitigation claims largely unverifiable from the outside: opaque validation as an industry norm. And the newest layer repeats the oldest pattern. The 2023–2026 audit studies of LLM-based résumé screening found demographic effects from names and proxies, along with sensitivities no assessment instrument should have: output shifting with prompt phrasing, with candidate order, with a rerun of the same input. The recurring failure is one failure: fluency mistaken for evidence, and correlation with the past mistaken for judgment about a person, the exact confusions Chapters 2 and 4 spent themselves teaching human assessors to resist, now available at scale.

The LLM workflow, and the assisted candidate

Two present-tense realities shape the disciplines below. First, generative tools are already inside every firm's workflow (summarizing interviews, drafting documents, structuring references), and the young evidence base says their outputs vary across runs, flatter the framing of the prompt, occasionally manufacture attributes, and inherit the biases of their training distributions: usable, under controls; never self-certifying. Second, the candidate is assisted too. The experimental finding is elegant and consequential: candidates who believe AI is evaluating them change how they present, more analytic, less warm, optimizing for the imagined machine, so the deployment of AI assessment changes the construct being measured. And candidate use of generative tools for applications and asynchronous interviews is rising fast enough that the design conclusion is already firm: assume assistance. Which quietly re-weights this book's own hierarchy: verification layers matter more, live encounter and enacted exercises gain evidentiary value, and unproctored asynchronous formats lose it.

The usage matrix

Permitted — assistance, under the disciplines below: transcription and meeting notes; market mapping and long-list research; document retrieval and summarization with source-return; translation assistance; scheduling and logistics; drafting support for documents whose human author reviews, revises, and owns them; structured comparison tables of verified facts.

Conditional — allowed only under named controls: LLM summarization of interviews and reference calls (source-return verification of every claim; extraction kept separate from evaluation; the tool never generates evaluative language about the person); CV parsing and knockout screening on objective, validated, job-related criteria (monitored for adverse impact, with a human-reachable appeal path); psychometric platforms (Chapter 22's five rules, in full); any decision-support scoring (the four oversight conditions below, plus the disclosure the law now requires); candidate-facing chatbots for process logistics (disclosed as automated, with immediate human escalation).

Prohibited — in this firm's practice, and in this book's counsel to any firm: AI scoring or ranking of candidates as the operative judgment; video, voice, or facial analysis for traits or emotions; personality inference from digital footprints or scraped social data; automated rejection without a human decision meeting the oversight conditions; generating the recommendation's reasons by machine — the rationale must not merely avoid restating the model, it must not be the model wearing the assessor's voice; and any undisclosed AI assessment of a person. The last item is absolute: a candidate assessed by a machine they were never told about has been wronged regardless of the machine's accuracy.

The eight disciplines

For everything in the permitted and conditional rows: (1) Fixed rubrics first — any tool touching candidate material operates against criteria written before use, never criteria emerging from the tool. (2) Extraction separated from evaluation — machines may pull facts, quotes, and timelines; conclusions about persons are drafted by humans from the extracted record. (3) Source-return of every claim — each machine-produced statement is traceable to, and checked against, the recording, the document, the witness; a summary sentence that cannot be returned to its source does not enter the file. (Readers of the corpus will recognize the practice by its proper name — locus-reinjection — and everyone else may simply recognize it as hygiene.) (4) One candidate at a time — no batch comparisons by LLM; ordering effects are documented, and comparison is Chapter 23's job under Chapter 16's rules. (5) Frozen configuration per campaign — model version and prompts locked for a search, so all candidates meet the same instrument. (6) Subgroup monitoring where volumes allow, and vendor audit rights where they don't. (7) Logs — what tool, on what material, at what step, per search: the record that makes rows one through six checkable. (8) The human rationale that does not restate the model — Chapter 24's memo, which exists on its own merits and now also happens to be what the regulators mean.

The regulatory floor, dated

As this book is completed, mid-2026, jurisdiction by jurisdiction — a floor, not advice; counsel per corridor.

European Union. The AI Act classifies AI systems for recruitment and selection — advertising, screening, filtering, evaluating candidates — as high-risk (Annex III), with the main obligations applying from 2 August 2026: risk management (Art. 9), data and data-governance quality (Art. 10), transparency to deployers (Art. 13), effective human oversight (Art. 14), deployer duties including use per instructions, monitoring, and human oversight assignment (Art. 26), and the affected person's right to an explanation of the role AI played in a significant decision (Art. 86). And the provision every search firm should read twice: Article 25 — a deployer that puts its name or trademark on a high-risk system, substantially modifies it, or repurposes a general tool into high-risk use becomes the provider, inheriting the full compliance stack. The firm that markets "our proprietary AI assessment," built on someone else's model, has very likely just volunteered for provider obligations it has not budgeted.

GDPR, and the judgment that closed the loophole. Article 22 restricts decisions based solely on automated processing with significant effects — and the Court of Justice's SCHUFA ruling established that a score which is decisive in practice for a downstream decision is itself the automated decision: routing the score through a vendor, or through a human who ratifies it, does not launder the responsibility. Profiling transparency, data-protection impact assessments, and minimization apply throughout.

United States. Title VII disparate-impact analysis applies to selection tools per EEOC guidance, four-fifths rule included, with employer responsibility extending to vendor tools. New York City's Local Law 144 requires annual independent bias audits and candidate notice for automated employment decision tools — with employment agencies in scope. Illinois runs two statutes: AIVIA on AI analysis of video interviews (consent, explanation, deletion), and HB 3773 — in force since 1 January 2026 — bringing AI discrimination in employment decisions under the Human Rights Act with notice duties. Maryland requires consent for facial recognition in interviews. The state patchwork is growing; multistate practice should assume the strictest applicable rule.

United Kingdom. No single AI statute; the floor is data-protection law (ICO guidance on AI and automated decision-making) plus equality law — functionally similar duties by another route.

Gulf and Central Asia. The data-protection regimes (Saudi PDPL, UAE frameworks, and their regional counterparts) are young, consent-oriented, and thin on AI-specific case law: the conservative posture — disclosure, consent, human decision, documentation — is both the compliant and the honorable stance while the law matures.

The practical compliance posture, for a firm working across all of the above: one standard, set at the highest common denominator, applied everywhere. Per-jurisdiction toggling of ethics is expensive, error-prone, and — the deeper point — incoherent for a firm whose product is judgment: the disclosure a Brussels candidate is owed by law, a Riyadh candidate is owed by the same professional standard that made the law look reasonable.

Meaningful oversight: the four conditions

Every conditional row above, and most of the law above, converges on one test — what makes a human in the loop a decider rather than a ratifier. Four conditions, all required: the human sees the evidence, not merely the score — the underlying material, at decision grain; understands the limits — trained on what this tool measures, misses, and gets wrong, and for whom; has authority and time to disagree — organizationally real, calendar real, and evidenced by a nonzero disagreement rate, because a reviewer who has never overruled the tool is a signature, not an oversight; and records reasons — the documented rationale, per discipline eight. Readers will notice these are not new to this book: they are Chapter 23's override log and Chapter 24's memo, discovered independently by the regulators — the rare and pleasing case where the compliance artifact and the craft's best practice are the same document.

Where this firm draws the line

One page, first person, offered as the standard to demand of anyone, us included. We use AI daily: for market mapping and research coverage, for transcription and note synthesis, for retrieval and drafting assistance, always under the eight disciplines, always logged. We do not and will not: let a machine score or rank a candidate as the operative judgment; analyze anyone's face, voice, or video for traits; infer personality from digital footprints; reject anyone by automation; or assess anyone with AI without disclosure. Every recommendation this firm issues is authored and signed by a human being who conducted or supervised the encounters, per Chapter 24; every significant tool use is disclosed on request; the override log and the calibration log exist and are run. And the portable version, for any reader engaging any firm, is three questions: What do your tools decide, and what do they merely prepare? Who signs the recommendation, and what did they personally see? Show me the disclosure you give candidates. Firms with good answers will not resent the questions. Firms that resent the questions have answered them.

Where the rules run out

The honesty paragraph, one last time. The legal material above was accurate when written and is aging as you read: treat it as a map's edition date, and confirm against current text and counsel before relying on any line of it. The LLM evidence base is young: reliability, bias, and mitigation findings are moving with the systems themselves, and a discipline defensible this year may be insufficient or superseded next. The matrix is doctrine applied to a moving field; the principles beneath it are the stable part, and they are not this appendix's to prove: the two registers, the instrument/judgment boundary, the answerable locus behind every claim about a person. For those, the corpus; for the law, counsel; for the practice, the twenty-five chapters behind you; and for the tools, the posture this book has taken toward every instrument it examined: use what earns its place, log what you use, and never let anything sign in your stead.

Notes and sources

Evidence grades: [M] meta-analytic/systematic; [L] peer-reviewed primary; [S] regulatory/practitioner, dated; [T] flagged doctrine.

• Algorithmic screening field evidence: Cowgill, "Bias and Productivity in Humans and Machines" (working-paper series and successors). [L]
• Mechanical combination: Chapter 23's sources (Kuncel et al., 2013; Grove et al., 2000). [M]
• Vendor audit: Raghavan, Barocas, Kleinberg & Levy, "Mitigating Bias in Algorithmic Hiring" (FAT* 2020). [L]
• Amazon case; HireVue facial-analysis retirement: contemporaneous investigative and regulatory records. [S]
• LLM résumé-screening audits (2023–2026): the emerging audit literature on name/proxy effects, prompt and order sensitivity. [L]
• The AI assessment effect: Goergen, de Bellis & Klesse (PNAS, 2025). [L]
• Candidate GenAI use in assessment: Robie et al. (IJSA, 2026). [L]
• EU AI Act: Regulation (EU) 2024/1689 — Annex III(4); Arts. 9, 10, 13, 14, 25, 26, 86; application dates per Art. 113. [S]
• GDPR Art. 22; CJEU SCHUFA (C-634/21, 2023). [S]
• US: EEOC Title VII guidance on selection procedures and AI (2023); NYC Local Law 144 and DCWP rules; Illinois AIVIA and HB 3773 (eff. 1 Jan 2026); Maryland HB 1202. [S]
• UK ICO guidance on AI and data protection. [S]
• Gulf/Central Asia data-protection regimes: Saudi PDPL and regional counterparts — thin AI-specific law, flagged. [S]
• The matrix, the eight disciplines, the four conditions as operationalized, the firm's line: this book's synthesis from the research and regulatory record. [T]