RESHUFFLE An interactive companion to the book
Reshuffle concept · deep dive

The Human-in-the-Loop Fallacy.

The Human-in-the-Loop Fallacy is the belief that the continued presence of humans in an AI-mediated workflow proves that humans remain central to value creation. It mistakes existence for position.

A human may remain in the loop. The question is where that human sits in relation to the algorithmic system. Does the human define the problem, own the customer, shape the workflow, and absorb accountability? Or does the human review, correct, validate, and patch the machine? The first position is above the algorithm. The second is below it.

See the position read → Place yourself on the matrix →
The framing

What does everyone get wrong about human-in-the-loop?

The phrase has institutional weight. AI safety guidance treats meaningful human oversight as a horizontal requirement. Regulation writes it into law for high-risk systems. CEO talking points and consulting decks borrow the same shape and apply it to labor markets directly.

Everyone agrees: humans stay in the loop, so jobs are safe.

Almost nobody asks: a loop terminating where?

The structural read Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design. Criticality gapDurability modifierTuareg trap (criticality-vs-compensation gap)Named after the Tuareg navigators of the Sahara - indispensable to caravan crossings but historically paid at commodity rates. A scarce capability is hired as commodity labour when the buyer refuses to acknowledge its criticality. The work looks defensible (the task can't be done without the worker) but compensation prices it as throughput.Where it shows up:Per-stamp architects-of-record; native-fluency translation post-editors paid per word; high-volume radiologists doing AI-overread at piecework rates; regulated-copy reviewers paid project-by-project to sign off on AI drafts. Substance testDurability modifierSurface-vs-substance validatorValidation work captures premium only when the validator tests substance against ground truth. It collapses when the validator can only check for surface coherence - because AI is structurally better than humans at producing surface coherence. The seat depends on whether ground-truth tests exist for the domain.Where it shows up:Citation-audit paralegals testing against actual case law (substance check); AI-output reviewers at content platforms judging coherence only (surface check); radiologists doing AI-overread against the image (substance check); compliance copy review without a statute-grounded checklist (surface check).
  1. Everyone says

    “AI augments. Humans stay involved.”

    NIST AI RMF (2023) and EU AI Act Article 14 (2024) write meaningful human oversight into law. McKinsey State of AI 2024 names augmentation as the dominant enterprise frame.

    But the loop terminates

    Involvement is not ownership. The translation post-editor is involved with the AI assistant. The Klarna gig CS agent at $41/hr was involved with the OpenAI assistant. Both are below the algorithm. Augmentation says nothing about where the output goes.

  2. Everyone says

    “Humans handle what AI can't.”

    Salesforce Agentforce launch (Sep 2024). CEO talking points across the Magnificent Seven. The consulting-deck residual-theory of work.

    But the loop terminates

    Starting with 'what AI can't' produces compensating-layer jobs - validation, exception handling, relationship management, ethics review. The map of those jobs is mostly below the algorithm.

  3. Everyone says

    “Humans validate AI output before it ships.”

    Anthropic RLHF era discourse, AI safety guidance, compliance-review job listings.

    But the loop terminates

    Dell'Acqua-Mollick 2023: GPT-4 users produced more polished output on every BCG task - including the tasks where the model was systematically wrong. Surface validation collapses as buyers learn this. Most validators do surface.

The structural cut

Where does the output go after the human touches it?

Same human. Same workflow. Two destinations. The destination is the position.

Human in the loop Client solving a new problem Above the algorithm Bureau Works founder · Aha! · senior architect on Cursor · M&A partner on Harvey AI pipeline someone else owns Below the algorithm Translation post-editor · Klarna gig CS · RLHF labeller · Workday recruiter

The presence of the human tells you nothing. The destination tells you everything. Everyone gets the question wrong.

The reframe

Where does the human actually sit?

Everyone gets augmented. Position decides outcome.

The premise of this map is that everyone in knowledge work is now augmented by AI. The real question is what AI does to your seat. Is AI inside the offering you sell - the thing the customer pays for? Or is AI inside the workflow you operate - the tool you use to do work that someone else owns? And does the position route the output to a client, or back into someone else's pipeline?

Craft

Trial litigator Trial litigator Courtroom is non-delegable. The bar associations have not allowed AI to argue cases. The litigator uses AI for research and prep, but the offering is the human advocacy seat. Surgeon Surgeon Procedure is the deliverable. The surgeon uses AI imaging and decision-support, but the OR is human-bound. AI has not entered the operating room as the actor. Sworn / court translator Sworn / court translator Court, asylum, and immigration translation require sworn signoff. Substance liability binds to the practitioner. AI tools speed prep but the signature is the offering. A1 authority binding firing. CPA (audit signoff) CPA (audit signoff) Big Four enterprise audit requires the CPA seal under PCAOB rules. AI accelerates the production work but the seal is what the client pays for. The seat is licensed-human-bound. M&A senior partner M&A senior partner Deal-thesis judgment is the offering. Capital-side counterparties pay for upstream inquiry, not memo production. Junior associates use AI; the partner's judgment is the sold thing.

AI as leverage

Vertical AI founder Vertical AI founder Owns the customer's new problem AND the AI stack inside the offering. The customer pays for an AI-augmented service the founder built. Bureau Works (Fairman), Aha! (de Haaff), Midjourney (Holz), Harrison.ai (Tran). Senior architect + Cursor Senior architect + Cursor Ships a software product that has AI inside it. Cursor/Copilot/Anysphere is an accelerator, but the product the customer pays for is the architect's design. Boilerplate compresses below them while they direct the system. M&A partner + Harvey M&A partner + Harvey Sells AI-augmented advisory: the deal thesis is the offering, Harvey accelerates production. A&O Shearman profit-share with Harvey signed Apr 2025 - the buyer reorganized comp around the AI-augmented offering. Frontier-lab eval engineer Frontier-lab eval engineer Authors the rubric the next AI model is graded against. The offering is the inquiry-framing - what the AI is for. Customers pay for the eval, not the model. Anthropic / OpenAI / frontier labs. Practitioner-vendor Practitioner-vendor Rebundled their own AI-augmented offering under their own brand. Owns the customer relationship and at minimum rents the AI stack inside their product. Christian Corah / PNC AI is canonical.

Commoditized labor

Junior paralegal + Harvey Junior paralegal + Harvey Uses Harvey for first-pass discovery, contract review, and citation work under a partner's signature. The firm owns the case, the partner captures the rent. The junior tier compresses as Harvey absorbs the substrate the partner used to pay them to produce. Junior SWE + Copilot Junior SWE + Copilot Accepts Copilot/Cursor/Cognition (Devin) diffs in someone else's repo. The senior architect owns the system; the junior is below the algorithm because their tier is the AI-compressible tier. Brynjolfsson-Chow-Hoffmann 2025 anchor. Workday-operating recruiter Workday-operating recruiter Operates the Workday AI scoring system that decides which candidates the firm sees. The firm owns the hiring outcome; the recruiter is the operator inside the scoring product. Mobley v Workday surfaced 1.1B rejected applications. Junior data analyst + copilot Junior data analyst + copilot Pulls and pivots data using analyst copilots. The firm owns the analysis; the junior analyst's tier is fungible. ChatGPT and Cursor compress the pull-and-pivot tier. Per-stamp AoR reviewer Per-stamp AoR reviewer Reviews AI-generated AEC documents for architect-of-record sign-off at per-stamp rates. The firm owns the project; the reviewer's seat is throughput-priced criticality. The Tuareg trap firing. Tier-1 CS + Agentforce Tier-1 CS + Agentforce Uses Salesforce Agentforce or Sierra/Decagon to deflect Tier-1 tickets. The firm owns the customer relationship; the agent's tier is what the AI deflects first. 70% containment baseline now achievable.

AI as discipline· The HitL trap

Translation post-editor Translation post-editor Post-edits machine-translation output at per-word piecework ($0.03-$0.05). The vendor MT (RWS, TransPerfect, Smartling, DeepL) is the offering; the post-editor's corrections feed the next model's training memory. RLHF / data labeler RLHF / data labeler Labels examples that train the AI offering. Output is the AI pipeline's input, with no paying client on the other side. The vendor (OpenAI, Anthropic, Scale, Appen) owns the offering; the labeler at $1-$2/hr is the residual. Gig CS overflow agent Gig CS overflow agent Handles the residual cases the vendor's AI cannot resolve. Klarna's Uber-style pilot ran this at $41/hr through May 2025 reversal. The OpenAI assistant is the offering Klarna's customers experience. AI-output reviewer paralegal AI-output reviewer paralegal Reviews AI-generated documents at content platforms where the AI output IS the product the platform ships. Surface coherence test, no substance check. The platform's AI is the offering. Content classifier reviewer Content classifier reviewer Reviews flagged content the surface classifier wasn't sure about. The vendor's classifier IS the moderation offering; the reviewer is the residual exception path. Ghost-work pattern Gray-Suri documented since 2019.

Hover any role chip for why it lands there. Click a quadrant for the practitioners and the durability test.

Above the algorithm · AI in the workflow

Craft

Authority binderWhere value re-anchorsRisk concentration / authority bindingComponents bearing institutional risk - signature, seal, regulatory accountability, professional liability - become more valuable, not less, because they absorb the new failure modes automation introduces. Insurance, regulators, and standards bodies price the residual risk into licensure language.Where it applies:Radiologists, surgeons, trial litigators, M&A partners, UHNW advisors, sworn translators, CPAs.Criticality gapDurability modifierTuareg trap (criticality-vs-compensation gap)Named after the Tuareg navigators of the Sahara - indispensable to caravan crossings but historically paid at commodity rates. A scarce capability is hired as commodity labour when the buyer refuses to acknowledge its criticality. The work looks defensible (the task can't be done without the worker) but compensation prices it as throughput.Where it shows up:Per-stamp architects-of-record; native-fluency translation post-editors paid per word; high-volume radiologists doing AI-overread at piecework rates; regulated-copy reviewers paid project-by-project to sign off on AI drafts.

AI is in your daily workflow as a tool, but the offering you sell is your human act, signature, or institutional seat. The customer pays for the act (the surgery, the courtroom appearance, the sworn attestation, the CPA seal) - AI just helps you prep. The institutional bind keeps you above the algorithm.

The durability test

Holds because the regulator refuses to let AI hold the seat. The FDA's AI/ML device approval pathway still routes the final read through a radiologist. The PCAOB still binds audit signoff to a CPA. Courts still bind immigration translation to a sworn human. The seat erodes the day the regulator publishes a framework allowing AI to sign directly - watch the FDA AI/ML pathway, FAA autonomous-flight certification, and bar association rulings on AI legal practice.

Practitioners in this quadrant today
  • Trial litigators Courtroom is non-delegable; AI accountability docket adds adversarial work
  • Surgeons Procedure is the deliverable; AI has not entered the OR as actor
  • Sworn / court translators Court, asylum, immigration translation requiring sworn signoff; substance liability binds to the practitioner
  • CPAs (audit signoff) Big Four 'Client Zero' enterprise sell-through requires CPA seal; AI accelerates production but the seal binds
  • M&A senior partners Deal-thesis judgment; capital-side counterparties pay for upstream inquiry, not memo production
Above the algorithm · AI in the offering

AI as leverage

Customer ownershipWhere value re-anchorsCustomer-problem captureThe practitioner moves from selling component-tasks to owning the customer's new problem - an outcome the buyer can't get elsewhere, anchored to a constraint the practitioner uniquely holds. AI is ingredient inside the offering; the practitioner owns the customer relationship.Where it applies:Senior-tier roles in every exposed knowledge profession - senior bankers, consultants, M&A partners, UHNW advisors, agency owners, brand-anchored designers.Question framingWhere value re-anchorsInquiry-framing premiumWhen AI answers well-posed questions at parity, rent moves upstream to whoever poses them. The worker captures premium by owning the question-frame the customer didn't know they needed - not by owning the customer but by holding problem-definition authority.Where it applies:Product managers, management consultants, data analysts deciding what to measure, market researchers, M&A bankers framing deal theses, strategy partners.Capital tierDurability modifierSovereign-vs-rented capitalPractitioners who own a customer relationship split into four tiers based on what capital they hold: those who own the AI stack inside their offering (most durable), those renting it from a third-party provider, those who own the distribution channel but not the AI, and those who own neither. The capital tier determines how durable the position is to provider price or access changes.Where it shows up:Wherever a practitioner owns a customer relationship - e.g. Bureau Works (Gabriel Fairman) owns the AI stack inside the offering; many freelance translators rent the AI stack from a third-party provider.

AI is inside the offering you sell. You own the customer relationship, the question the AI is for, or the capital stack the AI runs on. The AI multiplies what you sell instead of replacing you.

The durability test

Test: who decides what your AI economics look like next quarter? Bureau Works owns its own MT layer, so its margin survives an OpenAI price change. Founders building thin wrappers on GPT-4 had to renegotiate their P&L in 60 days when OpenAI launched GPT-4 Turbo Nov 2023. Sovereign capital is the difference between catching the expansion and reading about it in a vendor's pricing email.

Practitioners in this quadrant today
  • Gabriel Fairman Bureau Works - translator-founded platform with AI inside the offering
  • Brian de Haaff Aha! - bootstrapped product-management platform, AI inside the offering, no VC
  • Senior software architects + Copilot/Cursor System owners shipping faster while boilerplate compresses below them
  • M&A partners + Harvey Deal-thesis framing premium; AI inside the workflow they own. A&O Shearman profit-share with Harvey signed Apr 2025
  • David Holz Midjourney - ~107 employees, no VC, ~$500M ARR; AI inside the offering they own
  • Aengus Tran Harrison.ai / Annalise.ai - radiologist-founded; 35% NHS chest-X-ray share
Below the algorithm · AI in the workflow

Commoditized labor

Cost collapseTriggerCost decline at task levelInference cost falls such that the marginal cost of producing AI output approaches zero. The capability question is whether AI can do a task; the cost question is whether buyers can afford to use AI in place of human labour. Both must shift for adoption to crystallise.Where it applies:Roles with high-volume / low-margin work where the per-task AI cost has fallen below the per-task human wage.SubstitutionHow the work breaksDirect substitutionA task is done by AI at parity; the worker is no longer paid for that component. Pure absorption - no unbundling, no relocation.Where it applies:Discrete tasks where parity is clean - first-draft writing, basic image generation, routine cite-checking, simple translation, telemarketing scripts.Cascade displacementHow the work breaksCascade displacementWhen AI absorbs the work of one role, displaced workers flee to adjacent roles that haven't been hit yet - depressing wages in those adjacent roles even though the work there hasn't changed.Where it applies:Translators → copywriters → journalists. Paralegals → claims adjusters → administrative assistants. Junior bankers → strategy consultants.

AI is in your daily workflow as a tool, but the firm above you owns the offering. You're a fungible operator inside a traditional pyramid where the AI compresses your tier. Your output goes to a paying client - through your firm's hierarchy - but your role is increasingly substitutable.

The durability test

The cost curve has crossed the wage line for your tier. Two paths from here: absorption (your seat goes away as AI gets better) or relocation as cleanup (you migrate into the discipline quadrant on the left). Brynjolfsson-Chow-Hoffmann 2025: developers aged 22-25 lost ~20% employment share post-ChatGPT. The senior tier above you captures the differential.

Practitioners in this quadrant today
  • Tier-1 customer support (with deflection AI) Post-Agentforce/Sierra/Decagon era CS; the AI handles 70%+ containment, the human handles the rest
  • Junior software engineers (with Copilot) Routine code production accepting Copilot/Cursor/Cognition diffs in someone else's repo; cascade displacement firing
  • Junior paralegals (with Harvey) Routine discovery and contract review under partner signature; Harvey/Spellbook absorb the substrate
  • Junior data analysts (with copilots) Routine pull-and-pivot work that analyst copilots absorb
  • Workday-operating recruiters Operating the AI scoring system that decides which candidates the firm sees; Mobley v Workday discovery surfaced 1.1B rejected applications
  • Per-stamp architect-of-record reviewers Reviewing AI-generated documents for sign-off at per-stamp rates; criticality high, compensation throughput
Below the algorithm · AI in the offering

AI as discipline· The HitL trap

Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design.Criticality gapDurability modifierTuareg trap (criticality-vs-compensation gap)Named after the Tuareg navigators of the Sahara - indispensable to caravan crossings but historically paid at commodity rates. A scarce capability is hired as commodity labour when the buyer refuses to acknowledge its criticality. The work looks defensible (the task can't be done without the worker) but compensation prices it as throughput.Where it shows up:Per-stamp architects-of-record; native-fluency translation post-editors paid per word; high-volume radiologists doing AI-overread at piecework rates; regulated-copy reviewers paid project-by-project to sign off on AI drafts.Substance testDurability modifierSurface-vs-substance validatorValidation work captures premium only when the validator tests substance against ground truth. It collapses when the validator can only check for surface coherence - because AI is structurally better than humans at producing surface coherence. The seat depends on whether ground-truth tests exist for the domain.Where it shows up:Citation-audit paralegals testing against actual case law (substance check); AI-output reviewers at content platforms judging coherence only (surface check); radiologists doing AI-overread against the image (substance check); compliance copy review without a statute-grounded checklist (surface check).

AI is the offering - someone else's vendor product is what the customer pays for. You operate inside that vendor's AI as residual labor. The output you produce feeds back into the vendor's pipeline rather than going to a paying client. Augmented, throughput-priced, the canonical HitL trap.

The durability test

Two failure modes firing. First: criticality real, compensation throughput - Translation post-editors at $0.03-$0.05 per word despite native fluency, Klarna gig CS at $41/hr handling AI residual. Second: if the work is grading AI output for surface coherence, it dies as buyers learn Dell'Acqua-Mollick 2023: GPT-4 produced more polished output even when systematically wrong.

Practitioners in this quadrant today
  • Translation post-editors Machine-translation post-edit at per-word piecework rates; translation memory feeds the next model
  • Data annotators / RLHF labelers Labeling examples that train the AI they are paid to correct; canonical ghost-work configuration
  • Klarna gig customer-service agents Pre-reversal: ~$41/hr cleanup for the AI assistant handling two-thirds of chats. May 2025 walk-back reabsorbed the function in-house
  • AI-output reviewer paralegals Reviewing AI-drafted documents at content platforms; surface coherence test, no substance check
  • Content moderation reviewers Reviewing classifier-flagged content; ghost-work pattern documented by Gray-Suri
The closed loop

What if your human-in-the-loop work trains the AI that absorbs it?

The matrix above shows snapshot position. The dynamic underneath is more damning.

Every human-in-the-loop worker generates training data through the act of being in the loop. Pick a job below to see the loop play out.

Mechanism Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design. Substance testDurability modifierSurface-vs-substance validatorValidation work captures premium only when the validator tests substance against ground truth. It collapses when the validator can only check for surface coherence - because AI is structurally better than humans at producing surface coherence. The seat depends on whether ground-truth tests exist for the domain.Where it shows up:Citation-audit paralegals testing against actual case law (substance check); AI-output reviewers at content platforms judging coherence only (surface check); radiologists doing AI-overread against the image (substance check); compliance copy review without a statute-grounded checklist (surface check).
THE LOOP continuous
01
You correct the AI
Edit machine-translation output to acceptable client quality. Native fluency in source + target; sector judgment on tone and terminology.
02
Corrections log as TM
Every accepted edit becomes a translation-memory pair. Stored in the LSP's TM stack (Smartcat, Phrase, RWS Trados).
03
Next model trains on TM
Vendor MT models retrain on the corrected pairs. DeepL, RWS, TransPerfect MT improve on the exact pairs you fixed.
04
Per-word rate falls
The model now produces fewer errors on the segments you used to cleanup. Volume drops. Per-word rate drops. Seat erodes.
THE LOOP continuous
01
You rank model outputs
Label two AI responses for preference, helpfulness, harmlessness. Appen / Scale / CrowdGen contract workflow.
02
Preferences become rewards
Your rankings turn into reward signals for RLHF. The model learns to optimize for the preference distribution you authored.
03
Next model passes more cases
Fewer outputs need labeling because the model now ranks closer to the preference. Required label volume per model shrinks.
04
Annotator headcount shrinks
Vendor contracts compress. Per-task rate stays low ($1-$2/hr in Kenya / Philippines). Per-model labor allocation falls per generation.
THE LOOP continuous
01
You flag AI drafts
Review Harvey / Spellbook / Robin AI output for surface coherence, tone, completeness. Approve or flag.
02
Flags become training data
Acceptance and rejection signals feed the vendor's RLHF and supervised-fine-tuning cycles.
03
Surface coherence improves
Dell'Acqua-Mollick 2023: the model already produces more polished output than humans check. Your flags accelerate this.
04
Surface seat collapses
Substance-validation seats (citation-audit) hold; surface review is structurally what AI does better. Surface review headcount cuts first.
THE LOOP continuous
01
You audit AI citations
Test every AI-cited case against actual Westlaw / Lexis records. Post-Mata v Avianca (June 2023) the seat appreciated as a substance-validation premium.
02
Corrections feed RAG
Vendor records which citations failed substance check. Retrieval-augmented systems retrain on the verified-good set.
03
Hallucination rate falls
Next-gen Harvey / Spellbook citation-grounding gets stricter. Substance violations become rarer and easier to flag automatically.
04
Substance seat narrows
Eventually substance audit becomes a checklist the model runs itself. The substance premium holds longer than surface, but the loop is the same.
THE LOOP continuous
01
You author the rubric
Design the eval the next model is graded against. MMLU-Pro, ARC-AGI, SWE-bench, your own bespoke benchmark for the lab.
02
Model trains against rubric
Lab uses the rubric for RLHF reward shaping and instruction tuning. The model optimizes against your scoring function.
03
Model passes
Within 6-18 months of a benchmark's release, frontier models saturate it. Your rubric stops discriminating between top models.
04
Author the next harder one
The seat survives only as long as you can author benchmarks that out-pace the model. Each cycle compresses.
THE LOOP continuous
01
You handle the residual
Take the third of chats the OpenAI assistant could not resolve. $41/hr 1099, no benefits, Klarna Uber-style pilot Feb 2024 - May 2025.
02
Transcripts become training
Every resolved chat became supervised-learning data for the next assistant update. Vendor (OpenAI) owns the corpus.
03
Assistant containment grows
Two-thirds containment moved toward three-quarters as the assistant learned from your transcripts. Residual shrinks.
04
Gig pool reabsorbed
May 9 2025: Klarna reverses, brings CS back in-house at $35/hr W-2. The gig pool that trained the assistant is no longer needed at the same scale.

The HITL fallacy assumes the human is the safety perimeter around the AI. The loop is the opposite. The human is the training signal for the AI - and the better the signal, the smaller the perimeter the next model needs.

The pattern, role by role

Is human-in-the-loop one thing happening to one role?

No. Seven configurations firing across thirteen roles. Most cells fire below the algorithm. The fallacy isn't isolated. It's a pattern.

The seven configurations Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design. Criticality gapDurability modifierTuareg trap (criticality-vs-compensation gap)Named after the Tuareg navigators of the Sahara - indispensable to caravan crossings but historically paid at commodity rates. A scarce capability is hired as commodity labour when the buyer refuses to acknowledge its criticality. The work looks defensible (the task can't be done without the worker) but compensation prices it as throughput.Where it shows up:Per-stamp architects-of-record; native-fluency translation post-editors paid per word; high-volume radiologists doing AI-overread at piecework rates; regulated-copy reviewers paid project-by-project to sign off on AI drafts. Substance testDurability modifierSurface-vs-substance validatorValidation work captures premium only when the validator tests substance against ground truth. It collapses when the validator can only check for surface coherence - because AI is structurally better than humans at producing surface coherence. The seat depends on whether ground-truth tests exist for the domain.Where it shows up:Citation-audit paralegals testing against actual case law (substance check); AI-output reviewers at content platforms judging coherence only (surface check); radiologists doing AI-overread against the image (substance check); compliance copy review without a statute-grounded checklist (surface check).
54
cells fire across the grid (of 91)
78%
of those fire below the algorithm
9
cells hold above; almost all are substance, eval, or failover under regulatory bind
Cleanup Surface Substance Training Eval Edge case Fail-over
Translators below Translators · Cleanup Translation post-editing; the canonical cleanup HITL below Translators · Surface review Compliance copy review at LSP QA tier above Translators · Substance validation Sworn / court translation under signature liability below Translators · Training data TM corrections feed vendor MT below Translators · Edge case Low-resource language pairs
Paralegals below Paralegals · Cleanup First-pass doc review below Paralegals · Surface review AI-output review at content platforms above Paralegals · Substance validation Citation-audit specialists post-Mata v Avianca below Paralegals · Training data Edits feed Harvey, Spellbook, EvenUp retrain cycles below Paralegals · Edge case Junior tier handling escalations Harvey cannot resolve
Customer-support agents below Customer-support agents · Training data Chat transcripts feed assistant retrain below Customer-support agents · Edge case Klarna gig CS overflow at $41/hr; canonical below Customer-support agents · Fail-over Salesforce HITL pool inside Service Cloud
Copywriters below Copywriters · Cleanup AI-draft post-edit at content mills below Copywriters · Surface review Brand-voice + tone check at agencies below Copywriters · Training data Corrections feed agency in-house models below Copywriters · Edge case Compliance-copy + regulated-domain copy
Performance marketers below Performance marketers · Cleanup Andromeda + Advantage+ creative cleanup below Performance marketers · Surface review Brand-safety + creative review below Performance marketers · Training data Conversion signals feed platform models contested Performance marketers · Eval criterion Cross-platform incrementality measurement below Performance marketers · Edge case Anomaly review on auto-bid swings
Radiologists above Radiologists · Substance validation FDA AI/ML pathway overread; image against impression contested Radiologists · Training data Annotated overreads feed Aidoc / GE Imaging models above Radiologists · Edge case Uncertain cases the model declines above Radiologists · Fail-over FDA mandates human-in-the-loop fallback
Junior software engineers below Junior software engineers · Cleanup Copilot / Cursor diff review below Junior software engineers · Surface review PR review of AI-generated code contested Junior software engineers · Substance validation Senior architect review of AI architecture choices below Junior software engineers · Training data Accept/reject signals feed Copilot/Cursor models below Junior software engineers · Edge case Debugging the cases AI cannot trace
Junior bankers below Junior bankers · Cleanup Deck polish on AI-drafted decks below Junior bankers · Surface review Memo polish on AI-drafted memos below Junior bankers · Edge case Manual research the AI cannot source
Journalists below Journalists · Cleanup Wire-copy edit on AI-drafted news below Journalists · Surface review Style check on AI summaries above Journalists · Substance validation Investigative fact-checking against ground truth below Journalists · Training data Edits feed editorial AI training below Journalists · Edge case Beat-specific knowledge AI cannot replicate
Claims adjusters below Claims adjusters · Surface review Routine claim review on AI-triaged claims above Claims adjusters · Substance validation Substance validity check + fraud review below Claims adjusters · Training data Decisions feed insurer underwriting models below Claims adjusters · Edge case Escalations AI cannot adjudicate above Claims adjusters · Fail-over Regulatory bind on decision review
Content moderators below Content moderators · Surface review Classifier-flagged content review (Sama, Telus) below Content moderators · Training data Labels feed surface classifier retrain below Content moderators · Edge case Worst-of-worst edge content the model declines
Financial analysts below Financial analysts · Cleanup Model polish on AI-drafted analysis below Financial analysts · Surface review Memo polish on AI-generated reports above Financial analysts · Substance validation Audit-prep substance validation below Financial analysts · Edge case Manual research AI cannot source
Recruiters below Recruiters · Surface review Resume review on AI-scored candidates below Recruiters · Training data Hire/no-hire signals feed Workday + LinkedIn HA below Recruiters · Edge case Workday escalations
● Below the algorithm - the configuration fires as HitL trap ◆ Contested - the configuration is splitting ▲ Above the algorithm - the configuration holds (substance / regulatory bind)

Hover any marked cell for the source.

The HITL trap is not a translator problem or a paralegal problem or a CS-agent problem. It is seven configurations firing across the knowledge-work corpus. The fallacy assumes "human-in-the-loop" labels a safe seat. The grid shows what it labels: a pattern of below-the-algorithm work distributed across the roles where AI has already crossed the cost curve.

The descent

How did the human-in-the-loop end up below the algorithm?

Four roles where the descent is documented, dated, and load-bearing in the index. Pick a role and watch the algorithm line clear each rung of the bundle. Each chapter cites the platform event, the wage delta, and the named platform or practitioner that absorbed the rung.

Pick a role

Constraint that held the bundle: Bilingual human capital with sector-specific accumulation. The buyer paid $0.10-$0.20 per word plus project-management overhead.

6

Direct client relationship

5

Project pricing + scoping

4

Terminology + glossary ownership

3

Cultural-context judgment

2

Comprehension + production

1

CAT-tool / QA execution
The mechanism

What does "human in the loop" mean in practice?

There is no single 'human in the loop.' There are seven.

Each configuration is a distinct pattern of how a human gets relocated as cleanup labour for an AI pipeline. The configuration determines whether the practitioner sits above or below the algorithm. Most configurations land below.

Cleanup

Below

Human cleans AI output before it reaches the buyer.

Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design.
Example
Translation post-editors

Post-edit machine-translation output at per-word piecework rates. The output then feeds the next model's training corpus.

Surface coherence check, no substance test. The post-edited output gets logged as translation memory and becomes the next model's training data. The worker's expertise transfers to AI capital while their per-word rate stays at $0.03-$0.05.
Brynjolfsson 2023 m06 attrition study; Gray-Suri 2019 ghost-work corpus

Surface review

Below

Human grades AI output for coherence and tone.

Substance testDurability modifierSurface-vs-substance validatorValidation work captures premium only when the validator tests substance against ground truth. It collapses when the validator can only check for surface coherence - because AI is structurally better than humans at producing surface coherence. The seat depends on whether ground-truth tests exist for the domain.Where it shows up:Citation-audit paralegals testing against actual case law (substance check); AI-output reviewers at content platforms judging coherence only (surface check); radiologists doing AI-overread against the image (substance check); compliance copy review without a statute-grounded checklist (surface check).
Example
Compliance-copy reviewers

Scan AI-drafted documents for surface coherence without a statute-grounded checklist. AI is structurally better at producing coherent text than humans are at checking it.

Collapses on the Dell'Acqua-Mollick finding. BCG consultants using GPT-4 in the 2023 experiment produced more polished output on every task - including the tasks where the model's answer was wrong. The day a buyer runs an internal version of that experiment, the surface-review headcount has nothing to show.
Dell'Acqua-Mollick 2023 m03 (BCG x HBS Working Paper 24-013)

Substance validation

Above

Human tests AI output against ground truth.

Substance testDurability modifierSurface-vs-substance validatorValidation work captures premium only when the validator tests substance against ground truth. It collapses when the validator can only check for surface coherence - because AI is structurally better than humans at producing surface coherence. The seat depends on whether ground-truth tests exist for the domain.Where it shows up:Citation-audit paralegals testing against actual case law (substance check); AI-output reviewers at content platforms judging coherence only (surface check); radiologists doing AI-overread against the image (substance check); compliance copy review without a statute-grounded checklist (surface check).
Example
Citation-audit paralegals; AI-overread radiologists; sworn translators

Verify AI-generated citations against actual case law. Compare AI radiology impressions against the underlying image. Attest fidelity to source text under legal liability.

Holds because the test is against ground truth, not the AI's coherence. The canonical case is Mata v Avianca (June 2023): the attorney filed an AI-generated brief, the citations sounded plausible, none of the cases existed, the court sanctioned the attorney. The substance-audit seat became more valuable the day that ruling published.
Choi-ChatGPT 2023 m03; Mata v Avianca, S.D.N.Y. June 22, 2023

Training data

Below

Human labels examples that train the AI.

Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design.
Example
Data annotators, RLHF labelers

Label examples, write preferences, write demonstrations. Direct ghost-work configuration: the worker's output is the AI pipeline's input, with no client on the other side.

Appen revenue $407M FY24; Scale AI valued at $13.8B May 2024. The capital captured by the platforms is built on labellers in Kenya and the Philippines whose criticality to the AI capital stack is total, and whose wages are piecework. Data Labelers Association (Kenya) and Code AI (Philippines) are the early formal-labor organising responses.
Gray-Suri 2019 Ghost Work; AI Now Institute Algorithmic Management 2023

Eval criterion

Above

Human writes the rubric / golden set the AI is measured against.

Question framingWhere value re-anchorsInquiry-framing premiumWhen AI answers well-posed questions at parity, rent moves upstream to whoever poses them. The worker captures premium by owning the question-frame the customer didn't know they needed - not by owning the customer but by holding problem-definition authority.Where it applies:Product managers, management consultants, data analysts deciding what to measure, market researchers, M&A bankers framing deal theses, strategy partners.
Example
AI evaluation engineers, eval ops leads

Author the criteria, build the golden set, design the test suite the model is measured against. Owning the eval is owning the question the AI is for.

Holds when the rubric is non-trivial and contested. Anthropic and OpenAI eval engineers writing the benchmarks the next model is graded against; frontier-lab red-teamers designing the safety tests. The seat decays only if benchmark authoring becomes mechanically scored (the GLUE-then-SuperGLUE pattern) and the value migrates to the next harder rubric.
Korinek-GenAI 2023 m03; emerging eval-engineer practitioner records at frontier labs

Edge case

Below

Human handles cases the AI can't.

Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design.
Example
Escalation-tier customer support, exception handlers

Take the residual cases the AI declines or fails on. Compensation prices the per-case throughput, not the criticality. Workflow owned by the vendor.

The Klarna pattern from Feb 2024 to May 2025. AI handled two-thirds of chats; the gig overflow handled the rest at ~$41/hr. When CEO Sebastian Siemiatkowski announced the reversal in May 2025, the gig agents had been the load-bearing layer the entire time - and they were absorbed back into Klarna headcount only because the framing flipped, not because the work had changed.
Klarna OpenAI Feb 2024 press; Klarna-Fortune May 9 2025 reversal; Brynjolfsson 2023 m02

Fail-over

Below

Human is the fallback when the AI breaks.

Worker feeds AIHow the work breaksWorker reframed as AI cleanup labourWhen AI absorbs a task the worker used to be paid for, the firm doesn't necessarily fire the worker. It relocates the worker as cleanup labour for the new automation - reduced agency, throughput-priced compensation. The worker's output now goes back into the AI pipeline rather than to a client.Where it applies:Median-tier work across knowledge roles - translation post-editing, AI-output review at content platforms, AI-assisted ticket handling in customer support, AI-rendered draft touchup in design.
Example
Reliability oncall for AI-served products

Stand by to take over when the AI fails. The human's presence is what lets the AI ship in production; the human is structurally on call rather than producing the offering. Throughput-priced.

Bayer Calantic announced Sep 2025 discontinuation of its radiology AI orchestration platform. The radiologist fail-over layer that had been priced as standby capacity through the pilot didn't move to above-the-algorithm work when the platform failed - the function was absorbed back into incumbent imaging-OEM workflows (GE HealthCare Imaging 360, Siemens AI-Rad Companion). Fall-back time is not platform equity.
Bayer Calantic discontinuation Sep 2025; Klarna May 9 2025 reversal
Where the capital went

What was the human-in-the-loop conversation worth in capital terms?

While the popular conversation argued whether humans would stay in the loop, the capital answered. Six AI-native vendors raised or compounded into the tens of billions on the work that used to belong to translators, paralegals, CS agents, copywriters, and engineers. The labour those vendors run on was priced as throughput.

Latest disclosed ARR or valuation, AI-native vendors · log scale
$100M $500M $1B $5B $10B Sierra Enterprise CX $100M ARR in under 2 years; $10B valuation Nov 2025 Decagon Enterprise CX $4.5B valuation tender offer Mar 2026; Deutsche Telekom deployment Cursor Software eng $2B ARR fastest-growing standalone tool in the corpus Harvey AI Legal practice $11B valuation 100K+ lawyers, 1,300 orgs, 42% of AmLaw 100 Lovable Non-eng software $400M ARR $0 to $400M ARR in 18 months; $6.6B valuation Bureau Works Translation Profitable, founder-led AI-inside-offering; owned MT stack; bootstrapped Sources: vendor-map cross-corpus (corpus/overview/structural/_vendor-map.md): TechCrunch, Sacra, Contrary Research, Sierra customer pages, Decagon press, Harvey blog (Mar 2026), public ARR disclosures.
What the labour above ran on · 2024-26 disclosed rates
  • Translation post-editor $0.03-$0.05 per word 2024-26 mid-market rate
  • Data annotator (Appen) $1.32-$2.00/hr Kenya / Philippines baseline
  • Klarna gig CS agent ~$41/hr Pre-May 2025 reversal
  • Per-stamp architect-of-record review $80-$120 per stamp AEC AI-draft sign-off
The spectrum

Is your work an input to the AI, or an output to a client?

The line between above and below is not a switch. It is a position on a spectrum. Where on the spectrum does the customer pay for what you ship, versus where does the AI pipeline consume what you produce?

Input to the AI your work feeds the pipeline Output to the client your work is the offering
Data annotator
RLHF labeller; Appen / Scale
Translation post-editor
$0.03-$0.05 per word; output is training memory
Klarna gig CS agent
Pre-May 2025; ~$41/hr; vendor owns workflow
Workday recruiter
Operates the scoring system; 1.1B rejections in Mobley discovery
AI-output reviewer
Paralegal surface-grading AI drafts
Citation-audit paralegal
Substance check against case law (post-Mata 2023)
Eval engineer
Frontier-lab rubric author; the question is the offering
Senior architect + Cursor
Owns system; AI accelerates the build
Bureau Works founder
Owned MT + customer contract; AI inside the offering
The diagnosis

Why does the framing keep producing roles below the algorithm?

The residual theory produces four canonical roles. Three of them land below the algorithm by default.

When the framing starts with 'what can AI not yet do?' it produces a predictable set of compensating-layer jobs. Each one looks defensible because it answers a real gap in the model. But most of them sit below the algorithm unless a specific condition is met.

Validator

below

"Humans check AI output to make sure it's right."

Anchors to above on Substance testDurability modifierSurface-vs-substance validatorValidation work captures premium only when the validator tests substance against ground truth. It collapses when the validator can only check for surface coherence - because AI is structurally better than humans at producing surface coherence. The seat depends on whether ground-truth tests exist for the domain.Where it shows up:Citation-audit paralegals testing against actual case law (substance check); AI-output reviewers at content platforms judging coherence only (surface check); radiologists doing AI-overread against the image (substance check); compliance copy review without a statute-grounded checklist (surface check).

The test must be against ground truth, not the AI's coherence. Mata v Avianca (S.D.N.Y. June 22 2023) is the canonical sorting event: the attorney filed an AI-drafted brief with fabricated citations that sounded plausible, the court sanctioned him, citation-audit specialists became the new scarce role. If the validator only grades how well the AI writes, Dell'Acqua-Mollick 2023 predicts the seat collapses.

Below the algorithm AI-output reviewer paralegals at content platforms; compliance-copy reviewers without statute checklist Above the algorithm Citation-audit paralegals comparing AI output against actual case law; AI-overread radiologists comparing impression against underlying image; sworn translators attesting fidelity to source under legal liability

Relationship manager

below

"Humans handle the empathy AI lacks."

Anchors to above on Customer ownershipWhere value re-anchorsCustomer-problem captureThe practitioner moves from selling component-tasks to owning the customer's new problem - an outcome the buyer can't get elsewhere, anchored to a constraint the practitioner uniquely holds. AI is ingredient inside the offering; the practitioner owns the customer relationship.Where it applies:Senior-tier roles in every exposed knowledge profession - senior bankers, consultants, M&A partners, UHNW advisors, agency owners, brand-anchored designers.

The contract has to run through the practitioner, not the platform. Gabriel Fairman at Bureau Works owns the customer relationship and the AI stack inside the offering. The Klarna gig CS agent (pre-May 2025) did not - the contract sat with Klarna, the workflow sat with OpenAI, and the agent carried the residual emotional labour at ~$41/hr until the framing flipped.

Below the algorithm Klarna gig customer-service agents; recruiters operating Workday algorithmic scoring Above the algorithm Account leads who own the contract and the renewal; vertical-AI founders (Bureau Works, Aha!, Midjourney) whose customer relationship is direct

Accountability officer

above (rare) or below (common)

"Humans bear liability AI cannot."

Anchors to above on Authority binderWhere value re-anchorsRisk concentration / authority bindingComponents bearing institutional risk - signature, seal, regulatory accountability, professional liability - become more valuable, not less, because they absorb the new failure modes automation introduces. Insurance, regulators, and standards bodies price the residual risk into licensure language.Where it applies:Radiologists, surgeons, trial litigators, M&A partners, UHNW advisors, sworn translators, CPAs.

Formal authority must bind to the seat. Radiologists hold the final read under the FDA AI/ML device pathway; CPAs hold audit signoff under PCAOB rules; trial litigators hold the courtroom; sworn translators hold court / asylum / immigration attestation. An AI-ethics-officer title without enforcement authority collapses on the first incident - watch IBM, Microsoft, and Google's 2023-2024 'responsible AI' layoff rounds for what a non-binding accountability seat is worth.

Below the algorithm Risk-and-compliance reviewers without statutory authority; AI-ethics-officer titles without enforcement power Above the algorithm Radiologists under FDA AI/ML device approvals; CPAs under PCAOB audit signoff; trial litigators in courtroom; sworn translators in court / asylum / immigration

System architect

below (default) or above (rare)

"Humans design the agent workflows AI runs inside."

Anchors to above on Question framingWhere value re-anchorsInquiry-framing premiumWhen AI answers well-posed questions at parity, rent moves upstream to whoever poses them. The worker captures premium by owning the question-frame the customer didn't know they needed - not by owning the customer but by holding problem-definition authority.Where it applies:Product managers, management consultants, data analysts deciding what to measure, market researchers, M&A bankers framing deal theses, strategy partners.

The practitioner must own the question, not the configuration. Senior software architects who decide what the system is for capture rent; prompt engineers configuring vendor workflows don't. M&A partners framing the deal thesis Harvey accelerates capture rent; Workday operators clicking through scoring screens don't. The test: if you were removed, would the AI produce work for the wrong question or no work at all?

Below the algorithm Prompt engineers configuring vendor workflows; agent ops staff inside vendor products Above the algorithm Senior software architects deciding what the system is for; AI eval engineers authoring non-trivial golden sets; M&A partners framing the deal thesis the AI then accelerates
First error

What does the fallacy do to roles?

At the level of job design, the fallacy produces roles that sit below the algorithm. Humans are asked to review outputs, handle exceptions, correct errors, label edge cases. These jobs preserve employment and often reduce leverage. The worker is present; the workflow, customer relationship, pricing power, and learning loop sit elsewhere.

The real question for any role is not whether a human remains in the loop. It is where that human is positioned to own the constraint, the context, the accountability, or the outcome.

The four above-the-algorithm anchors Authority binderWhere value re-anchorsRisk concentration / authority bindingComponents bearing institutional risk - signature, seal, regulatory accountability, professional liability - become more valuable, not less, because they absorb the new failure modes automation introduces. Insurance, regulators, and standards bodies price the residual risk into licensure language.Where it applies:Radiologists, surgeons, trial litigators, M&A partners, UHNW advisors, sworn translators, CPAs. Customer ownershipWhere value re-anchorsCustomer-problem captureThe practitioner moves from selling component-tasks to owning the customer's new problem - an outcome the buyer can't get elsewhere, anchored to a constraint the practitioner uniquely holds. AI is ingredient inside the offering; the practitioner owns the customer relationship.Where it applies:Senior-tier roles in every exposed knowledge profession - senior bankers, consultants, M&A partners, UHNW advisors, agency owners, brand-anchored designers. Validation premiumWhere value re-anchorsValidation migrationWhen the doing-side of work is automated, the buyer's new scarcity is the certified fit of AI output to the buyer's risk-tolerance, regulator, or downstream system. Value migrates from producer to validator. The validator captures real leverage only if the validation tests substance against ground truth - AI is structurally better than humans at producing surface coherence.Where it applies:Citation-audit paralegals, AI-overread radiologists, compliance-copy reviewers, AI-evaluation engineers, sworn translators. Question framingWhere value re-anchorsInquiry-framing premiumWhen AI answers well-posed questions at parity, rent moves upstream to whoever poses them. The worker captures premium by owning the question-frame the customer didn't know they needed - not by owning the customer but by holding problem-definition authority.Where it applies:Product managers, management consultants, data analysts deciding what to measure, market researchers, M&A bankers framing deal theses, strategy partners.

The job-design implications page works through the position test role by role, names the structural condition each above- the-algorithm seat rides on, and walks through the four seats practitioners are reaching today.

Read the job-design implications Coming soon
Second error

What does it do to organizations?

At the level of organizational design, the same framing produces a deeper error. Firms redesign around what AI can and cannot do inside yesterday's workflow. Agents take over the scalable work. Humans fill the gaps. The org becomes a map of machine capability and human leftovers.

That is automation with a human safety layer. It is not the organization a new capability portfolio makes possible.

A genuinely AI-native organization starts from a different question. Given the combination of human and agent capabilities now available, what new problems can the firm solve that it could not solve before? The org chart follows the frontier of solvable problems, not the residual limitations of the tools.

The org-design implications page works through the precedent (cloud at Netflix, computerized reservation systems at the airlines), names the AI-era parallels firm by firm (Sierra, Decagon, Harvey at A&O Shearman), and surfaces the canonical residual-theory failure (the Klarna walk-back).

Read the org-design implications Coming soon
Self-diagnostic

Where do you sit on the matrix?

Six questions. The first asks whether AI is inside your daily work. The next five ask where you sit in relation to the workflow, the customer, and the offering. The result places you on the matrix above and names the durability test firing on your seat.

Question 1 of 6

Is AI inside your daily work?

Whether you use it directly (Copilot, Cursor, Harvey, Smartcat, Workday) or operate downstream of it.

Question 2 of 6

Where does your output go after you produce it?

The next consumer of your work - the AI pipeline, another reviewer, or a paying client.

Question 3 of 6

What does the buyer actually pay you for?

The unit of value the engagement transacts on.

Question 4 of 6

Who is your buyer-of-record?

Who the contract or invoice runs through.

Question 5 of 6

Do you hold a constraint the system can't replicate?

A signature, seal, licence, regulatory seat, or institutional authority.

Question 6 of 6

Who decides which problems the AI gets routed to?

Question framing - who chooses what the AI is for.

Your quadrant
-

-

-

Durability flag -

Practitioners in the same quadrant

    Where to next:

    What this means for designing your role What this means for designing your organization
    Jevons Misunderstanding The sister fallacy. Efficiency expands the market, but the gains route through position, not occupation. The Jobs Reshuffle Index Per-role analysis of how AI is reshuffling specific knowledge-work jobs. Reshuffle The book, the framework, and everything we've built around it.