Narrative-first, evidence-driven

Back to case studies [pdf]

Receipts: tracked in Methods and Sources by type: Primary documents | Official data | Independent analysis

A practical series for An Economy for Everyone (E4E).

These are not “AI takes” and they are not a catalog of every industry AI touches. They are bounded case studies chosen to keep scope sane while extracting portable lessons that apply across similar domains.

AI changes the game in two ways. Sometimes it amplifies an old failure by making it faster, cheaper, and harder to contest. Sometimes it unlocks a new operating model by making coordination, targeting, or content production possible at scales that were previously impractical. In both cases, the right question is not “is AI present?” but “what power does it shift, and who gains from that shift?”

The goal is shared reality and repeatable civic action:

• reduce the monthly squeeze
• keep systems contestable
• prevent quiet capture at the top
• keep real choice (exit) alive

• The common spine
• Why these are industry-specific
• How to read / use this series
• What to watch for as you read
• The case studies
• What “good” looks like

The common spine

Across domains, the same pattern repeats:

When decisions become fast, cheap, and opaque, “human review” becomes a rubber stamp. That’s how leverage moves from people to institutions through defaults, convenience, and time pressure.

Two tests show up everywhere:

Test 1: Human Command (minimum floor)

If AI affects a life outcome, you get:

• notice
• reason
• appeal
• records
• a human override

Test 2: Exit / captivity

If you can’t exit, you’re captured. When exit is not realistic, governance has to be stronger, not weaker.

Why these are industry-specific

We use specific industries as stress tests because they:

• make the mechanism concrete (real decisions, real harm, real incentives)
• provide measurable proxies (appeals, reversals, time-to-action, churn, gatekeeping)
• keep the research bounded
• let us generalize responsibly (“this pattern appears here; here’s why it should appear elsewhere”)

Each case study is meant to produce:

• a clear mechanism (“what’s happening and why”)
• a minimum floor (what must be true for contestability)
• levers (what to change, who can change it, and how to verify progress)

This also helps separate two different kinds of AI risk:

• old patterns getting scaled up to machine speed
• genuinely new capabilities that create new failure paths

You need both categories in view. If you miss the first, you treat today’s harms like they came out of nowhere. If you miss the second, you assume the future will just be the past turned up louder.

How to read / use this series

1. Read the one-paragraph “scene” at the top of each study (the lived reality).
2. Identify the mechanism (the repeatable pattern).
3. Apply the Human Command test and the Exit test.
4. Use the levers section to choose one action:
   • a personal / interpersonal move (repeatable by normal people)
   • a community / policy / procurement lever (repeatable by institutions)

This series is designed to be modular. Each case study uses a shared library of Mechanisms and Mechanism Modules. Not every case uses every mechanism equally, but the vocabulary stays stable across the series.

What to watch for as you read

A recurring pattern in this series is that AI is not automatically pointed in a good direction.

In every case study, the same core capability can usually be used in at least two ways:

• to reduce friction, widen access, improve accountability, and make a system fairer and easier to live with
• or to speed up extraction, hide responsibility, weaken contestability, and make the system more stressful, more opaque, and harder to escape

That does not mean “AI is neutral” in the lazy sense. It means the initial deployment choice usually reveals the real priority.

The default deployment path often leans toward short-term wins:

• faster throughput
• lower labor cost
• tighter control
• more hidden pricing power
• more nominal review with less real accountability

Those wins can be real in the short run and still be disastrous in the long run if they hollow out learning, weaken appeals, trap people in captive systems, or let the gains flow upward without improving life for the people affected.

So while reading each case, note three things:

1. What is the system optimizing for right now? Cost, speed, control, and margin? Or fairness, reliability, resilience, and better outcomes for the people using it?

2. How could the same capability be used in a better way? Not in theory — through what guardrails, incentives, ownership rules, disclosures, or procurement terms?

3. Is this a short-term win creating a long-term institutional loss? Faster denial, thinner apprenticeship, weaker trust, more capture, or more downstream stress for households and communities?

This is one of the main reasons the series exists. The goal is not just to describe AI harms after they scale. It’s to identify where the technology could make systems fairer, easier to navigate, and more accountable and where current incentives push it toward the opposite result.

The case studies

1) Lower walls, harder gates

What it studies: AI’s effect on market competition and new-entrant survival across software, creative, and professional services markets, specifically the gap between lowered production costs and unchanged distribution, certification, trust, and credential gates.

Why it’s useful: it surfaces the bottleneck-shift pattern across two gate types:

• In software and enterprise tech: the scarce resource becomes channel access such as procurement bundles, certification floors, and platform API dependency
• In professional services: the scarce resource becomes credibility, coverage, and credentials such as trust-based distribution, liability coverage, and regulatory standing

Generalizes to: any market where production cost deflation meets procurement bundling, regulatory certification, or trust-based distribution lock-in; including cloud services, healthcare IT, government contracting, consulting, legal, and financial services.

Primary levers: antitrust enforcement on bundling-as-foreclosure, FedRAMP reform at scale, documented alternatives evaluation as a buyer discipline, pooled insurance structures for small professional services firms, published AI governance standards from bar associations and FINRA, acqui-hire scrutiny under HSR merger standards.

2) The Moving Breadbox

What it studies: AI’s effect on the build-vs-buy threshold for internal software and the risk that point-solution vendors lose not to rival vendors but to their own customers’ ability to build.

Why it’s useful: it captures a different competition pattern from Lower Walls. There is no outside gate to clear. The buyer is internal. When build costs fall enough, demand disappears instead of shifting to another vendor.

Generalizes to: reporting and analytics tools, workflow automation, internal portals, integration connectors, vertical-niche point solutions, and any software category whose value proposition depended mostly on building being too expensive.

Primary levers: explicit build-versus-buy reviews, lightweight governance for citizen-developer tools, ownership and maintenance requirements for internal replacements, and apprenticeship policies that keep the savings from thinning the junior pipeline.

3) IT ladder collapse

What it studies: what happens to skill formation when “learning work” is automated away and entry-level ladders shrink.

Why it’s useful: it captures a second-order effect many miss: institutions can lose the ability to do the work even if they want to later. When apprenticeship collapses, oversight becomes fake.

Generalizes to: claims adjusters, caseworkers, compliance analysts, junior reporters/editors, entry-level accountants, paralegals, and any profession with a training pipeline.

Primary levers: protected learning-work quotas, apprenticeship funding, review rotations, “manual flight checks,” measurable hiring/mentorship ratios, and procurement rules that require retained human capability.

4) Content flood and gate shift

What it studies: near-zero cost content production and the resulting shift of power to distribution, ranking, and trust layers.

Why it’s useful: it demonstrates a general rule: when production becomes cheap, the scarce resource becomes attention and trust, so power moves to whoever controls ranking, verification, and gatekeeping.

Generalizes to: app stores, marketplaces, job boards, social platforms, academic publishing, reviews/ratings, and any domain where discovery is the choke point.

Primary levers: provenance and labeling, contestable ranking/takedown, anti-spam friction, preservation of credentials through pipelines, transparency requirements, and competition constraints on gatekeepers.

5) Claims and eligibility

What it studies: automated or semi-automated decisions about eligibility, coverage, payment, and access (insurance, benefits, healthcare admin, financial access).

Why it’s useful: this is where “rubber-stamp review” becomes visible in data:

• low appeal rates
• high overturn rates when appealed
• vague reason codes
• asymmetric logs

Generalizes to: government benefits, disability decisions, fraud flags, credit/underwriting, provider payment, any system where a “no” is cheap and contesting it is costly.

Primary levers: notice/reason specificity, appeal SLAs, audit logs, independent review, procurement requirements, portability/exit constraints.

6) Personalized pricing and steering

What it studies: individualized offers, routing, eligibility nudges, and price discrimination in captive or semi-captive markets.

Why it’s useful: it shows the shift from “pricing as information” to “pricing as extraction” when:

• comparison shopping is hard
• the offer is personalized
• the logic is hidden
• switching costs are high

Generalizes to: housing, insurance premiums, healthcare routing, lending, utility add-ons, consumer marketplaces, and any “surge” or segmentation system.

Primary levers: all-in pricing, disclosure at point of decision, non-discrimination audits (including proxy tests), contestable adverse actions, portability and switching tools, limits on exploitative personalization in captive markets.

7) Surveillance and coercion

What it studies: cheap sensing, cheap scoring, and cheap enforcement, especially where consequences arrive without adjudication.

Why it’s useful: it distinguishes two failure modes:

• accuracy failures (false positives / false negatives)
• rights failures (even if accurate: no contestability, no exit, no audit)

It also provides a clear test: scale without accountability produces coercion.

Generalizes to: workplace monitoring, gig deactivation, fraud detection, benefits enforcement, immigration enforcement, “risk scoring” in many forms, and private-sector data broker ecosystems.

Primary levers: independent corroboration rules, audit logs, query constraints, appeal rights with timelines, prohibited-use boundaries, kill switches, transparency reporting, and limits on warrantless purchase of sensitive data.

8) Physical world control

What it studies: AI’s shift from recommendation to direct actuation in utilities, logistics, buildings, and transport — where the affected person cannot exit, the operator holds the logs, and nominal human oversight exists without practical override.

Why it’s useful: it is the most direct application of the Human Command test in the series: when AI steers the physical systems you cannot avoid — your workplace pace, your building access, your city’s roads — the question of who can override it and who eats the failure becomes immediate and material.

Generalizes to: any physical infrastructure where AI controls access, pace, routing, or resource allocation and where the affected party has no realistic exit: grid management, warehouse and gig logistics, residential building systems, autonomous vehicles, and AI transit.

Primary levers: mandatory log access rights for affected parties and regulators, real human review requirements for deactivation and access decisions, national AV liability assignment (UK ADSE model), OSHA enforcement authority over algorithmically caused injury, and worker deactivation rights ordinances modeled on Seattle’s.

9) Creator backlash, betrayal, and authenticity collapse

10) Young worker ladder shift

What it studies: how AI-enabled workflow changes are tightening some early-career gates by weakening old proof-of-work signals while leaving accountability with the human.

Why it’s useful: it turns a vague “kids should just learn AI” story into a more precise labor-market and governance claim: cheap first-pass output changes what counts as proof, but oversight, liability, and real judgment do not disappear.

Generalizes to: junior analyst roles, entry-level professional services, early-career knowledge work, apprenticeship-heavy fields, and any job family where first-draft work used to double as training.

Primary levers: apprenticeship protection, hiring transparency for true entry-level roles, real human-review standards in high-stakes workflows, and better measurement of entry-cohort conditions in AI-exposed occupations.

What it studies: what happens when creative work is used to build generative systems before consent, compensation, provenance, and contestability are in place.

Why it’s useful: it shows that creator backlash is not just a culture-war reaction to new tools. It is a governance stress test where training-data use, labor devaluation, discovery flooding, trust failures, and institutional accommodation all converge.

Generalizes to: any market where cheap synthetic output competes against the people whose work trained it and where distribution, labeling, and trust are controlled by platforms or institutions with weak accountability.

Primary levers: consent before training, consumer-facing provenance labels, contestable ranking and takedown systems, compensation pathways for licensed training data, and contract floors for replica and replacement risk.

What “good” looks like

What good looks like is not “AI everywhere” or “AI nowhere.” It is AI used in ways that lower friction for normal people, preserve real choice, keep human command meaningful, and make sure gains are actually shared instead of captured upstream.

A system that uses AI well tends to have:

• visible and actionable reasons at decision time
• cheap, real appeals (not symbolic)
• logs that can be inspected by an auditor and, when appropriate, by the affected person
• meaningful human overrides with authority and time
• clear exit paths or stronger governance where exit is impossible
• evidence that gains are shared (lower prices, better access, reduced admin drag) rather than captured as margin

Back to case studies