Narrative-first, evidence-driven

Back to case studies [pdf]

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

Note (Early Release): This is the initial release of our AI case-study work. A full AI case-study series is in progress and is expected to be complete by March 31, 2026. We released this early because the topic is important enough to share now.

Label convention for this case

• confirmed (source) = empirical claim grounded in named sources
• plausible = supported directional claim with incomplete or still-maturing evidence
• inference = synthesis drawn from multiple sources or patterns
• working frame = normative or design standard used to evaluate the case

What they did

From 2022 through early 2026, firms, governments, and workers integrated generative AI into writing, customer support, coding, search, workflow routing, and early agentic task execution. The practical shift was not “AGI arrived.” It was that a general-purpose cognitive tool moved into ordinary work much faster than labor-market, governance, and training systems adapted.

Why it worked (or didn’t)

It worked, in the narrow sense, because generative AI is cheap to try, easy to deploy over existing digital infrastructure, and useful for many language-heavy tasks without requiring every firm to build its own software stack. That made adoption fast. (confirmed (source) - NBER rapid adoption paper, OECD 2026)

It did not automatically preserve broad human relevance because the same mechanism that made adoption easy also brought substitution pressure quickly to roles with standardized cognitive tasks and fragile entry ladders. (inference)

The strongest near-term pattern is not full human replacement. It is uneven transformation. (confirmed (source) - ILO/NASK, NBER, Stanford SIEPR) Some workers become more productive. Some firms reduce low-level work or hiring. Some teams use AI as leverage while others use it as a headcount gate. (plausible)

The deeper issue is institutional, not only technical. If AI improves output while shrinking junior pathways, concentrating gains, and moving practical authority into opaque systems, then productivity can rise while the middle weakens. That is the central risk being tested here. (plausible)

Mechanism evidence

What actually moved:

• U.S. adult use: nearly 40% of adults age 18-64 had used generative AI by late 2024; 23% of employed respondents had used it for work in the prior week; 9% used it every work day. (confirmed (source) - NBER)
• Work-hour penetration: generative AI was already assisting an estimated 1% to 5% of all work hours, with time savings equal to roughly 1.4% of total work hours. (confirmed (source) - NBER)
• OECD adoption: more than one-third of individuals across the OECD used generative AI tools in 2025; firm-level AI use in reporting OECD countries rose to 20.2% in 2025 from 14.2% in 2024 and 8.7% in 2023. (confirmed (source) - OECD)
• Customer-support productivity: access to a generative AI assistant increased productivity by 14% on average and by 34% for novice and lower-skilled workers. (confirmed (source) - NBER)
• Knowledge-work pattern: workers in a six-month field experiment spent about two fewer hours per week on email and less time working outside normal hours, but individual access alone did not clearly change total task composition. (confirmed (source) - NBER)
• Early-career labor effect: workers ages 22-25 in highly AI-exposed occupations experienced a 13% relative decline in employment after broad generative-AI adoption. (confirmed (source) - Stanford SIEPR)
• Exposure/adaptation split: of 37.1 million workers in the top quartile of AI exposure, 6.1 million are in occupations with low expected adaptive capacity, concentrated in clerical and administrative roles. (confirmed (source) - NBER)

Time horizon: Early. The strongest evidence runs from 2022 through early 2026. This is long enough to establish rapid adoption and early labor-market pressure. It is not long enough to establish the final wage, productivity-sharing, or institutional equilibrium.

Counterfactual: No clean counterfactual exists. The best available evidence is a mix of field experiments, labor-market exposure studies, and institutional reporting. We can see rapid deployment and some early effects. We cannot yet prove the medium-run distribution of gains.

Load-bearing variable: The load-bearing variable is whether institutions create enough new labor-complementary work, enforce meaningful human accountability, and preserve entry pathways fast enough to offset substitution pressure. If not, human relevance becomes narrower, more elite, and more conditional.

Label uncertainty: rapid adoption is (confirmed (source) - NBER rapid adoption paper, OECD 2026). Early productivity gains are (confirmed (source) - NBER field experiments). Broad long-run shared gains are unknown.

Tripwire update rule: This remains a generative-AI diffusion case, not an AGI-arrival case, unless the economic pattern changes enough to justify a stronger frame.

Treat this case as moving into an AGI-like economically phase only if several of these become true at once:

• sustained autonomous multi-step task completion in production settings without narrow human handholding (plausible threshold, not yet met)
• measurable replacement of junior task bundles across multiple occupations, not just isolated task assistance (unknown)
• stable firm-level output with lower headcount in roles that previously required junior-to-mid skill ladders (unknown)
• routine use of agentic systems in high-stakes business processes with limited human override in practice (plausible, uneven)
• strong evidence that new task creation is not keeping pace with displacement in exposed occupations (unknown)

These are economic tripwires, not metaphysical ones. They tell the repo when this stress test should be updated from “rapid diffusion with early labor effects” to “frontier capability is materially reordering labor demand.” (working frame)

Guardrails

The guardrail question here is not only whether models are safe. It is whether humans remain meaningfully in command.

Three oversight concepts matter:

• human-in-the-loop: a person can intervene in the decision cycle. (confirmed (source) - EU trustworthy AI guidance)
• human-on-the-loop: a person supervises operation and can intervene if needed. (confirmed (source) - EU trustworthy AI guidance)
• human-in-command: a person or institution decides whether the system should be used at all, in what contexts, and with what override rights. (confirmed (source) - EU trustworthy AI guidance)

That hierarchy matters because nominal human review is not enough. If the system is too fast, too opaque, too cheap to overrule, or too deeply embedded in workflow, then a person can remain formally present while becoming practically irrelevant. (inference)

The most relevant guardrails in this case are:

• notice when AI is used in hiring, promotion, dismissal, task assignment, or workplace monitoring (confirmed (source) - OECD workplace guidance, EU AI Act workplace treatment)
• understandable reasons and a real path to challenge outcomes (confirmed (source) - OECD Employment Outlook 2023, OECD AI Principles)
• traceability, record-keeping, and clear assignment of responsibility (confirmed (source) - NIST AI RMF, OECD AI Principles)
• human accountability for consequential decisions (confirmed (source) - OECD Employment Outlook 2023, EU AI Act Article 14)
• limits on algorithmic management where workers cannot contest hidden scoring or opaque ranking (plausible)
• preserved training pathways so junior work is not eliminated before replacement ladders exist (plausible)

Workplace AI Bill of Rights - plain-language version

• notice: people should know when AI is being used in hiring, evaluation, scheduling, or discipline
• reason: people should be able to get a plain-language explanation of consequential decisions
• contest: people should have a real appeal path, not only a help desk
• log: organizations should keep records sufficient to reconstruct what the system did
• override: qualified humans should be able to stop, reverse, or escalate consequential outputs
• audit: independent review should be possible for high-stakes systems
• retaliation protection: workers should be able to question or challenge AI-mediated decisions without punishment

Observability is part of this guardrail stack, not separate from it. Security without observability is weak because institutions cannot see misuse, drift, failure, or overreach. Trust without observability is also weak because end users cannot inspect provenance, limits, confidence, review history, or recourse. (inference)

Useful operating patterns include kill switches, circuit breakers, evaluations, scoring thresholds, audit logs, escalation paths, fallback modes, and drift monitoring. They help institutions see, challenge, and stop consequential failures. (working frame)

How generative AI changes market power

• vendor lock-in: firms can become dependent on a small number of model vendors and surrounding workflow tools. In practice, switching costs can include prompts, eval harnesses, fine-tuning data, agent tooling, and policy logs that do not port cleanly. (plausible)
• procurement opacity: public and private buyers may not know enough about model behavior, data practices, or failure history to make informed comparisons. In practice, the buyer may see a benchmark sheet and a sales promise, but not enough incident history or audit detail to compare vendors on governance quality. (plausible)
• audit asymmetry: vendors and employers can often see more of the system than workers, end users, or regulators can. In practice, the institution controlling the tool may know the prompts, thresholds, and logs while the affected person only sees the result. (plausible)
• proprietary scoring: AI-mediated ranking in hiring, performance management, and workflow assignment can shift power toward institutions that control the model and away from people being scored. In practice, a worker can be sorted by a model-mediated score they cannot inspect or challenge in full. (plausible)
• headcount gating: firms may use AI not only to augment work but to justify delayed hiring, especially in junior roles. In practice, “use the tool first” can become a default answer before a team is allowed to add people. (plausible)

Enforcement check: This is where the case is weakest. Right now, a lot of U.S. policy is “go faster”; a lot of EU policy is “go, but prove you can control it.” (inference) In the U.S., the dominant 2025 federal policy direction favored acceleration, infrastructure, and adoption more than strong labor guardrails. That means enforcement of broad human-protective guardrails is uneven and still forming. (confirmed (source) - White House AI Action Plan, OECD/EU governance materials)

Where it broke (or where it’s under strain)

The first strain point is the entry ladder.

The most credible early warning is not mass unemployment. It is that AI may remove or compress the junior tasks through which people become useful, trusted, and eventually senior. (confirmed (source) - Stanford SIEPR, NBER AI labor-market paper) If firms keep senior decision-makers while reducing junior analyst, support, or drafting work, then the profession may survive while the pathway into it weakens. (plausible)

The second strain point is bargaining power.

If firms can raise output with fewer early-career workers, or use AI as a gate before approving new headcount, workers become easier to sort, monitor, and replace at the margin. Gains can then flow upward without broad wage security following. (inference) Historical analogs from automation and early industrialization support that warning. Productivity alone does not protect labor. (confirmed (source) - Acemoglu/Restrepo 2019, Acemoglu/Johnson 2024)

The strongest counter-argument is that this may still become a complementarity story rather than a hollowing-out story.

Some of the best evidence points that way. In customer support, novice workers saw large productivity gains. In product-development work, AI helped individuals perform more like teams and reduced silo costs. Separate research also suggests that some human leadership and judgment skills remain load-bearing in AI-rich environments. (confirmed (source) - NBER customer-support paper, NBER cybernetic teammate paper, NBER leadership-with-AI-agents paper) On that reading, the likely medium-run outcome is not broad human irrelevance. It is a reorganization of work in which humans remain central as supervisors, integrators, relationship managers, and accountable decision-makers. (plausible)

The problem is that this stronger pro-complementarity case still does not answer the distribution question. It shows that some humans remain central. It does not yet show that broad entry paths, bargaining power, or shared gains will survive at scale. (inference / unknown)

The third strain point is pseudo-oversight.

A human can appear to be in the loop while the system has already ranked the candidates, summarized the evidence, proposed the action, and set the pace. In that setup, the human may function more as formal sign-off than as a true decision-maker. That preserves formal accountability while weakening real accountability. (plausible)

The fourth strain point is scope.

This is not only a coding or office-work story. The risk is broader institutional drift toward systems that make AI more authoritative while making workers and end users less able to inspect or challenge what the system is doing. When that happens, human relevance erodes even if many jobs still exist on paper. (plausible)

Scale check: This is already a system-level stress test, not a narrow product case. The adoption pattern is broad enough across office, service, and managerial workflows that the key question is no longer whether AI matters. It is whether the surrounding institutions are strong enough to shape how work, accountability, and gains are distributed.

Market verdict (or public verdict)

The market verdict so far is favorable to adoption and unfavorable to complacency.

Firms are adopting because the tool is useful, cheap to test, and capable of delivering measurable productivity gains in some settings. That is real. (confirmed (source) - NBER rapid adoption paper, NBER field experiments, OECD 2026)

But the public-interest verdict is not stable. We do not yet know whether the gains will flow into median wages, better working conditions, shorter hours, cheaper services, or simply higher margins with thinner entry-level hiring. (unknown)

The best current reading is that adoption is being rewarded faster than broad human-protective guardrails are being built. That is not yet proof that the resulting equilibrium will be good for broad human relevance.

Policy environment

The current policy environment is split.

One branch, especially in the United States, is oriented toward capability, infrastructure, diffusion, and geopolitical competition. (confirmed (source) - White House AI Action Plan)

Another branch, stronger in OECD and EU governance material, is oriented toward trustworthy use, human oversight, workplace accountability, challenge rights, and high-risk controls for employment-related systems. (confirmed (source) - OECD AI Principles, OECD Employment Outlook 2023, EU AI Act)

A new 2026 signal is that governance conflict is no longer only conceptual. Public reporting and statements around frontier-lab/defense engagement indicate active disputes over autonomous-use boundaries and domestic-surveillance constraints. Treat that as a confirmed direction in policy conflict, not as proof that the most extreme uses are already deployed at scale.

For this pattern to scale without hollowing out broad human relevance, policy would need to do more than encourage adoption. It would need to preserve entry ramps, require human accountability in consequential domains, limit opaque workplace scoring and surveillance, and push productivity gains toward shared gains rather than only cost reduction. (plausible)

What would need to be true at policy level:

• consequential employment decisions cannot be meaningfully out of the loop
• workers can know when AI is being used and challenge outcomes
• firms cannot erase training pipelines through routine decisions while claiming future workers will somehow appear later
• observability and auditability are treated as prerequisites for trust, not optional extras
• fallback security and retraining systems are strong enough that adaptation is real rather than rhetorical

US Analog (if non-US case)

Closest US analog: not a single company or law, but the emerging pattern across U.S. white-collar and administrative work from 2022 through 2026.

Data maturity: Early.

What we can measure now:

• adoption rates
• firm-reported use across business functions
• field-experiment productivity gains
• early-career employment effects in exposed occupations
• exposure and low-adaptation clusters in clerical and administrative work
• stated employer plans for upskilling, redeployment, and workforce reduction

Scoreboard for 2026-2028

• early-career job postings in highly exposed occupations
• internal promotion rates from junior to mid-level roles in exposed functions
• wage dispersion within exposed occupations
• median compensation growth vs. measured productivity growth in exposed firms or functions
• time-to-competency for new entrants using AI-heavy workflows
• ratio of augmentation use cases to headcount-reduction use cases in employer disclosures
• frequency of AI use in hiring, performance, scheduling, and discipline
• appeal and override rates for AI-mediated decisions
• vendor contestability: can organizations switch providers without major operational lock-in?
• audit-log completeness for high-stakes systems
• worker-reported clarity: do affected people know when AI was used and how to challenge outcomes?

What we must not overclaim yet:

• that AGI is imminent
• that current early-career losses will necessarily become broad mass displacement
• that productivity gains will translate into broad wage growth
• that keeping a human formally present is the same as meaningful human control
• that “learn AI” is a sufficient policy response
• that current governance conflict alone proves broad long-run labor displacement

What good looks like - one institutional example

A hiring pipeline that keeps humans in command would do six things:

1. Tell applicants where AI is used and where it is not.
2. Keep a durable log of ranking, filtering, and override events.
3. Limit autonomous rejection in high-stakes screening.
4. Give hiring staff authority to reverse model suggestions without penalty.
5. Give applicants a real appeal path with human review.
6. Track whether the system is narrowing entry routes for junior candidates over time.

That is not a full solution to labor displacement. It is a minimum example of governance that keeps speed from replacing accountability by default.

For that example to be enforceable, two more things have to be true:

1. Someone with standing can audit it: internal audit, an independent reviewer, a worker representative, or a regulator.
2. Failure has consequences: pause the system, roll back the workflow, notify affected people, and remediate before reuse.

North Star verdict

This case complicates the loop unless institutions change course.

Generative AI does not, by itself, reinforce security -> choice -> competition -> shared gains -> more security. The current evidence is consistent with a world in which adoption outpaces protections for entry ramps, bargaining power, and accountability. Productivity up while middle-class leverage goes down is a fail. The technology can still fit the loop, but only if human accountability, training pathways, observability, and shared-gain institutions are treated as core design requirements rather than afterthoughts.

System lesson in one sentence:
Humans stay broadly relevant under generative AI only when institutions make augmentation, accountability, and shared gains stronger than substitution pressure and opaque control.

Research gaps

Rule: If any [RESEARCH GAP] is load-bearing for a claim in the top sections, do not publish that claim until the gap is closed.

1. [RESEARCH GAP: Distribution] No medium-run evidence yet shows whether generative-AI productivity gains will raise median wages, shorten hours, or mainly increase margins and top-end earnings.
2. [RESEARCH GAP: Entry ladder] No robust multi-year evidence yet shows whether early-career contraction in exposed occupations is a temporary adjustment or the start of a more durable ladder-collapse dynamic.
3. [RESEARCH GAP: Agentic workflows] The labor effects of agentic workflows are still less mature than the labor effects of chat-style assistance; enterprise direction is clear, but measured outcomes are still thin.
4. [RESEARCH GAP: Enforcement map] U.S. enforcement practice for workplace AI guardrails remains fragmented; a stronger jurisdiction-by-jurisdiction map would sharpen the policy section.

Back to case studies