AI

AI at Work: Savings, Resistance, and the Productivity Illusion

AI scales enterprise operations → workers push back and metrics stay flat

Level 1

AI at Work, Workers Push Back

Enterprises are deploying AI agents to automate worker tasks at scale, with some firms like Salesforce reporting $100M in annualized cost savings. Yet a sweeping study of 6,000 executives finds nearly 90% of firms report no measurable impact on employment or productivity. Meanwhile, Chinese tech workers are building tools to sabotage AI workflow-capture systems their employers are mandating.

Bullets

  • Salesforce's Agentforce handled 3 million support conversations and cut support costs by $100M annually
  • A study of 6,000 executives found 90% say AI had no impact on productivity or employment in three years
  • Chinese workers are creating counter-tools to feed AI workflow-capture systems useless data
  • Worker confidence in AI utility dropped 18% globally even as usage rose 13% in 2025

Key Points

  • Enterprise AI is delivering isolated wins but broad productivity metrics remain flat
  • Worker resistance to AI workflow automation is becoming organized and technical
  • The gap between AI investment narrative and macro data is reviving Solow's 40-year-old productivity paradox

Timeline

Feb 2026

NBER study of 6,000 executives published, finding 90% report no AI impact on productivity or employment

Apr 4 2026

Koki Xu publishes anti-distillation GitHub tool to sabotage AI workflow capture, drawing 5M likes

Apr 18 2026

Fortune reports Salesforce Agentforce has delivered $100M annualized savings and influenced 3,200+ opportunities

Apr 20 2026

MIT Technology Review reports Chinese tech workers training AI doubles amid growing backlash

Sources

MIT Technology Review

5 days ago

Fortune

7 days ago

Fortune

2 months ago

Level 2

The Productivity Gap Is Real

The collision between AI's enterprise promise and its measurable reality is now statistically documented. A small cohort of advanced adopters like Salesforce are unlocking genuine gains, but they represent the exception, not the rule. Worker resistance is no longer passive skepticism — it is becoming a technical, organized, and legally contested phenomenon. The macro data is failing to confirm what the earnings calls assert.

Key Points

  • 90% of surveyed executives report zero AI impact on productivity or employment, echoing Solow's 1987 IT paradox — AI investment is not yet showing up in the numbers that matter
  • Salesforce is an outlier, not a benchmark — its $100M in savings and 3,200 influenced opportunities reflect deep product integration unavailable to most enterprises
  • Worker resistance has crossed from sentiment into engineering — the anti-distillation tool is a technical countermeasure, not a protest sign, signaling a new phase of labor-AI conflict
  • Declining worker confidence in AI utility, down 18% globally despite rising usage, points to a widening gap between deployment mandates and on-the-ground effectiveness
  • The productivity unlock may be a timing issue: historical precedent from the 1990s IT boom suggests gains can lag investment by a decade or more

Sources

MIT Technology Review

5 days ago

Fortune

7 days ago

Fortune

2 months ago

Level 3

What Actually Changes

The enterprise AI deployment wave is producing two distinct classes of firms: a small vanguard capturing compounding gains in cost and revenue, and a large majority investing heavily with little measurable return. Simultaneously, a new labor dynamic is emerging where workers are not just resisting AI but actively engineering countermeasures, raising novel legal, ethical, and operational challenges for HR, legal, and tech leadership. The productivity paradox is restructuring who captures value from AI — and it is not yet workers, and not yet most firms.

Key Points

  • A structural two-tier enterprise landscape is forming between AI-native operators and AI-adopting laggards
  • Worker resistance to AI workflow capture is evolving from cultural to technical, creating compliance and IP risks for enterprises
  • Productivity gains from AI remain concentrated in efficiency metrics, not yet in GDP-level or employment-level data

Timeline

Late 2022

ChatGPT launch triggers enterprise AI investment cycle; Federal Reserve later records 1.9% cumulative productivity uptick from this baseline

2025

AI adoption focus is operational: automate support, cut costs, reduce repetitive labor; Salesforce deploys Agentforce across help portal

Feb 2026

NBER study finds 90% of 6,000 executives report no AI impact on productivity or employment over three years

Apr 2026

Chinese tech workers respond to boss-mandated workflow automation with viral counter-tools; Salesforce reports Phase 2 revenue impact from AI agents

Key Actors

Salesforce

AI-native enterprise vanguard

Enterprise CRM deploying Agentforce at scale, reporting $100M savings and 3,200+ influenced opportunities

Koki Xu

Worker-side technical resistance

AI product manager and law graduate who built the anti-distillation tool on GitHub to sabotage workflow capture

Hancheng Cao

Academic labor-AI analyst

Emory University professor studying AI and work, argues firms gain structural knowledge advantage from workflow mapping

Daron Acemoglu

Skeptical productivity researcher

Nobel laureate economist whose 2024 MIT study found only 0.5% productivity gain over a decade from AI

Torsten Slok

Macro AI skeptic

Apollo chief economist invoking Solow's paradox to describe AI's absence in macro productivity data

What This Means

AI investment thesis faces a credibility stress test

Markets

With $250B invested in AI in 2024 and 90% of executives reporting no measurable productivity impact, markets are pricing a narrative that the macro data does not yet support. A J-curve delay is plausible but not guaranteed, and earnings call AI mentions are not a substitute for margin data.

Workflow codification becomes a new enterprise battleground

Tech

Tools like Colleague Skill and OpenClaw are normalizing the extraction of implicit employee knowledge into machine-readable blueprints. The technical arms race between enterprise workflow capture and worker anti-distillation tools will force platform vendors to add integrity verification and access control layers.

Vanguard case studies are raising the bar for AI ROI storytelling

Startups

Salesforce's $100M savings and revenue-from-sawdust narrative is setting investor and board expectations that most AI-adopting companies cannot currently meet. Startups selling enterprise AI tools will face intensifying pressure to demonstrate not just efficiency gains but pipeline and revenue attribution.

Sources

MIT Technology Review

5 days ago

Fortune

7 days ago

Fortune

2 months ago

winners

  • Deep-integration AI operators like Salesforce that use their own tools at enterprise scale and can redirect human capacity to revenue generation
  • Firms that map and codify employee workflows now, gaining structural knowledge assets regardless of near-term productivity outcomes
  • AI agent platform vendors whose tools are being mandated top-down across Chinese and Western enterprises

losers

  • Mid-tier enterprises that made large AI investments in 2024-2025 but cannot point to board-level revenue impact in 2026
  • Entry-level and mid-tier knowledge workers whose workflows are being codified into automatable blueprints
  • Workers in China facing top-down pressure to document and replicate their own roles, with limited legal protections over personality and judgment data

implications

  • HR and legal teams face new exposure: workflow capture tools that harvest personality traits, tone, and judgment may violate data privacy and labor laws in multiple jurisdictions
  • The Solow productivity paradox precedent suggests enterprise AI ROI may not materialize broadly until the early 2030s, compressing investor timelines
  • Worker counter-engineering — technically sabotaging AI training pipelines — will likely force enterprises to develop verification and integrity layers for AI training data

minority report

  • The flat productivity data may be a measurement artifact: if AI gains are concentrated in consumer-facing digital efficiency and informal task compression, standard GDP and employment surveys structurally cannot capture them yet
  • Worker resistance tools like anti-distillation may inadvertently accelerate enterprise AI maturity by forcing companies to build more robust, less personality-dependent automation systems
  • The 90% of executives reporting no AI impact may themselves be the laggards — selection bias in survey populations means the vanguard is systematically underrepresented in NBER-style studies

Level 4

Second-Order Shocks Ahead

The convergence of flat macro productivity data, organized worker resistance, and a small but compounding vanguard of AI-native firms is setting the stage for a structural divergence event in the enterprise landscape. As boards intensify pressure for AI to contribute to revenue rather than just cost reduction, firms without deep integration capabilities will face a credibility crisis. Worker counter-engineering will evolve from individual GitHub projects into institutionalized labor practices, potentially attracting regulatory and union attention in multiple jurisdictions. The next 24 months will determine whether the J-curve is real or a narrative convenience.

Key Points

  • Board pressure is shifting from AI cost reduction to AI revenue attribution, a threshold most enterprises cannot yet meet
  • The anti-distillation movement signals a coming formalization of worker AI rights as a labor relations category
  • Historical IT productivity lag precedent suggests a potential surge window of 2028 to 2032 if adoption patterns hold

Timeline

2025

Salesforce Phase 1: Agentforce handles 3M conversations, cuts support caseload 8%, generates $100M annualized savings

Early 2026

Salesforce Phase 2: AI agents deployed against dormant leads, influencing 3,200+ opportunities and generating closed revenue

Apr 2026

Anti-distillation tool goes viral in China with 5M likes; NBER study confirms 90% of executives report no AI productivity impact

Late 2026

Predicted: AI revenue attribution becomes standard board KPI; EU or China introduces worker data disclosure regulation

2028-2032

Projected window for broad AI productivity surge if J-curve holds, based on historical IT adoption lag precedent

Key Actors

Salesforce

AI-native enterprise vanguard

Demonstrated two-phase AI value creation from cost reduction to revenue generation, setting enterprise benchmark

Koki Xu

Worker-side technical resistance

Built anti-distillation tool that went viral with 5M likes, catalyzing a new worker-led technical resistance category

Erik Brynjolfsson

Optimistic productivity researcher

Stanford Digital Economy Lab director flagging potential reversal of productivity lag with Q4 GDP tracking at 3.7%

Daron Acemoglu

Skeptical productivity researcher

Nobel laureate finding only 0.5% AI productivity gain over a decade, anchoring skeptical macro forecast

Torsten Slok

Macro AI skeptic

Apollo chief economist proposing AI value creation follows a J-curve contingent on implementation quality

What This Means

Valuation divergence between AI vanguard and laggard firms is approaching

Markets

As AI revenue attribution becomes a board-level KPI, firms that cannot demonstrate the Salesforce model — cost savings plus new revenue from previously unaddressed segments — will face multiple compression. The 90% of executives reporting no impact are not yet penalized by markets, but that window is closing as investor sophistication on AI metrics increases.

Worker AI data rights are a regulatory gap about to be filled

Policy

The harvesting of employee personality traits, tone, and judgment patterns by tools like Colleague Skill has no clear legal framework in most jurisdictions. The EU AI Act, China's data security law, and US state-level privacy statutes are all candidates for rapid extension into this space, creating compliance exposure for any enterprise using workflow capture tools at scale.

New categories opening in AI training integrity and worker-side tooling

Startups

The anti-distillation movement signals demand for a class of tools that help workers manage their AI data exposure — and separately, enterprises will need audit and provenance layers to verify that their AI training data has not been corrupted. Both represent greenfield startup opportunities with clear enterprise and consumer buying motivations.

Detected Trends

AI Workflow Codification as Knowledge Asset Extraction

accelerating

Enterprises are systematically converting implicit employee knowledge — workflows, decision logic, personality traits — into machine-readable assets. This is happening faster in China due to top-down mandates but is spreading globally through tools like Claude Code and OpenClaw.

Technical Labor Resistance to AI Automation

emerging

Workers with technical skills are building counter-tools to subvert AI training pipelines. What began as a viral stunt in China is becoming a template for organized, engineering-driven resistance to workplace automation mandates.

AI Revenue Attribution Pressure

accelerating

Boards and investors are moving beyond cost reduction metrics and demanding that AI show up in revenue, pipeline, and growth data. Salesforce's sawdust model is setting a new standard that most enterprises are not yet equipped to meet.

Solow Paradox Redux

pending

Macro economists and executives are increasingly invoking the 1987 IT productivity paradox to describe AI's current absence in GDP, employment, and margin data. Whether a 1990s-style delayed surge follows depends on adoption quality, not just scale.

Sources

MIT Technology Review

5 days ago

Fortune

7 days ago

Fortune

2 months ago

second order

  • Enterprises mandating workflow codification will inadvertently create detailed internal knowledge graphs that become strategic assets — and acquisition targets — independent of AI productivity outcomes
  • As worker anti-distillation tools proliferate, enterprises will invest in AI training data auditing and provenance verification, spawning a new B2B security and compliance category
  • The two-tier firm split between AI vanguard and laggard will begin to show up in valuation multiples, customer retention data, and talent acquisition by late 2026 or early 2027

prediction

  • At least one major jurisdiction — likely the EU or a Chinese municipal government — will introduce regulations requiring employer disclosure when AI tools are trained on employee behavioral or personality data within 18 months
  • The Salesforce sawdust model — deploying AI agents against previously uneconomical market segments — will become a standard board-level AI ROI framework, replacing efficiency-only metrics by end of 2026
  • Worker anti-distillation tools will be adopted or replicated by at least one major labor union as a formal collective bargaining instrument within two years

minority report

  • The vanguard narrative around Salesforce-style AI ROI may itself be a form of enterprise marketing: $100M in cost savings and 3,200 influenced opportunities are outputs of a company selling AI products, creating a structural incentive to overstate results and set unreplicable benchmarks for the broader market
  • Worker resistance tools like anti-distillation may have negligible systemic effect — enterprises with sufficient data and compute can route around corrupted inputs, and the tools may only delay automation while reducing workers' perceived cooperativeness

Level 5

The Operator's Strategic Edge

The current AI enterprise moment is best understood as a knowledge extraction race disguised as a productivity initiative. Firms mandating workflow codification are not primarily optimizing for near-term automation — they are building proprietary training datasets, operational knowledge graphs, and decision-pattern archives that will compound in value regardless of whether current AI tools deliver on their productivity promise. The firms that win this decade will not be those that deployed AI earliest, but those that most systematically captured institutional knowledge before workers understood what was being taken. Worker resistance is rational and legally grounded, but technologically asymmetric. Operators who can see this dynamic clearly — and build the trust, governance, and incentive structures to align worker participation — will create durable moats. Those who simply mandate compliance will get corrupted training data, disengaged workers, and eventual regulatory exposure.

Key Points

  • Workflow codification is a knowledge asset extraction play, not just an automation play — the strategic value persists even if current AI tools underperform
  • The productivity paradox is a lagging indicator problem, not necessarily a technology failure — operators who act on J-curve logic now will be positioned for the surge window
  • Worker alignment is an enterprise AI moat — firms that capture authentic workflow data with consent and incentive will outcompete those relying on mandate and surveillance

Timeline

1987

Solow's productivity paradox first articulated: computer age visible everywhere except in productivity statistics

1995-2005

IT productivity surge materializes — 1.5% growth after decades of lag, validating J-curve logic for technology adoption cycles

2024-2025

Enterprise AI investment exceeds $250B; 374 S&P 500 companies mention AI positively in earnings calls; productivity data remains flat

Apr 2026

Vanguard firms show measurable AI revenue impact; worker resistance reaches technical and legal sophistication; NBER confirms paradox is live

2028-2032

Projected J-curve inflection window: firms with authentic workflow data assets and deep integration are positioned to compound gains

Key Actors

Salesforce

AI-native enterprise vanguard

Demonstrates the full arc from operational AI to revenue AI, using internal deployment as both product development and market proof

Koki Xu

Worker-side technical resistance

Crystallized the legal and ethical dimensions of AI workflow capture; her anti-distillation tool is a preview of institutionalized worker countermeasures

Hancheng Cao

Academic labor-AI analyst

Articulates the firm-side strategic logic of workflow mapping as knowledge asset creation, independent of automation outcomes

Erik Brynjolfsson

Optimistic productivity researcher

Provides the optimistic macro frame: Q4 GDP and productivity data may already signal the start of the J-curve inflection

Daron Acemoglu

Skeptical productivity researcher

Provides the skeptical anchor: even a 0.5% productivity gain over a decade is real but far below the narrative being sold to investors and boards

What This Means

AI valuation premiums will bifurcate on revenue attribution evidence

Markets

The market is currently pricing AI adoption narratives. As boards and analysts demand evidence of the sawdust model — AI generating revenue from previously unaddressed demand — firms that cannot produce it will face multiple compression. The vanguard cohort that can demonstrate compounding AI-driven revenue will command durable premiums. The divergence will be visible in earnings data by late 2026.

Worker AI data rights will become a major regulatory front within three years

Policy

The legal vacuum around AI tools that capture employee personality, tone, judgment, and behavioral patterns is not sustainable. The combination of viral worker resistance, cross-jurisdictional data privacy frameworks, and escalating labor-AI conflict will accelerate regulatory action. Enterprises using workflow capture tools at scale should begin building consent, disclosure, and data governance frameworks now, before compliance becomes mandatory.

Two high-conviction startup categories emerge from this event

Startups

First: AI training data integrity and provenance verification — enterprises need to know their workflow training data has not been corrupted by tools like anti-distillation. Second: worker-side AI data management — tools that help employees understand, control, and potentially monetize the workflow and behavioral data enterprises are extracting from them. Both categories have clear buyers, clear pain, and no dominant incumbent.

Detected Trends

AI Workflow Codification as Knowledge Asset Extraction

accelerating

The strategic value of mapping employee knowledge into machine-readable systems is compounding independently of near-term automation outcomes, making it a durable enterprise priority regardless of productivity headline data.

Technical Labor Resistance to AI Automation

emerging

Engineering-capable workers are building formal countermeasures to AI workflow capture, a development that will attract union, legal, and regulatory attention and reshape enterprise AI governance requirements.

AI Revenue Attribution Pressure

accelerating

Boards are demanding AI show up in growth metrics, not just efficiency. The sawdust model — using AI agents to pursue previously unaddressed revenue segments — is becoming the new benchmark for enterprise AI maturity.

Solow Paradox Redux

pending

Macro data continues to contradict enterprise AI narratives, but leading indicators from GDP and productivity research suggest a potential inflection. Whether it materializes depends on implementation depth, not deployment breadth.

Sources

MIT Technology Review

5 days ago

Fortune

7 days ago

Fortune

2 months ago

implications

  • Enterprises that treat workflow codification as a trust and compensation question — rather than a mandate — will produce higher-quality training data and face less regulatory and legal exposure when worker data rights frameworks arrive
  • The sawdust revenue model is the most replicable AI ROI framework available to enterprises today: identify previously uneconomical customer or market segments and deploy agents against them without cannibalizing existing human-led revenue
  • AI brain fry — productivity declining when workers use four or more AI tools — is an early warning that tool proliferation without workflow integration is a net negative, and enterprises should consolidate AI tooling around outcomes rather than adoption metrics

second order

  • If the J-curve thesis holds and a productivity surge materializes between 2028 and 2032, the firms that captured the richest workflow and decision-pattern datasets in 2025 and 2026 will have a compounding data moat that latecomers cannot replicate at any cost
  • The decoupling of job growth and GDP growth — already observed by El-Erian and Brynjolfsson — will intensify political pressure around AI labor policy, making the regulatory environment for workflow capture tools materially more restrictive within three to five years
  • IBM's decision to triple young hires despite AI automation capability reveals a hidden cost of over-automation: destroying the talent pipeline that produces future leadership, a risk most enterprise AI business cases do not model

minority report

  • The entire productivity paradox framing may be the wrong diagnostic: if AI is primarily compressing the cost of digital tasks — job hunting, scheduling, research, communication — and workers are reallocating that time to rest and social activity rather than more work, the outcome is a welfare gain, not a productivity failure, and GDP statistics are simply the wrong instrument to measure it
  • The Chinese anti-distillation movement may signal that the most AI-sophisticated workers — those capable of building counter-tools — are also the most valuable, and the firms aggressively pursuing workflow automation may be systematically destroying the exact judgment, creativity, and contextual reasoning that makes their workforce irreplaceable, producing short-term cost savings and long-term capability collapse