Skip to main content

Structured institutional evidence

AI Workforce Transition Tracker

See how AI is changing real work, how organizations are responding, and what the evidence actually supports.

For the conceptual framework behind these evidence cases, read the workforce transformation framework.

Dataset context

41 published evidence records

For workforce leaders, policymakers and researchers: compare documented changes in work and the limits of available evidence.

How we verify evidence

Current findings

What the reviewed evidence shows

Reviewed finding

AI is changing work before it clearly changes workforce size

Across the current evidence base, AI is already compressing, substituting and reorganizing specific tasks, while corresponding changes in workforce size are often not directly evidenced. Workflow change and headcount effects should be treated as separate questions.

7 supporting evidence cases

Explore evidence
Evidence boundary

This does not establish that AI generally has little employment impact, or that workforce reductions will not occur. It means only that within the current published evidence base, task and workflow changes are often more directly evidenced than net workforce effects.

Reviewed finding

Measured AI gains are clearest at the workflow level

The strongest quantified outcomes in the Tracker are tied to concrete work units—coding, review, planning, reporting and logistics. Institutions evaluating AI should therefore measure changes in specific workflows, not rely only on enterprise-wide adoption metrics.

6 supporting evidence cases

Explore evidence
Evidence boundary

These metrics come from different organizations, functions and measurement methods. They must not be aggregated into a common productivity rate or treated as an expected return from AI generally.

Reviewed finding

Human authority remains explicit in some high-stakes AI workflows

Several healthcare and public-sector cases combine AI delegation with explicit human control points. Where the evidence is strongest, humans retain confirmation, review or final-judgment authority.

5 supporting evidence cases

Explore evidence
Evidence boundary

Do not generalize this into a claim that human oversight is the norm across AI deployments. Human decision rights are not documented for many Tracker Signals.

Evidence explorer

Browse evidence cases

8 of 41 visible

Filter the evidence

Active

ZTE

China Communications equipment manufacturing

ZTE agent optimizes warehouse slotting at Nanjing Binjiang

Explicit outcome claims

AI mechanism → work change

AI agents execute multi-step work Workflow Redesign

How the organization responded

Not yet evidenced

What changed

The agent arranges frequently needed materials nearer picking positions overnight using the next day's production orders.

Key outcome

Reported picking travel distance — Company-reported before/after operational figure, not independently audited; the earlier Nanjing report describes a single robot but does not specify a period, so denominators cannot be reconciled. — Before more than 400,000 meters · After 160000 meters

Nanjing Iron and Steel

China Steel manufacturing

AI-assisted hazard reports connect classification, routing and safety response

Explicit outcome claims

AI mechanism → work change

AI-assisted work Workflow Redesign

How the organization responded

Not yet evidenced

What changed

Photo/voice reports feed AI classification and routing; staff remediate.

Key outcome

Reported fastest hazard closeout — Fastest-case testimony, not mean time or a controlled comparison — Before 2–3 working days · After about 1 hour

Changan Automobile

China Automotive

Engineers build an AI-assisted harness-selection workflow with technical verification

Explicit outcome claims

AI mechanism → work change

AI-assisted work Workflow Redesign

How the organization responded

Not yet evidenced

What changed

Mao Hai formalized harness rules and used AI to build an executable tool; structured inputs replace manual calculation, with engineer checking.

Key outcome

Reported harness-selection calculation time — Source-reported single-workflow comparison; elapsed/work time distinction unspecified, no causal control — Before 2880 minutes · After 5 minutes

Beijing Tiantan Hospital

China Healthcare

AI-generated radiology report draft with doctor review

Explicit outcome claims

AI mechanism → work change

Automation Workflow Redesign

How the organization responded

Not yet evidenced

What changed

AI generates an initial head-CT report draft that is then reviewed and finalized by a doctor.

Key outcome

AI report-generation time — Report generated in less than one minute

AdventHealth

United States Healthcare

Clinical and utilization-management workflow augmentation

Explicit outcome claims

AI mechanism → work change

AI-assisted work Workflow Redesign

How the organization responded

Not yet evidenced

What changed

AI assists documentation and utilization-review work while clinicians retain final judgment.

Key outcome

Utilization-review turnaround — Teams reported faster turnaround from a baseline of about 10 minutes per case

Qunar

China Travel & hospitality technology

AI-mediated customer-service training and QA

Explicit outcome claims

AI mechanism → work change

AI-assisted work Workflow Redesign

How the organization responded

Entry-Path Redesign

What changed

AI now mediates customer-service onboarding, live response support and automated quality assurance.

Key outcome

New-agent training duration — Training duration reduced from 15 days to 5 days — Before 15 days · After 5 days

Deutsche Telekom

Germany Telecommunications

AI agents monitor mobile networks and execute corrective actions in minutes

Explicit outcome claims

AI mechanism → work change

AI agents execute multi-step work Workflow Redesign

How the organization responded

Not yet evidenced

What changed

A multi-agent system now identifies network events, analyzes load and anomalies, recommends optimization actions and can autonomously execute corrective network changes that previously required manual staff analysis and implementation.

Key outcome

Selected network-management process duration — Activities previously taking employees around one hour can be completed in a few minutes

Futian District Government, Shenzhen

China Public sector & justice

AI digital employees draft, review and route administrative work

Explicit outcome claims

AI mechanism → work change

AI agents execute multi-step work Workflow Redesign

How the organization responded

Operating Model Redesign

What changed

Multiple administrative workflows now delegate drafting, review, routing or coordination work to AI digital employees under a human-guardian model.

Key outcome

Document-format correction accuracy — Reported as greater than 95%

Evidence landscape

Each case is counted once by its primary work change. Evidence stages describe what sources report, not independent proof of causation.

Work changeIn operationObserved work changeExplicit outcome claims
Task CompressionAI reduces the time or effort required within an existing task.0 cases2 cases5 cases
Task SubstitutionAI performs a task that was previously performed by people.1 cases1 cases4 cases
Workflow RedesignThe sequence, handoffs, or coordination of work changes around AI.3 cases6 cases8 cases
Role RedesignThe responsibilities or boundaries of a role change as work is reorganized.1 cases3 cases7 cases

Methodology and updates

The Tracker reviews public sources and distinguishes operational evidence, observed work change and reported outcome claims. Claims are not necessarily independently verified or caused by AI. Records are reviewed periodically as evidence becomes available.

Read methodology v1.0