HYPERNOVELTYINSTITUTE

Technical intelligence
for accelerating change.

Search the record

Home/Analysis

Daily Hypernovelty Lead · AI systems · September 27, 2026

AI Is Changing Work Faster Than Firms Can Measure It

An ILO brief finds striking AI productivity claims in 21 Chinese firms, but most lacked a systematic way to measure the changes. The missing record includes review work, quality, and who benefits.

At an office desk, a baseline ledger sits beside a growing tray of work awaiting human review.

The International Labour Organization's September brief on AI adoption in Chinese enterprises puts striking performance numbers beside an awkward admission. One insurance operation said 300 employees increased daily customer-query throughput from 6,000 to 15,000. A recruitment team reported cutting its cycle from 30 days to 13. A manufacturing site reported a 30% improvement in production efficiency. Yet most of the firms studied had no systematic framework for measuring AI's impact.

This evidence comes from interviews with representatives of 21 Chinese enterprises and a separate survey of 1,591 Chinese professionals. The firms ranged from an eight-person startup to a conglomerate with 270,000 employees, and every interviewed company reported active AI deployment or concrete plans to adopt it.

But those details come with limits. Researchers selected the 21 firms deliberately rather than randomly, so the interview findings cannot represent all Chinese enterprises. Managers and department heads supplied the productivity figures, which were not independently verified. No firm described a controlled evaluation that could separate AI's effect from other changes happening at the same time. The reported metrics also measure different things, including throughput, cycle time, staffing, and cost, so they cannot be added up into one clean productivity result.

Even with that uncertainty, work was changing inside the firms studied. The brief identifies centralized AI teams, business-level integration, and bottom-up employee adoption as three common organizational patterns. Firms described AI handling customer queries, screening resumes, processing documents, generating code, and supporting production decisions. Most remained in hybrid arrangements where people retained judgment and oversight.

The result is a closer look at what human oversight contains. It can protect quality, but it can also hide labor that a throughput metric misses. When a system processes more cases, someone may still have to check weak outputs, resolve exceptions, correct errors, or explain a decision. The ILO brief does not establish that this hidden load increased across the sample. It shows that most firms were poorly equipped to find out, especially when measuring job quality, working conditions, and changes in the mix of tasks.

Meanwhile, the worker survey adds another layer without settling the argument. Fifty-six percent of respondents viewed AI adoption as inevitable, 47% believed it would create more jobs than it displaced, and 39% expected income declines. These are perceptions recorded at one point in time, rather than observed changes in jobs or pay. The firm interviews also did not directly include worker voices, leaving management accounts to carry most of the evidence about workplace change.

This creates a practical problem for any organization buying or building AI systems. A faster process can be real while the explanation for the gain remains weak. The organization still needs to know whether quality held, whether rework moved to another team, whether employees absorbed more review work, and whether the improvement lasted after the early rollout.

Verification bottleneck

  • Deployment moved faster than baseline measurement. Teams need comparable before-and-after records for volume, time, errors, rework, and escalations before they attribute a change to AI.
  • Management reports moved faster than independent checks. Reviewers need the underlying period, workflow, staffing level, and concurrent changes behind each productivity claim.
  • Hybrid work moved faster than job-quality measurement. Employers, workers, and researchers need evidence about correction load, task changes, working conditions, and where the gains went.
  • Watch the next evidence layer. Longitudinal studies, worker-level data, representative samples, and controlled comparisons would show whether the reported patterns survive beyond this group of firms.

Opportunities

So a practical service could help small and midsize organizations create an AI adoption evidence packet before a rollout. The packet would freeze a baseline, define the intended benefit, record the exact workflow and staffing model, and track errors, rework, escalations, and review time alongside output.

That same evidence packet could support a lightweight hybrid-work audit. It would ask which tasks disappeared, which tasks moved, who now checks the system, and whether faster throughput created a queue somewhere else. The value would come from making the labor around the tool visible, without pretending to prove causation from a simple before-and-after comparison.

Instead of offering a standard or policy requirement, the ILO brief documents a narrower problem: organizations are already reorganizing work around AI while many still lack the evidence needed to say what improved, why it improved, and who benefited.

Sources

Back to the archive