When 20% of US Workers Delegate to AI, Governance Becomes an HR Problem

July 2026: a representative Epoch AI/Ipsos survey of 1,106 employed US adults found that one in five now hand at least one work task to an AI model that used to belong to a coworker or contractor. Software developers lead at 57%. Data analysts follow at 46%. The delegation is real. The org chart has not caught up.

The productivity story is thinner than the headline

Time savings show up in 53% of AI-heavy tasks. But roughly one in six takes longer than the human version. That is not a rounding error — it is a distribution with a long tail of failed handoffs, silent retries, and rework nobody logs. The U.S. Bureau of Labor Statistics still measures output per hour worked; it does not measure output per prompt refined. Executive dashboards inherit the same blind spot.

The mid-market pattern we see across CAI consulting engagements: individual contributors quietly route drafts, code reviews, and research through AI. Managers assume the person did the work. Auditors later cannot reconstruct who — or what — produced the artifact. That gap is where liability accumulates under the EU AI Act’s transparency obligations for GPAI-integrated workflows, and under ENISA’s guidance on AI supply-chain risk.

What executives should change this quarter

Three concrete moves, ordered by ROI:

# ai_delegation_policy.v3.yaml
task_disclosure:
  required_for: [client_deliverable, code_merge, financial_analysis]
  audit_retention_days: 730       # aligns with NIS2 Art. 21 evidence
provenance:
  log_model: true
  log_prompt_hash: true           # not raw prompt (PII risk)
  log_reviewer_id: true
rework_metric:
  threshold_hours_per_week: 4
  escalate_to: line_manager

First, make delegation visible. Not banned — visible. A one-line disclosure field on every deliverable, logged centrally, satisfies both internal QA and the audit trail regulators will request under NIS2 Article 21. Second, measure the rework tail, not the average. If your team celebrates the 53% who save time and ignores the one-in-six who lose time, you are optimizing the wrong distribution — a pattern our LLM agent failure taxonomy tracks in production systems. Third, route heavy inference through a governed layer with cost-aware model routing, so finance sees per-team spend before it becomes shadow IT.

CAI’s position

Delegation to AI is not a tooling decision. It is a labor and evidence question that HR, legal, and CISO have to co-own before the next audit — and long before the AI Act’s 2027 enforcement window closes. Most consultancies will sell you a training program. We prefer to install the disclosure plumbing first, then let training follow the data it produces. If your organization is past the pilot and into shadow adoption, talk to us about a four-week governance sprint.

Read further

• How the EU AI Act Omnibus tightens transparency in six days
• Six failure clusters every LLM agent team should track
• The 2027 AI Act delay and what compliance teams do now