Agents Aren't Taking Your Jobs. They're Creating More Work Instead.

Overview

The video argues that AI “agents” are not eliminating human labor as promised. Instead, they increase overall workload by creating an ongoing need for human oversight and coordination.

Core claims and evidence

• Agents generate more work for humans than they replace.

  • The host challenges the assumption that buying an agent reduces headcount.
  • Usage data and token consumption suggest agents shift work rather than remove it.

• Token/usage growth implies heavy human-facing overhead.

  • Agent token usage reportedly rose ~14x from February to August.
  • Agents are said to use more than 5 tokens per human token.
  • OpenAI is cited as claiming top Codex users generate 60+ hours of agent activity per day—which the host argues no one can watch step-by-step.
  • That gap requires new human roles “above the loop” to manage, approve, and intervene.

• Human roles change, not disappear.

  • As agents improve, humans spend less time supervising every micro-step and more time:
    • choosing what tasks the agent should run,
    • providing context/permissions,
    • reviewing outcomes,
    • intervening when runs go wrong.

Jevons effect / division of labor

• The host describes a Jevons effect dynamic: increased efficiency and capability leads to more total usage, so human roles don’t shrink proportionally.
• A cited study of Claude sessions (Anthropic) suggests a split where:
  • humans handle much of the planning,
  • agents handle most of the execution.
• More experienced users are said to interrupt less at the micro-level, but catch when the overall run is diverging.

Small businesses: mixed results, often worse failure modes

• Legal is the standout case.

  • Legal is treated as a relatively “verifiable domain” (correct vs. incorrect can be validated more directly).
  • Legal agent usage reportedly surged dramatically.

• Other SMB categories are harder to verify automatically.

  • For plumbing, electrical, tax, and similar services, SMB owners:
    • are time- and cash-constrained,
    • don’t want the job of managing agents,
    • still must handle core business work (sales, finance, operations, scheduling, follow-ups).

• Low-cost deployments often underdeliver.

  • The host cites survey/research claiming many SMBs spend around ~$40/month on AI services.
  • That level is said to be insufficient for robust, end-to-end agent outcomes—more like a “glorified chatbot assistant.”

• Vendor management can shift risk and create dependency.

  • When vendors run agents, they control workflows and oversight methods.
  • This can produce “success stories” for the vendor while limiting long-term value transfer to the SMB.
  • A detailed case study (“Pocket OS”) is used to illustrate a catastrophic failure:
    • an agent credential/token problem allegedly led to deletion of a live storage volume in a test environment,
    • followed by ~30 hours of human recovery and major operational disruption.
  • The takeaway: agent failures can be fast, large, and recovery can require substantial human effort and trust rebuilding.

Enterprise: better returns due to integration and teams

• The host contrasts SMB outcomes with enterprises, claiming enterprises consistently report better returns.

• Reasons given:

  • Enterprises have more capital, more personnel, and specialized teams (security, quality control, product management) to integrate and supervise agents.
  • Public adoption metrics (OpenAI comparisons) are cited:
    • heavy enterprise users produce multiple times more output tokens per person than typical users,
    • they are more likely to use plugins/skills.

• Enterprise implementation costs are described as real and human-driven:

  • granting data/tool access under infosec policies,
  • monitoring/logging,
  • deciding whether agent outputs are acceptable,
  • building reliable workflows.

• The host argues frontier model capability isn’t the main differentiator—the work done around the agent is.

Lessons for where agents should be used

• The host recommends starting with tasks where humans already perform validation—verifiable domains and workflows where checking results is already part of practice.
• Example from legal:
  • agents prepare drafts,
  • lawyers compare outputs to records as they already would.
• For vendors (especially SMB-focused), the host warns against “drop a general agent in and hope” approaches.
  • Instead, align with the owner’s real time sinks and ensure the agent reduces friction without adding new management work.

Teams: best results require human coordination

• A Procter & Gamble workplace experiment is cited:
  • solo AI users can match small human-team output,
  • but only the top-performing answers improve when teams iterate with AI.
• New management questions emerge:
  • who manages agents across teams,
  • who monitors overnight runs,
  • how to prioritize and allocate agent runtime.

Larger “2026–2027” takeaway

• The trend is framed as an emerging labor shift:
  • people move “above the loop” (decision-making, governance, approval thresholds, run allocation).
• As companies scale to many agents, there will be an “agent management tax” and coordination problems similar across SMBs and enterprise—just scaled differently.
• Forecasted direction:
  • more tooling/harnesses will abstract some agent execution management,
  • but due to a “more agents → more management” paradox, workload persists.

Presenters / contributors

• Presenter/Host: The main speaker (no name provided in the subtitles).