Agentic dark economy: forecast to 2035

In the study, agents run whole departments before whole companies. It gives mine haulage 80% odds of material scale by 2030 and a fully autonomous company 18%.

The numbers

  • $230B

    agent operations software and services in 2030, roughly, in the base case

    A modeled scenario, range about $90B to $410B

  • 88%

    of surveyed organizations used AI in 2025

    Stanford AI Index 2026. Agent deployment stayed in single digits in nearly all functions

  • 12%

    success on real household robot tasks

    Stanford AI Index

Show all figures
  • $1.7B

    equity round for one industrial AI group

    Funding does not validate unit economics

Odds of material scale by 2030

  • odds that dark mine and quarry haulage reaches material scale by 203080%
  • odds for food production hubs60%
  • odds for a fully autonomous company, held back by law, banking, liability and trust18%

The AI’s judgments, not measurements. Material scale means at least 10% production use among eligible organizations, or more than $10B a year in global spend.

How sure the study is

  • The 2030 rankingmedium confidence
  • The 2035 penetration and spending numberslow confidence

I wanted to know where “dark” operations, work with the lights off and agents doing the steps, will show up first. I asked an AI to build a forecast from adoption, reliability, regulation and security evidence, and to separate hard evidence from vendor claims.

These scenarios are a way to think through what might happen. The AI estimated the odds from the sources below; I have no evidence that those odds are calibrated. The spending range comes from a simple calculation: take the published size of the software and IT services market, assume a growth rate, then estimate the share spent on agents.

What I found

  • The study puts dark departments first, dark industrial cells second and dark companies last. Software operations, back office, customer operations and research lead. A fully autonomous company gets 18% odds of reaching material scale by 2030, held back by law, banking, liability and trust. Material scale means at least 10% production use among eligible organizations or more than $10B in yearly global spend.
  • Capability is running ahead of deployment. The 2026 Stanford AI Index reports AI use at 88% of surveyed organizations in 2025, with agent deployment in the single digits across nearly all functions. McKinsey’s 2025 survey found no single business function above 10% of organizations scaling agents. Leading agents reached 66.3% accuracy on OSWorld, a benchmark of computer tasks across operating systems, and Stanford notes they still fail roughly one attempt in three on structured benchmarks.
  • In the study, physical autonomy wins where the environment is constrained. The International Federation of Robotics counts 4.66 million industrial robots in operation in 2024, and by the study’s reading most of them sit on mapped, controlled sites. The Stanford AI Index reports 12% success on real household robot tasks. On that basis the AI gives dark mine and quarry haulage 80% odds of material scale by 2030 and food production hubs 60%.
  • In the study’s base case, the software and services stack for agent operations reaches about $230B in 2030, with a range of about $90B to $410B. It excludes robot hardware, land and other physical capital, so it is not the same market as industrial AI operators.
  • The study examines the case of one large industrial AI group. Its $1.7B equity round, confirmed by the company and by TechCrunch on 22 July 2026, shows investors will finance the thesis, but funding does not validate unit economics. The study treats the group’s food robotics results as vendor-reported, and on the company site its transport business was still marked “announcement soon” in November 2026.
  • The operating model the study expects to last is exception-based autonomy: bounded authority, safety checks, independent verification, a durable receipt, an exception queue and a named accountable human. In that model, what gets measured is human minutes and recovery cost per finished outcome, not demo completion.
  • The study rates its 2030 ranking at medium confidence and its 2035 penetration and spending numbers at low confidence.

How it was made

The study is dated 3 August 2026 and rests on 18 public sources, among them the Stanford AI Index, McKinsey, METR, the ILO, NIST and the International Federation of Robotics, plus company disclosures. The AI gathered the sources, built the scenario tables and drafted the text. I asked the question. Figures were current at that date. Survey adoption is self-reported.

Sources

The question
Which parts of business and physical work will be run by AI agents with people only handling exceptions, and by when?
How it was made
AI forecast study built on 18 public sources.