The Models Are Ready. The Factory Floor Isn't.

Kimberly Tan at a16z published something this week that deserves attention from anyone building or buying industrial AI. “AI Adoption by the Numbers” compiles hard data on where enterprise AI is actually working, and where it is not. The headline numbers are striking: 29% of the Fortune 500 are live, paying customers of an AI startup. Revenue is real. Adoption is real. But the distribution tells a more important story.

Nearly all the revenue is in coding, legal, customer support, and search. The industries adopting fastest are tech, legal, and healthcare. Manufacturing and industrial operations are nowhere on the chart.

That is not because the models are not good enough for industrial work. The a16z data proves the opposite.

The Capability Gap Is Closing. The Adoption Gap Is Not.

a16z overlays enterprise AI revenue against GDPval, a benchmark from OpenAI that measures how well models perform economically valuable tasks compared to human experts. The picture is revealing.

Industrial engineers went from a 17% win rate against human experts to 44% in six months. Mechanical engineers went from 25% to 44%. These are among the fastest capability improvements across every occupation tracked. The models are getting better at industrial reasoning faster than almost anything else.

But there is no corresponding revenue breakout. No billion-dollar industrial AI startup. No factory-floor equivalent of Cursor or Harvey or Glean.

The models can do the work. Something else is stopping them from doing it at scale in the physical world.

What a16z Says Is Blocking Adoption

a16z identifies four properties that make an industry easy for AI to enter: text-based work, rote and repetitive tasks, a natural human-in-the-loop workflow, and clearly verifiable outputs. Code runs or it does not. A support ticket is resolved or it is not. A legal brief is drafted or it is not. There is a tight feedback loop and a low cost of failure.

Industrial operations have none of these properties. The work interacts with the physical world. It involves multiple stakeholders across shifts, departments, and geographies. It operates under heavy regulatory and compliance requirements. And the outputs are difficult to verify after the fact, because the decision happened at 2 AM on a plant floor and nobody recorded the reasoning.

These are not implementation details. They are structural barriers. And they explain the gap between what the models can do and what enterprises are willing to let them do in a refinery, a mine, or a processing plant.

The Barrier Is Not Intelligence. It Is Infrastructure.

This is the insight buried in the a16z data that most people will miss.

When AI adoption is easy, the model capability is sufficient on its own. You give a developer Cursor, they write better code, the output is testable, and the organisation can verify ROI in a week. The deployment infrastructure barely matters because the feedback loop handles it.

When AI adoption is hard (physical world, multi-stakeholder, regulated, verification-difficult), the model capability is necessary but not sufficient. You can have a model that reasons as well as an industrial engineer. It still cannot act in your plant without governance architecture that makes its decisions auditable. It still cannot coordinate across your maintenance, production, and safety systems without an orchestration layer. It still cannot operate on the factory floor at 2 AM without edge deployment that works without cloud connectivity.

The missing piece is not a better model. It is the deployment infrastructure that makes a capable model safe, governed, and verifiable in environments where the consequences of a wrong answer are measured in equipment damage, safety incidents, and regulatory exposure.

What I See in Every Customer Conversation

I see this gap in practice every week. The operations teams I work with are not waiting for better AI. They have access to the same foundation models everyone else does. What they do not have is a way to deploy those models into their operational environment with the governance, connectivity, and auditability that their safety managers, regulators, and boards require.

The deployment barriers a16z identifies from their enterprise data are the same barriers I encounter on every project:

Physical-world interaction requires edge-native deployment that works inside air-gapped environments, on constrained hardware, without cloud dependency.
Multi-stakeholder coordination requires multi-agent orchestration, where teams of AI agents coordinate across systems in real time, not a single copilot answering questions.
Regulatory compliance requires governance architecture that evaluates agent actions against engineering constraints before execution, not after.
Verification difficulty requires decision traces that capture not just what an agent did, but why it did it, what information it considered, and what alternatives it rejected.

None of this is model capability. All of it is deployment infrastructure.

The Next Wave Is Already Visible

a16z closes their analysis by pointing to domains where model capabilities are improving fast but revenue has not broken out yet. They flag long-horizon agents, computer use on legacy systems, and modalities beyond text as the enabling capabilities arriving now.

Industrial operations sits squarely in that description. The models are reaching production-viable capability for industrial reasoning. The enabling infrastructure (agents that run continuously, connect to legacy OT systems, and operate beyond text in the physical world) is what unlocks the revenue.

The first wave of enterprise AI captured the easy cases: text-based, single-user, verifiable, low-risk. The next wave captures the hard cases: physical, multi-system, regulated, consequence-heavy.

That is where the value is largest. It is also where the deployment barriers are highest. Whoever builds the infrastructure to clear those barriers captures the wave.