The primary operators of industrial machinery will no longer be human. They will be AI agents. A shift that began in software is now crossing into the physical world.
On the internet, the center of gravity has already moved from people to agents. Agents now make up more than half of all web traffic and write a substantial share of the production code deployed across modern software services.
As a16z notes, this is what the early stage of a platform shift looks like: real usage is surging, while the business models built on top of it are still taking shape. The web gave businesses an online storefront; mobile put that storefront in every consumer’s pocket. Now, as agents move past answering questions to completing end-to-end tasks, purchasing power is shifting from humans to software. Agents decide which tools to adopt, which services to subscribe to, and where to spend money. a16z calls this distribution layer—where agents emerge as the new first-class customer—the Business-to-Agent (B2A) economy.
What has played out in digital markets is a preview of what awaits heavy physical industry. Once an agent steps onto a factory floor, however, its job changes completely. Here, AI agents aren't customers buying services; they are operators running machinery. Alongside B2A, we are watching the dawn of Machine-to-Agent (M2A).
From advising on a screen to running the machine
Until now, nearly every industrial site was engineered around human eyes and human hands. Operations were held together by HMI screens, SCADA dashboards, alarm lists, shift handover notes, and the hard-won tacit knowledge of twenty-year veterans.
Conventional industrial AI could predict failures and flag anomalies, yet acting on those signals remained strictly a human job. Even when a system threw an alert, an engineer still had to walk the floor, trace the root cause, and get the line back up.
M2A flips that loop. AI agents plug directly into machine data, isolate what went wrong, decide the next move, and fire off control commands themselves. Humans step back from watching screens line by line, setting operational targets for the agent and verifying outcomes instead.
Built around platforms like Palantir’s AIP, early deployments already show agents moving past passive warnings into active execution. At Wendy’s, AI agents track real-time inventory and demand across stores nationwide, triggering stock reallocations and inventory adjustments the moment a supply imbalance appears. At United Airlines, agents synthesize vast maintenance logs with live aircraft sensor feeds to catch early signs of wear and schedule proactive part replacements before a failure grounds a flight. At Panasonic Energy, agents cross-check live equipment telemetry against historical fault records and technical manuals, pinpointing root causes and guiding technicians through recovery steps on the spot.
The shift goes beyond operational coordination. Agents are also beginning to take direct control of physical processes. In CoreWeave data centers, Phaidra’s AI agents watch power spikes build across server racks in real time and adjust coolant distribution setpoints before heat can spread. Imubit puts AI agents on Oxbow’s high-temperature calcining lines, forecasting thermal shifts and feedstock swings to write optimal operating parameters every minute. In commercial greenhouses, autonomous growing provider Blue Radix runs AI agents that continuously weigh indoor sensors, weather forecasts, and energy prices, sending heating, ventilation, and irrigation commands straight to the climate computer around the clock.
Why factories move slower than the web
Even as change takes hold on the shop floor, adoption moves far slower than it did on the internet. The underlying foundation is built differently.
On the web, AI agents have already moved beyond clicking through browser screens like humans, connecting directly to external tools through dedicated channels like APIs and MCP (Model Context Protocol). Speed comes easy because every endpoint speaks a common language.
Physical facilities work the opposite way. First, communication protocols are fragmented. A typical plant is a patchwork of machinery delivered by dozens of vendors over decades—a 1990s PLC sitting right next to a modern robotic cell, each speaking its own dialect like Modbus, Profinet, or OPC UA.
Second, the data itself arrives stripped of operational context. Machines spit out raw memory addresses or cryptic tags like DB10.DBD4, while the manuals and schematics defining what those numbers actually represent sit offline, disconnected from the data network. With data detached from context, an agent can read the numbers and still have no way to interpret them on the spot.
Third, timing and determinism play by different rules. Production machinery is deterministic: it must execute the exact same motion inside a fixed cycle with zero margin for error. AI agents take time to reason and produce probabilistic outputs, making it a non-starter to drop them straight into a high-speed real-time control loop.
Handing over control when mistakes have physical consequences
Clearing technical hurdles still leaves the question of authority. For decades, industrial control networks treated air-gapped isolation—shutting external software out completely—as the baseline for safety. Because process parameters are themselves trade secrets, plants hesitate to grant outside systems even read access.
Handing over write control demands far greater caution. Online, a bad agent call gets undone with a refund. On a factory floor, a bad call wrecks equipment or injures people. Full autonomous control in high-temperature, high-pressure, or heavy-machinery processes remains largely confined to pilot trials for this exact reason.
Bringing AI agents onto the floor safely therefore comes down to a hard split between read and write permissions, backed by real-time execution guardrails and hardware interlocks. Because an agent optimizing for a goal can look for ways around access limits, the system must inspect and validate every control command at every step before it executes.
In practice, an agent might be allowed to tune coolant flow or pressure inside a verified safe envelope, while any command that bypasses a safety interlock or triggers an emergency stop is blocked cold at the lower-level controller. Robotics cell recovery startup Dawn Industries initially required human engineers to approve every recovery plan its agent proposed, opening up automated execution only on workflows with proven safety boundaries. South Korea’s SK Innovation follows the same logic with its "Dual Brain" setup, pairing AI decision-making with human engineering sign-off.
Connecting agents without ripping and replacing legacy hardware
Making permission splits and safety controls work in practice starts with accepting the physical reality of existing plants. Machinery and controllers that have run for decades cannot simply be ripped out—the capital cost and downtime from halting a line are too steep.
M2A will therefore evolve not by replacing legacy equipment and lower-level controllers, but by placing an intermediate connectivity layer on top of them—absorbing diverse industrial protocols and translating them into standardized interfaces that AI agents can consume. For years, keeping outsiders out was a machine’s defining virtue; now, future competitiveness will hinge on how cleanly a machine or solution opens up standard interfaces so agents can read state and issue commands without friction.
Standardization around this middle layer is already taking concrete shape across multiple fronts. Anthropic introduced the Model Hardware Standard (MHS) to give AI agents one consistent spec for controlling physical hardware across manufacturers. Interoperability architectures bridging existing industrial standards and AI agents are moving forward in parallel: the OPC Foundation launched its "UA-for-AI" initiative to map OPC UA and industry-specific information models into formats built for MCP and AI workflows, laying down a standard semantic layer between agents and factory automation. Industrial IoT edge data platform Litmus shipped an official MCP server wiring heterogeneous shop-floor telemetry and edge device management straight to AI agents. Meanwhile, frontline operations platform Tulip argues that forcing every machine onto a single communication protocol is impossible—pointing instead to mapping inconsistent vendor tags, such as SV_01 and Target_Temp, into shared shop-floor concepts like "Target Temperature" or "Fault Alarm."
To hold up on a real factory floor, this connectivity layer must satisfy two requirements. First is vendor-agnostic openness: because a single plant runs equipment from FANUC, Siemens, Mitsubishi, and others side by side, the layer must ingest and normalize multi-vendor protocols inside one unified system. Second is building an industrial ontology that gives raw data structural and process context. Equipment hierarchies and causal relationships across processes—previously buried in manuals and schematics—must be modeled as structured data and tied to memory addresses (DB10.DBD4) and tags, letting an agent grasp the physical meaning of a reading and act safely.
Once this standardized foundation sits over existing equipment, plant operations grow sturdier over time. Machine run data, agent-resolved fault histories, and near-miss records all accumulate inside a single system of record. Just as corporate business processes accumulated inside ERP systems in the past, the tacit knowledge of veteran technicians and on-the-ground operating know-how will accumulate as data atop the M2A layer.
Operational know-how as data, and an early-stage market
Capturing operating know-how as data is becoming more urgent as reliance on individual skill hits its limit. As one reliability engineer put it, veterans are retiring while new hires are still learning the ropes, leaving a missing middle generation to bridge the two. In South Korea, the picture is no different: no matter how automated a line gets, unexpected disruptions still fall back on human intuition, yet passing that experience down is getting harder. Across many sites, M2A goes beyond boosting productivity—it turns personal judgment and troubleshooting know-how into durable system data that keeps operations steady.
Still, real-world adoption of AI agents on the factory floor remains at the starting line. According to the "2024 Smart Manufacturing Innovation Survey" by South Korea’s Ministry of SMEs and Startups, only 0.1% of roughly 160,000 small and mid-sized manufacturers have adopted manufacturing AI.
Just as standard APIs and MCP reshaped the internet into the B2A economy, physical industry will hit its true inflection point once scattered machine data and operational context are unified and wired into AI agents. Moving past the era of humans watching screens to run machines, the era of M2A—where AI agents run the floor—is arriving.