For decades, information technology and operational technology have largely occupied different corners of the enterprise. 

IT has powered the digital side of business i.e. applications, cloud platforms, data centres, networks and cybersecurity. Operational technology, or OT, has powered the physical side i.e. factory machinery, industrial robots, sensors, laboratory instruments and production lines. 

The two worlds have evolved with different priorities. IT is built for scale, connectivity and rapid change. OT is built for safety, reliability and uninterrupted operation. A software error may disrupt a workflow; a faulty command to a machine can stop a production line, damage equipment or put people at risk. 

Now, artificial intelligence is beginning to bring these worlds together. 

Anthropic’s recent research preview of the Model Hardware Standard, or MHS, points to this direction. The proposed framework is designed to give AI agents a common way to discover, understand, monitor and operate programmable physical devices. Instead of creating a custom integration for every robot, sensor or laboratory instrument, the aim is to establish a shared interface through which AI systems can work with hardware safely. 

The development is significant because it addresses a stubborn problem that has limited automation for years: getting different machines and systems to work together. 

The challenge of disconnected systems 

Modern factories and laboratories are rarely built around a single technology provider. A production line may include programmable controllers, cameras, conveyors, robotic arms, quality-inspection stations and supervisory software from several vendors. Each may use its own protocols, software and data formats. 

Some devices have modern application programming interfaces. Others rely on older software, proprietary tools or interfaces that require an operator to work through a screen manually. 

This has created what many industrial teams describe as “islands of automation”. Machines may be automated individually, but people are still needed to move information between systems, resolve exceptions, interpret logs and make decisions when something goes wrong. 

Anthropic says integrating hardware often takes weeks or months because devices do not naturally communicate with each other. MHS seeks to reduce that effort by introducing a common driver model. In simple terms, it gives a device a standard way to describe what it can do, what state it is in and what limits should govern its operation. 

That matters because the real challenge is often not controlling one machine. It is coordinating several machines, software platforms and human decisions across a complete process. 

The principle  

The concept is easier to understand through examples most people already know. 

The Java Virtual Machine, or JVM, offers a related example from software. Java applications can run on different operating systems because the JVM provides a common runtime layer between the application and the underlying platform. 

MHS applies a similar principle to physical operations. It does not make every machine identical. Instead, it aims to give AI agents a controlled and consistent context to work with different kinds of equipment while keeping the device-specific complexity below the surface. 

For an AI system, that could mean reading a temperature, checking whether a machine is available, adjusting an approved parameter or launching an authorised procedure. The machine remains governed by its own capabilities and safety limits, but the AI can work through a common interface rather than a different bespoke connection every time. 

AI as an orchestrator, not an unchecked operator  

The prospect of AI operating physical equipment can sound unsettling. It should. 

No responsible organisation should hand over emergency controls, safety interlocks or millisecond-level machine decisions to a general-purpose language model. Those controls need to remain deterministic, tested and dependable. 

But there is a layer above direct machine control where AI could be genuinely useful. 

Consider a supervisor responding to an equipment alarm. They may need to examine sensor readings, maintenance history, quality data, inventory levels and standard operating procedures before deciding whether to pause a line, call an engineer or continue production. 

Today, much of this coordination happens in a person’s head. 

An AI agent could help by pulling together the relevant information, checking what actions are permitted, recommending the next step and documenting the decision. In limited, well-defined circumstances, it could also execute approved actions. 

Anthropic’s early MHS trials offer a glimpse of this possibility. In one laboratory workflow, an AI agent coordinated a liquid handler, a robotic arm and a plate reader. The important development was not merely that the system could issue commands. It could observe results, plan the next step and coordinate a multi-device process that would otherwise require a person to connect the dots between separate tools. 

This is where IT, OT and AI begin to converge. 

IT provides connectivity, data, identity management, cybersecurity and governance. OT provides the equipment, sensor data and physical constraints. AI can serve as the reasoning and orchestration layer between them. 

Safety must come first 

The risks are clear. A wrong answer from a chatbot may be frustrating. A wrong command sent to a robot or industrial process could lead to a safety incident, damaged equipment or a spoiled batch of products. 

That is why the interface layer is as important as the AI model. 

A robust AI-to-hardware system needs to communicate more than commands. It must understand the state of a device, the actions that are allowed, the actions that need human approval and the safeguards that must never be bypassed. 

This requires strong identity controls, limited access rights, separation between IT and OT networks, detailed command logs, device authentication and reliable emergency-stop processes. These are not features to add after an AI pilot succeeds. They are the foundation that makes a pilot safe enough to attempt. 

Anthropic has positioned MHS as a research preview and says it is working with partners on safety evaluations before a wider release. That caution is sensible. A successful demonstration in a controlled lab does not automatically mean a system is ready for a busy factory floor, where communication failures, sensor errors and unexpected conditions are unavoidable. 

Start with practical use cases 

For most companies, the answer is not to pursue a fully autonomous factory from day one. A more sensible route is to begin with a narrowly defined process that involves disconnected systems, considerable human coordination and clear safety boundaries.  

The first steps are practical: 

• Maintain a clear inventory of connected assets, interfaces, owners and safety constraints. 

• Prioritize equipment with documented, secure and supportable interfaces. 

• Bring IT, OT, cybersecurity, data and business teams into one operating model. 

• Define exactly what an AI system can recommend, what it can do independently and what must always be approved by a human. 

• Measure results in operational terms: reduced downtime, faster cycle times and fewer manual handoffs. 

The future of enterprise AI will extend well beyond chatbots and back-office workflows. It will increasingly reach the systems that make, move, test and inspect physical things. 

The promise of standards such as MHS is not that they remove complexity. It is that they place complexity in a reliable layer, making it easier and safer for people and systems above that layer to work together.