Only 12% of Manufacturers Fully Integrate AI: Inside the OT/IT Divide

1 Οκτώβριος 2026

Why it matters now: Manufacturers are racing to attach artificial intelligence to production, yet a new analysis published on 1 October 2026 finds that only 12% of AI-using manufacturers have fully integrated AI into their business systems. The barrier is structural rather than algorithmic: legacy automation, fragmented data and the persistent OT/IT divide — the gulf between operational technology such as PLCs and SCADA, and enterprise information technology — continue to stall the smart factory.

The timing is uncomfortable. AI budgets are expanding, edge computing has matured, and controller vendors are shipping hardware that speaks cloud-native protocols. The control layer, however, remains the least integrated element of the enterprise stack.

The 12% Reality: AI Reaches the Plant Floor, Rarely the Enterprise

Adoption and integration are not the same achievement. Most manufacturers have now piloted AI somewhere — visual quality inspection, maintenance forecasting, scheduling optimisation. Very few have wired those models into the transactional systems that actually run the business.

That distinction matters because integration is where the returns sit. An unconnected model produces a dashboard that someone must interpret. An integrated model adjusts a setpoint, raises a work order, or alters a purchase commitment without waiting for human translation.

Analyst Insight: The 12% figure measures plumbing, not enthusiasm. Manufacturers have largely bought the intelligence; they have not yet built the pipes that carry data out of the controller and decisions back into it. Until those pipes exist, AI remains a reporting tool rather than a control strategy.

The OT/IT Divide: Where the Smart Factory Stalls

OT and IT were engineered around opposing priorities. OT privileges determinism, uptime and machine safety across decades-long asset lifecycles. IT privileges scalability, agility and open standards on refresh cycles measured in months.

When those worlds meet, friction is inevitable — and it is both technical and organisational. Plant engineers guard availability; enterprise IT guards the network perimeter and the integrity of corporate data.

Three obstacles recur in the report: legacy automation systems with proprietary interfaces, fragmented and ungoverned production data, and the cultural wall between plant engineering and IT teams.

Data Snapshot: The Recurring Barriers to AI Integration

Barrier	Why It Blocks Full AI Integration
Legacy automation systems	Proprietary protocols and closed engineering environments prevent contextualised data from leaving the controller.
Fragmented data	Tag names, units and asset models differ per line and per vendor, so downstream models cannot be trained at scale.
OT/IT organisational divide	Differing ownership of uptime versus network security slows approval of any new data path.
Security governance	Inbound connectivity into the plant network is widely restricted, limiting closed-loop AI use cases.
Talent and skills gaps	Fewer engineers straddle IEC 61131-3 programming and modern data engineering practices.

PLCs Sit at the Centre of the Integration Battle

The programmable logic controller is simultaneously the richest source of operational data and the hardest asset to integrate. It holds cycle times, fault codes, torque curves and process values that AI models need — yet that data is frequently trapped behind vendor-specific interfaces and legacy serial links.

This is why the controller has become the front line of the convergence debate. Whoever owns the path between the PLC and the cloud owns a disproportionate share of the smart factory value chain.

From Machine Controller to Edge Data Node

The industry's answer has been the edge-capable controller: hardware that preserves deterministic local control while streaming contextualised data upward. Containerised applications, native MQTT publishing, buffered store-and-forward during network outages and physical digital-I/O tapping of brownfield machines are now shipping features rather than road-map promises.

Open Protocols as Competitive Terrain

OPC UA has become the de facto bridge across the divide. Its companion specifications move semantic models — not merely raw tags — from plant to enterprise, while certificate-based, outbound-only connectivity satisfies IT governance requirements.

The strategic consequence is straightforward: vendors that confine data inside closed ecosystems increasingly find themselves excluded from enterprise data architectures altogether.

Market Trend: Interoperability is shifting from a marketing claim to a procurement criterion. As AI programmes move from pilot to production, controller selection is being evaluated on data egress capability, protocol openness and lifecycle cybersecurity — not on scan time and I/O count alone.

Frequently Asked Questions

Why has only 12% of AI-using manufacturers fully integrated AI?

Because integration requires data to move reliably between the control layer and business systems, and that path is blocked by legacy equipment, inconsistent data models and the security and organisational split between OT and IT teams. Piloting AI on an isolated data set is comparably easy; industrialising it across a plant estate is not.

What exactly is the OT/IT divide?

It is the separation between operational technology — PLCs, SCADA, drives and machine control, where real-time behaviour and safety dominate — and information technology, where servers, cloud platforms, MES and ERP systems reside. The divide is technical, in that protocols and networks differ, but it is equally organisational, in that different teams own each domain.

Can legacy PLCs participate in a smart factory architecture?

Yes, in most cases. Existing controllers can be connected through protocol-converting gateways that translate Modbus, Profibus or proprietary serial data into OPC UA or MQTT, or through digital-I/O taps at the signal layer where no network interface exists. The practical constraint is designing those links as outbound-only and buffered, so plant availability is never compromised by an IT-side outage.

Why is OPC UA so central to AI integration?

OPC UA provides more than transport. It carries information models that give raw values meaning and structure, which is what machine learning pipelines require in order to generalise across assets. Combined with established security mechanisms, it allows the same data set to serve plant operations and enterprise analytics without duplicating infrastructure.

What should buyers prioritise when selecting a controller today?

Deterministic control performance remains non-negotiable, but it is no longer sufficient. Buyers should assess native support for open protocols, edge compute capacity for containerised workloads, data buffering during connectivity loss, long-term firmware and security support, and the cost of licensing any software needed to expose controller data.

The Analyst Outlook: Convergence Becomes a Procurement Criterion

The 12% figure is best understood as a lag indicator. It describes where integration stands now, not where the market is heading. The technical building blocks for closing the gap — edge controllers, open protocols, semantic data models and outbound-only secure connectivity — are already commercially available.

What remains is execution: treating the OT/IT divide as an engineering problem to be designed around rather than a boundary to be defended. Manufacturers that standardise data egress at the controller level will find that AI integration stops being a special project and becomes an architectural default.

For automation vendors, the competitive logic is equally clear. Controllers that function as isolated islands will be specified less often, however strong their control performance. In the next procurement cycle, the decisive question will not be how fast a PLC scans, but how cleanly its data crosses the line.