← All updates
Industrial Technology

ISA 95 Has Not Been Replaced. AI Just Added Layers It Never Described.

An analyst proposal adds new layers to the industrial AI conversation. Here is what it means for plant systems and asset data.

Adil Najeeb, CMRP, CRL ·

Summary

A widely shared post this week announced that the ISA 95 pyramid is getting new layers, with an eight level industrial AI stack drawn on top of the familiar five. The diagram is useful. The headline needs one correction before it travels any further: ISA has not revised ISA 95. The stack is a proposal from IoT Analytics, built from interviews at IMTS 2026, and it sits on top of the standard rather than replacing it. For plant leaders, the real story is not the new layers at the top. It is the layer in the middle that every one of those layers depends on: your systems of record, and the asset data inside them.

1. What Happened?

IoT Analytics published its IMTS 2026 recap on October 19, 2026, under the title "A New Industrial AI Tech Stack Is Forming." IMTS ran in Chicago from September 14 to 19 and drew 91,295 registrants. Working from more than 30 interviews at the show, the firm argues that industrial AI is settling into an eight layer stack. Machines and production lines remain the base. Controllers are becoming a software defined edge. A new layer called Industrial DataOps handles connectivity and gives raw machine data its context. Above that sit the systems of record: ERP, MES, PLM, CMMS, and the historian. Then comes a layer ISA 95 never described, a knowledge graph and ontology that models assets, processes, and the relationships between them. Interfaces such as MCP connect agents to software, to other agents, and to machines. Agents plan and carry out work across those systems. At the top, people review and approve.

The proposal is already visible in product announcements. At the show, PTC, Rockwell Automation, Limble, MachineMetrics, and Vention all presented MCP servers. AWS described an orchestrator that queries MCP servers for each plant system and assembles a knowledge graph on demand. This week the pattern continued: Rockwell introduced FactoryTalk Conductor, an orchestration platform for AI agents, on October 21, with commercial availability planned for the first quarter of 2027. Schneider Electric, NVIDIA, and HPE announced industrial AI agents for process plants on October 22.

One fact matters before any of this is repeated in a budget meeting. ISA, the organization that owns ISA 95, has announced no revision of the standard this month, this quarter, or this year. Its most recent ISA 95 publication is ANSI/ISA 95.00.01 2025, announced in April 2025, which updated models and terminology in Part 1. The eight layer stack is an analyst framework. Treat it as a map of where vendors are heading, not as a standard anyone must comply with.

2. How Does It Work?

Read the stack as a dependency chain, because that is how it behaves in a plant.

The bottom two layers produce signals. Machines, robots, and lines do the physical work, and the edge layer runs controllers and computing close to the equipment. Industrial DataOps is the plumbing: it collects data from machines and control systems and attaches context, so a temperature value arrives knowing which asset it belongs to, which line that asset sits on, and what product was running.

Systems of record are where the plant keeps its official truth. The historian holds time series data. The CMMS holds assets, work orders, and maintenance history. The ERP holds cost and materials. Nothing in the new stack removes these systems. The agents above them read from them, and increasingly write back into them.

The knowledge graph is the genuinely new idea. It is a structured model of the plant: this pump feeds that vessel, this motor belongs to this asset, this asset sits in this area, this failure mode applies to this class of equipment. An agent answering a question about a pump does not search documents and guess. It walks relationships that were defined in advance. That is what makes its answers consistent enough to act on.

Interfaces are the agreed connection points between the graph, the agents, and the systems of record. MCP is the one gathering momentum: a CMMS vendor exposes an MCP server, and an authorized agent can read work orders, pull analytics, and suggest repairs through it. Agents then do the running around: gathering evidence, drafting plans, coordinating across systems. People stay at the top, reviewing and approving before work is released.

3. Why Does It Matter to Industrial Operations?

Because the stack is honest about something vendors often skip: every layer depends on the layers beneath it. An agent is only as good as the knowledge graph beneath it. The graph is only as good as the records beneath that. And the records are your asset register, your naming conventions, your work order history, and your failure codes.

That is a maintenance problem before it is an AI problem. If the same pump carries one name in the CMMS, a tag number in the historian, and a nickname on the floor, the graph inherits three identities for one asset. If work orders close without findings or failure codes, the agent that suggests repairs from historical resolutions is learning from a thin and lopsided record. This week offered a clean illustration: L2L launched a platform whose AI suggests equipment fixes drawn from past resolutions. The feature is sound. Its ceiling is set entirely by the quality of closeout data the plant has been writing for years.

The week also showed the gap between demonstration and deployment. Siemens and NVIDIA described extensive AI use at Siemens' own Erlangen electronics plant, with claims of up to 85 percent less engineering effort. Reporting by TechTarget carried a caution from Interact Analysis: those figures describe potential at a Siemens owned showcase plant, not savings proven across multi vendor brownfield sites. Most plants are brownfield sites. The transferable work is not the model. It is the data readiness and integration work underneath, rebuilt site by site.

4. Practical Industrial Applications

Three applications follow directly from the stack as published, and none of them requires science fiction.

First, alert to planned work. A condition monitoring alert fires on a critical asset. An agent gathers the asset record, its criticality ranking, recent work history, and current operating context from the historian, then drafts a planned job with a parts list for a planner to review. A person approves it. The value is not autonomy; it is that the planner starts from a prepared package instead of a bare alarm.

Second, one answer per asset. A supervisor asks why a line keeps losing an hour every Tuesday. The agent walks the graph across the CMMS, the historian, and production records, and returns the pattern with the evidence attached. Today that investigation takes a reliability engineer an afternoon of exports and cross referencing. The graph is what makes the short version possible.

Third, the morning meeting. An orchestrator pulls open backlog ranked by asset criticality, overdue preventive work, and parts availability, and presents the short list worth discussing. This is decision support, not decision making, and it is the level of AI assistance most plants can absorb safely right now.

5. Implementation Challenges and Risks

Data readiness is the largest. Asset hierarchies, naming, and documentation are inconsistent in most plants, and the stack does not forgive that; it amplifies it, because errors in the record layer become confident answers at the agent layer.

Integration is the second. Every site connects differently, and this week a Gartner analyst quoted in trade coverage made the same point about scaling AI across plants: the work is rebuilding integrations and contextual data sets for each site. Budget for that work, not just for licenses.

Governance and safety come next. Interfaces that let agents read plant systems can, if configured carelessly, let them act on plant systems. No agent should create, approve, or close safety critical work without a person in the loop, and every agent action needs a log a supervisor can audit. Voluntary guidance published on October 16 by NIST, NSA, and CISA on readiness for AI agent incidents points the same direction: prepare, detect, respond, and keep the ability to shut an agent down.

Finally, there is a quieter risk in the diagram itself. Teams may treat an analyst proposal as an official standard and build compliance work, roadmaps, and vendor scorecards on it. The map is good. It is still a map drawn by one research firm in October 2026, and it will change.

6. What Should Plant Leaders Do Next?

None of the following requires buying a platform.

Clean the asset register. One asset, one name, one hierarchy, consistent across the CMMS and the historian. This is unglamorous work and it is the foundation layer of everything above it.

Fix work order closeout. Require findings and failure codes on completed work, at least on critical assets. Every closed work order is training material for whatever assistant your plant adopts in the next three years.

Rank assets by criticality, and keep the ranking current. Agent attention, planner attention, and spare parts money should all follow the same ranking. If the three lists disagree today, they will still disagree when an agent reads them.

Ask AI vendors three questions. What sits under your interface layer: is there a knowledge graph, or does the agent read raw tables? Can the agent write back into the CMMS, and under whose approval? Can we see and audit the reasoning behind a recommendation? The answers separate architecture from advertising.

Start with read only assistance. Let agents gather, summarize, and draft. Keep approval, release, and closure with people until your own evidence says otherwise.

7. KANE Industrial Perspective

Our read of the eight layer stack is that it is most valuable read from the bottom up. The industry conversation starts at the top, with agents, because agents are new and demonstrable. The dependency order runs the other way. In plant after plant, the binding constraint on industrial AI is not model capability. It is whether the asset data, documentation, and work history in the systems of record are accurate enough to build on. That is why our AI readiness work begins with asset data and documentation, and why our criticality work matters here: the ranking is what tells every layer above it where the consequences live.

We are also skeptical of the percentages travelling with this week's announcements, and we recommend clients be skeptical with us. Vendor figures from showcase plants describe what is possible under ideal conditions with the vendor's own engineers on site. They are not a forecast for a brownfield plant with a mixed fleet and a lean maintenance team. The honest claim is narrower and still significant: the architecture is converging, the interfaces are standardizing, and plants that fix their record layer now will adopt the top layers faster and cheaper than plants that do not.

8. Conclusion

ISA 95 has not been replaced, and no new standard arrived this week. What arrived is a credible map of the layers industrial AI is adding above the plant systems you already own, published by IoT Analytics on October 19, 2026, and echoed in product announcements from Rockwell and Schneider Electric within days. The map's message for maintenance and operations leaders is direct: the top of the stack gets the attention, but the bottom decides whether any of it works. Clean asset names, trustworthy records, disciplined closeout, and a current criticality ranking are the readiness program. They were worth doing before anyone drew the diagram. The diagram just made the cost of skipping them easier to see.

Sources and Further Reading

Frequently Asked Questions

Has ISA 95 been replaced by a new eight layer standard?
No. The eight layer stack is a proposal published by the research firm IoT Analytics in October 2026. ISA has not revised ISA 95 to include it. The most recent ISA 95 update is ANSI/ISA 95.00.01 2025, published in April 2025.

What is a knowledge graph in a plant context?
It is a structured model of the plant itself: assets, processes, and the relationships between them. It lets an AI assistant answer questions by following defined relationships, for example which motor drives which pump and what work has been done on it, instead of searching documents and guessing.

Which layer should a plant fix first?
The systems of record layer: the CMMS, the historian, and the asset data inside them. Agents and knowledge graphs build on those records, so errors and gaps there limit everything above them.

Does our plant need to buy an AI agent platform now?
Not to get started. Cleaning the asset register, enforcing work order closeout with failure codes, and keeping the criticality ranking current are no purchase steps that raise the value of any platform evaluated later.

What is MCP in plain terms?
It is a standard way for AI agents to connect to software systems. A vendor exposes an MCP server for its product, and an authorized agent can then read data, and where permitted take defined actions, through that connection instead of a custom integration.