Honeywell Is Rebuilding Itself Around Industrial Autonomy

The new Honeywell is smaller, more focused, and making a deliberate bet: as industrial AI moves from advising operators to acting on operations, deep process knowledge and proximity to control may matter more than ever.

Honeywell is becoming a very different company.

Over the past several years, it has separated Aerospace and Advanced Materials, sold its personal protective equipment business, exited Warehouse and Workflow Solutions, and most recently completed the sale of Productivity Solutions and Services.

The numbers make the transformation hard to miss.

The former Honeywell generated $38.5 billion in revenue in 2024. The newly independent Honeywell Technologies is guiding to roughly $20 billion in 2026 revenue.

So yes, Honeywell has become substantially smaller.

But describing the strategy as “shrinking” misses what is actually happening.

While reducing its corporate perimeter, Honeywell says it has deployed approximately $11.5 billion since 2023 into acquisitions that remain inside the new automation company.

Those investments have added compressor controls, OT cybersecurity, building access systems, LNG process technology, pumps and compressors, battery-safety sensing, and catalyst and process technology.

That does not look like indiscriminate cost cutting.

It looks like concentration.

Honeywell is giving up breadth to go deeper into a particular set of capabilities: process technology, sensing, safety, cybersecurity, automation, industrial software, and control.

And Honeywell has now given that strategy a name:

Automation to autonomy.

For Industrial AI, the interesting question is whether Honeywell has chosen the right place in the technology stack to make that bet.

One of the most revealing parts of Honeywell’s restructuring is what it chose to sell.

Warehouse and Workflow Solutions was not some obsolete industrial business. Honeywell had warehouse execution software, robotics, material handling, sortation, and software capable of combining information about labor, equipment, and processes.

Productivity Solutions and Services included mobile computing, scanning, printing, and frontline data capture.

Both businesses could easily fit within a broad definition of industrial automation.

Honeywell sold them anyway.

That tells us something about the company it is trying to become.

Honeywell is not simply collecting every technology that touches industrial operations. It appears to be specializing within automation itself.

Look at what it has been buying.

Compressor Controls brought turbomachinery control and optimization.

The former Air Products LNG business added process technology and cryogenic equipment.

Sundyne brought pumps, compressors, and deep installed-base equipment expertise.

Johnson Matthey Catalyst Technologies added catalysts, process licensing, and chemical-process knowledge.

SCADAfence strengthened OT cybersecurity.

Li-ion Tamer added specialized sensing for battery thermal runaway.

Viewed together, those acquisitions point toward a company going deeper into the physics, chemistry, equipment, sensing, safety, and control of industrial processes.

That is a very different strategy from simply buying more software.

And it could prove extremely important as Industrial AI gets closer to physical operations.

Honeywell does not need to own the world’s most capable foundation model.

Neither does Siemens. Neither does Schneider Electric.

Industrial companies can increasingly source intelligence from Microsoft, Google, AWS, NVIDIA, OpenAI, and others.

Honeywell itself is doing exactly that.

Its partnership with Google Cloud combines Gemini and Vertex AI with data from Honeywell Forge to create industrial agents and operational insights.

That points to a broader shift.

The model itself may not be the scarce asset.

Industrial context, operational authority, and the ability to act safely may be much harder to reproduce.

This is where Honeywell’s portfolio starts to become strategically interesting.

A general-purpose AI model can reason about a compressor.

Honeywell may know the compressor’s operating state, its history, the process around it, the control strategy, the safety envelope, the conditions under which it becomes unstable, and the actions experienced operators normally take when something starts to go wrong.

Add process licensing, catalyst knowledge, sensing, control history, alarm history, and equipment expertise, and this becomes more than simply having more data.

It gets closer to understanding why the physical process behaves the way it does.

This is where Honeywell’s approach begins to look different from some of the other Industrial AI architectures we have examined.

Cognite starts very explicitly with industrial data contextualization.

Its architecture connects time series, assets, engineering documents, ERP information, 3D models, P&IDs, files, and other industrial information into a structured knowledge graph that AI can reason over.

Schneider Electric’s proposed $3.1 billion acquisition of Cognite would place that capability alongside AVEVA’s engineering and operational software and Schneider’s automation estate.

Siemens is taking another route.

Intelligence Center X brings together industrial data, lifecycle context, AI, Graph Studio, AI Studio, Mendix applications, workflows, agents, and humans within a governed orchestration environment.

Siemens therefore appears to be working from broad engineering and lifecycle context toward AI-driven decisions and execution.

Honeywell seems to be starting somewhere else.

Its strongest differentiation may be the context closest to the physical process.

Experion knows the control environment.

Operators create a history of interventions and responses.

Sensors provide physical state.

Honeywell UOP contributes process technology.

Johnson Matthey adds catalyst and chemistry expertise.

Compressor Controls and Sundyne bring rotating-equipment knowledge.

And Honeywell’s LNG portfolio can now span parts of process design, process technology, equipment, control, optimization, and operations.

That creates an interesting possibility.

Honeywell may be building industrial context from the physics and process outward, while others are building from enterprise and engineering information inward.

That is not inherently better.

But it is different.

And as AI moves closer to physical action, that difference could become important.

One reason I initially questioned Honeywell’s Industrial AI architecture was Forge.

Historically, it was easy to think of Honeywell Forge mainly as an industrial SaaS, analytics, and asset-performance portfolio.

Honeywell’s 2026 positioning is broader than that.

The company now describes Forge as an open, hardware-agnostic intelligence layer that connects installed assets, contextualizes information through domain-specific models, spans IT and OT, and combines AI with deterministic models and operational constraints.

Honeywell says Forge connects approximately 6.4 million assets across 324,000 sites and more than 32,000 customers.

Those are Honeywell’s numbers, but the strategic direction is clear.

Honeywell is not conceding the context layer.

It is trying to build one.

The more difficult question is what kind of context Forge can ultimately provide.

Honeywell’s public material shows asset and process context, operational history, domain models, IT/OT connectivity, digital twins, analytics, and AI integration.

What is much less visible is a reusable semantic architecture spanning product genealogy, quality, maintenance, engineering configuration, ERP, supply chain, lifecycle information, documents, and plant operations.

That is an area where Cognite is unusually explicit.

It is also something Siemens is increasingly emphasizing through Graph Studio, Intelligence Center X, and its broader engineering-lifecycle portfolio.

There are signs Honeywell is moving further in this direction. Reporting from its 2026 user conference described a Honeywell Data Fabric architecture involving ontologies, knowledge-graph standards, and agent connectivity.

But the public technical picture is still developing.

That leads to a more useful question than, “Does Honeywell have context?”

Of course it does.

The real question is:

Can Honeywell make its unusually deep process knowledge computationally available across the broader enterprise context required for increasingly autonomous decisions?

The clearest evidence of Honeywell’s movement toward autonomy is not Experion Cognition.

It is Experion Operations Assistant.

The product grew partly out of work including Honeywell’s collaboration with Chevron around AI-assisted alarm management.

The underlying idea is compelling.

Industrial plants generate enormous volumes of alarms and operational history. Experienced operators also accumulate years of knowledge about which combinations of signals actually matter, which abnormalities tend to precede failures, and which interventions have historically brought the process back under control.

Operations Assistant uses historical and real-time operational data, together with site-specific knowledge and language-model capabilities, to identify developing conditions and advise operators.

At TotalEnergies’ Port Arthur refinery, Honeywell reported a pilot in which five potential events were predicted roughly 12 minutes before alarms occurred.

But the most important architectural detail is what happened next:

The operators took corrective action.

That places Operations Assistant squarely within a commercially credible Industrial AI pattern:

Observe → Predict → Recommend → Human decides → Existing systems execute

Honeywell commercially launched Operations Assistant in March 2026.

This is not autonomous manufacturing.

But it may be an important foundation for it.

And that brings us to Experion Cognition.

Honeywell introduced Experion Cognition in June 2026 as an AI-enabled control-system platform designed to move beyond operator assistance.

Honeywell says Cognition can make recommendations and automated decisions, detect and mitigate abnormal situations, and use AI agents to perform cognitive tasks on behalf of operators.

The distinction matters.

Prediction is one thing.

Recommendation is another.

Decision is another.

Execution is something else again.

Once AI moves from telling an operator what it thinks should happen to participating in decisions that change a physical process, Industrial AI enters an entirely different governance problem.

Honeywell may have an important structural advantage here.

Experion Cognition is not being developed as an AI layer sitting outside the automation environment and looking in.

It sits inside the Experion PKS ecosystem, close to the DCS, process models, operational history, operators, and conventional control technologies responsible for running the plant.

But this is also where the public evidence gets much thinner.

Honeywell says Cognition can make automated decisions.

It says agents can act on behalf of operators.

What Honeywell has not yet explained publicly, at least in enough technical detail, is what happens after an agent reaches a decision.

Can an agent change a setpoint?

Can it invoke procedural automation?

Does a deterministic policy engine validate the action first?

Does an operator need to approve certain classes of decisions?

Can an agent call an APC or MPC layer but not the controller directly?

Which decisions can be automated?

Which can never be delegated?

How are actions permissioned and audited?

What happens if the AI or network becomes unavailable?

Where is the boundary between Cognition and an independent safety system?

These are not minor implementation details.

They are the architecture of industrial autonomy.

Honeywell’s highest-profile Cognition example is Borouge International’s Ruwais facility.

The work matters because the tests took place in a live production environment.

Borouge has described potential improvements of up to 20 percent in plant efficiency, a 20 percent reduction in downtime, and up to a 15 percent reduction in operating costs.

But the wording is important.

Borouge describes the work as a proof of concept and the benefits as potential improvements.

It also describes a future progression toward a full-scale AI-driven control room.

So the evidence does not support saying Honeywell has already deployed autonomous AI control broadly at industrial scale.

It supports a narrower claim, but still a significant one:

Honeywell has demonstrated its agentic automated-decision concept in a live process environment and is now trying to industrialize it.

That is meaningfully beyond a PowerPoint architecture.

It is not yet proof of the final destination.

There is another useful caution here.

Borouge is also working with Yokogawa on autonomous control-room concepts.

Honeywell is not alone in this race.

The phrase “autonomous operations” can sometimes obscure more than it explains.

Industrial plants already contain a great deal of autonomy.

Controllers run continuously.

Interlocks protect equipment.

Advanced process control optimizes processes.

Model predictive control manages interacting variables.

Procedural automation sequences operations.

Safety instrumented systems provide independent protection.

AI is entering an environment that was already highly automated long before generative AI arrived.

Honeywell’s own autonomous LNG architecture reinforces this point. It combines Experion, procedural automation, APC/MPC, digital twins, asset-performance technologies, and AI/ML.

So the future is unlikely to look like:

AI replaces control.

It is more likely to look like:

AI operates within an increasingly sophisticated envelope of deterministic control, optimization, procedures, permissions, and safety constraints.

That may be where Honeywell has its strongest structural position.

Honeywell does not need AI to become deterministic.

It needs to determine where probabilistic intelligence can safely participate in a deterministic operational environment.

The Industrial AI architecture then becomes:

Intelligence → Context → Decision → Governance → Execution

The hardest transition may not be Intelligence to Context.

It may be Decision to Execution.

And Honeywell has spent decades building the control technologies that sit at that boundary.

It would be tempting to summarize the competitive landscape like this:

Honeywell owns execution.

Siemens and Schneider own context.

That would be too simplistic.

All three companies have major automation estates and substantial access to physical execution.

All three are building context.

All three are adding AI.

The more useful distinction is where each appears strongest today.

Siemens may have the broadest engineering and lifecycle position. Its challenge is whether it can coherently bring Mendix, RapidMiner-derived capabilities, Altair, knowledge-graph technology, engineering software, industrial data, automation, and agents together into one usable operating architecture.

Schneider Electric, assuming the Cognite transaction closes, would combine AVEVA’s engineering and operational footprint with one of the industry’s most explicit industrial contextualization architectures.

Its challenge is integration.

Buying Cognite does not automatically turn Cognite, AVEVA, EcoStruxure, and Schneider’s automation portfolio into one system.

Honeywell starts closer to process behavior and control.

Its challenge is almost the inverse: can deep domain and control context become broad enough to support decisions that increasingly depend on information outside the control room?

The competition may therefore have very little to do with who owns the largest AI model.

The more consequential question may be which direction is easier to travel.

Can Siemens and Schneider take broad context and move safely downward toward physical execution?

Or can Honeywell take deep process and control context and expand outward toward the wider enterprise?

That is a much more interesting race.

There is also an organizational argument behind the breakup.

Honeywell CEO Vimal Kapur has described the new company as having a simplified business model and a revamped innovation machine.

There is at least some early evidence of product momentum.

Honeywell moved from alarm-management collaboration in 2024, through Operations Assistant pilots, to a commercial release in March 2026, followed by Experion Cognition in June.

At the same time, it has assembled new process technologies, sensing, cybersecurity, and equipment businesses into the retained automation portfolio.

But this part of the thesis deserves caution.

A simpler portfolio does not automatically create a faster company.

And an acquisition does not become integrated simply because a press release describes the synergies.

Honeywell still has to demonstrate that concentration translates into faster product development, reusable architectures, common data models, stronger software integration, and quicker deployment across its customer base.

There is another part of this that I find particularly interesting.

Honeywell manufactures sophisticated products itself, including advanced sensors and semiconductor devices.

If the company can apply the same Industrial AI architecture inside its own manufacturing network, its factories could become powerful validation environments for the strategy.

Public evidence about Honeywell’s internal manufacturing architecture remains surprisingly limited.

That is something I will be watching.

Manufacturers evaluating Honeywell’s autonomy strategy should look past the word autonomous and ask more precise questions.

Where does the AI obtain context?

What information is persistent, and what is retrieved dynamically?

What decisions can an agent make?

Which actions require human authorization?

What deterministic constraints sit between AI and the process?

What happens when the AI is wrong?

What happens when the context is incomplete?

What gets recorded for audit?

Where does inference run?

What remains operational if the AI layer disappears?

And perhaps most importantly:

Who has authority to act?

Those questions will ultimately separate Industrial AI architectures built for real manufacturing from systems that simply demonstrate impressive reasoning.

Three things will determine whether Honeywell’s strategy becomes as consequential as its portfolio transformation suggests.

First, Experion Cognition has to move from proof of concept to repeatable production deployment.

Prediction and operator assistance are already becoming commercial capabilities. Automated industrial decision-making requires a much stronger body of evidence.

Second, Forge has to become a more explicit context architecture.

Honeywell has extraordinary process and operational knowledge. The question is whether that knowledge becomes computationally accessible across engineering, maintenance, quality, production, and enterprise systems rather than remaining embedded in individual products and human expertise.

Third, Honeywell has to prove organizational velocity.

The company has spent years reshaping the portfolio.

Now it has to show that a narrower Honeywell can integrate acquisitions and deliver coherent Industrial AI systems faster.

Honeywell’s transformation has changed how I think about its Industrial AI position.

The obvious interpretation is that Honeywell dismantled a conglomerate and emerged as a smaller automation company.

That is true.

But I think it is incomplete.

Honeywell appears to be concentrating around something more specific: the technologies that understand, sense, protect, optimize, and control physical processes.

That creates a fascinating Industrial AI bet.

Honeywell’s argument is not that context matters less than control.

It is that context closest to the physical process may become disproportionately valuable once AI is expected to act.

There is a real risk to that strategy.

Process context alone may not be enough.

The best industrial decision can depend on information sitting far outside the DCS: engineering revisions, quality results, maintenance history, product genealogy, materials, orders, schedules, and enterprise constraints.

Siemens and Schneider are moving aggressively toward that broader context.

Honeywell is trying to broaden Forge without losing its depth near the process.

So the decisive question is not whether Honeywell has context while its competitors have control.

They all have both.

The question is this:

Can Honeywell turn unusually deep process knowledge and proximity to control into trustworthy autonomous execution faster than Siemens and Schneider can connect broader lifecycle and enterprise context to their own automation systems?

We do not know the answer yet.

But after Honeywell’s restructuring, that is increasingly the right question to ask.

About Industrial AI Review

Industrial AI Review is an independent publication covering industrial AI, manufacturing technology, automation, and industrial software.

Industry. Intelligence. Impact.

industrialaireview.com