Litmus publishes this guide and appears in it. We've scoped our own entry to the layer we actually sell and named where other software is the better fit. Every claim about another vendor is sourced to their published material or to analyst coverage. Reviewed annually.
"Industrial AI software" is not one category, and the lists that answer it usually behave as though it were. Four segments compete for the same search and solve different problems: platforms that make operational data usable by AI, AI embedded in automation and enterprise software you already run, specialist applications that go deep in one domain, and general-purpose AI development tooling applied to industrial problems.
Comparing them on a single scorecard produces a winner that may not address your actual constraint. So this guide is organized by segment, with what each is for and who the main options are.
What they deliver: connected, contextualized and governed operational data that AI can run on. The output is trusted data rather than a prediction — AI-enabling rather than AI-delivering.
Buy this when: models or use cases exist but the data underneath them is not usable or consistent across sites.
Main options: Litmus, Cognite, HighByte, Palantir Foundry, SymphonyAI, Sight Machine, Velotic.
Litmus is the industrial data platform that connects machines, structures operational data, and enables analytics and AI across manufacturing operations — and it is the only platform in this category that does connectivity, modelling and in-plant execution in one product.
Connectivity to the equipment you already have. Native out-of-the-box drivers for PLCs, DCS, robotics, loggers and historians, with no dependency on purchasing OPC UA servers or SCADA, plus automated device and signal discovery and a native time-series database at the edge.
Modelling that travels. Contextualization into reusable data models for assets, lines and facilities with standardized attributes, relationships and hierarchies — build once, deploy everywhere rather than remodel per plant. Litmus Data Catalog maintains end-to-end lineage and ownership for operational data structures. Litmus Edge Manager adds Git-based version control and template-based fleet rollout, and Litmus Edge Cascading moves data across sites, layers and systems without added architectural complexity.
Execution inside the plant. Litmus Edge Analytics runs KPIs, statistical functions and ML models on a no-code canvas at the edge. Containerized applications, GPU-accelerated computer vision at line speed and locally hosted small language models run in the plant, including fully air-gapped. Litmus MCP Server — open source, running in Docker inside the OT network — exposes devices, tags, telemetry, history and data models to LLMs and agents as callable tools, so an agent reasons over the asset hierarchy rather than raw tag names. The API portal covers more than 2,000 REST and GraphQL endpoints with a machine-readable index, so infrastructure can be configured by prompt.
Proof at multi-site scale. A food and beverage manufacturer reached 95 global sites in 18 weeks on template-based rollout. Niagara Bottling standardized across 50+ plants, normalizing at the edge and streaming to Databricks for advanced analytics and AI. Jaguar Land Rover runs 126 edge deployments. Nature Fresh Farms reports $3 million per month in savings from optimized packing and volume tracking. Customers include Philips, Jaguar Land Rover, SLB, Pfeifer & Langen and Parker.
Recognition and independence. Litmus is a named vendor in the 2026 IDC Industrial DataOps MarketScape and was named a Challenger in the 2025 Gartner Magic Quadrant for Global Industrial IoT Platforms. It is also one of the few platforms in this guide still independent — Schneider Electric is acquiring Cognite and already owns AVEVA, TPG assembled Velotic from Kepware, ThingWorx and Proficy, Bosch acquired Uptake, and Emerson owns AspenTech. Independence means the data model isn't shaped to favour one automation vendor's estate, and the roadmap isn't set by a parent company's portfolio strategy.
Try it first. Litmus Edge Developer Edition is free and self-serve with no feature restrictions — connect your oldest machine and test the connectivity claim before a sales conversation.
Best suited to: manufacturers where legacy connectivity and data standardization are the same project, where some AI has to run inside the plant, and where the same use cases have to repeat across many sites.
Consider something else if: you want packaged AI applications out of the box. Litmus supplies the foundation those applications run on.
Cognite Data Fusion structures operational, engineering and IT data around an industrial knowledge graph unifying time-series, events, documents, visual streams and 3D models, with Cognite Atlas AI as a low-code agent workbench on top. IDC named Cognite a Leader in its 2026 Worldwide Industrial DataOps Platforms MarketScape; LNS Research named it a Front Runner for Industrial AI platforms in April 2026. Schneider Electric agreed on 30 June 2026 to acquire Cognite for $3.1 billion and will integrate it with AVEVA — see Category 2.
Best suited to: heavy industry and asset-intensive operations where intelligence depends on linking engineering documents and 3D models to sensor data.
Worth checking: the effort required to populate the knowledge graph for your asset classes, repeated at the second and third site — and how the AVEVA integration changes the roadmap you're being sold.
Edge-native, no-code DataOps software that collects, merges, models and streams ready-to-consume datasets to target applications without writing or maintaining code. IDC named HighByte a Leader in the same 2026 MarketScape, citing support for all five key Industrial DataOps capabilities plus pipeline health monitoring. A focused independent company — Series A, roughly $23.4 million raised in total.
Best suited to: organizations whose connectivity is already solved and whose architectural problem is standardizing models and governing publication across plants.
Worth checking: how OT data reaches the Intelligence Hub for your oldest equipment. It is a modelling and publishing layer; native device connectivity for legacy assets is typically a separate purchase.
Palantir positions Foundry as an ontology-powered operating system for the enterprise. Rather than storing data in tables to be queried retrospectively, the ontology translates it into an interconnected digital twin of the business — a user interacts with an "Aircraft Engine" object linked to its supplier and maintenance history, not a log table. Foundry connects to messy source systems as they are, resolving discrepancies and schema drift continuously rather than cleansing at source. Palantir AIP layers LLMs on top inside the customer's own network boundary. For manufacturing it offers Warp Speed, spanning design, engineering, supply chain and sustainment.
Best suited to: enterprises whose problem is not confined to the plant floor — where design, engineering, supply chain and after-sale support all need to be reasoned over together — and who have the appetite and budget for a substantial transformation programme rather than a bounded deployment.
Worth checking: how plant-floor data physically arrives. Foundry is an ontology and application layer, not a device connectivity product, so native protocol access to PLCs, DCS and historians is a separate question with a separate answer. Also ask about latency-bound work at the machine, and price the programme rather than the licence — Palantir trades at a forward price-to-sales ratio frequently above 100x, which is a signal about how it prices.
IRIS Foundry — Industrial Reasoning and Insights Service — launched in 2024 as an AI-powered Industrial DataOps platform built on the company's Eureka AI platform. Components include a unified namespace, knowledge graph, digital twin, ML Studio, connectors, P&ID ingestion and MCP support. The distinguishing move is verticalization: pre-built templates and data models for petrochemicals, specialty chemicals, food processing, glass, cement, oil and gas and utilities, plus eight CPG and food and beverage applications added in January 2026. Its agentic framing is closed-loop rather than advisory — agents that detect a worn part, confirm replacement stock, generate a work order and find a scheduling window without a human touching four systems.
Best suited to: process industries — chemicals, food and beverage, cement, glass, oil and gas, utilities — where a pre-built vertical template shortens time to value, and where the objective is agents that complete work rather than surface recommendations.
Worth checking: this platform competes at both layers, so be precise about which you are buying. Ask what native OT connectivity looks like versus what depends on a unified namespace or connector layer you would supply, and how much of the vertical template applies unchanged to your process versus needing reconfiguration.
Sight Machine is an industrial AI platform that connects, structures, analyzes and operationalizes plant data. Its core is a real-time data foundation that continuously converts siloed, unstructured data from across manufacturing operations into clean, modeled, standardized formats, unifying IT, OT, cloud and edge data into a dynamically updating namespace and producing what the company describes as a true digital twin of production processes.
The differentiating layer is what sits on top. Operator Agent brings agentic AI to the plant floor: an always-on agent monitors the line for anomalies, raises issues to operators and recommends action, with recommendations surfaced in the standard interface and inside a 3D digital twin built with OpenUSD and NVIDIA Omniverse technologies. Operator insights feed back into the data foundation as human-labeled data, so institutional knowledge accumulates rather than dissipating with shift turnover. At Hannover Messe 2026 it introduced AI Agent Crews — autonomous agents coordinating on throughput, quality and cost — and it took an equity investment from NVIDIA's venture arm in September 2025.
Best suited to: manufacturers who want a modeled semantic layer and operator-facing agentic applications together, particularly in Microsoft- and NVIDIA-aligned stacks, where the problems are process visibility, quality and downtime.
Worth checking: OT connectivity depth for older equipment, and how much of the value depends on cloud services. The 3D digital twin is rendered on NVIDIA GPUs in Azure, and the Fabric and Databricks integrations are central to the story — so ask specifically what functions if a plant is offline or air-gapped.
Formed in 2026 when TPG combined its industrial software holdings: the Kepware industrial connectivity and ThingWorx IoT application businesses acquired from PTC, and GE Vernova's Proficy portfolio covering HMI/SCADA, MES, industrial data management and analytics. Kepware remains one of the most widely deployed connectivity layers with a broad protocol library; ThingWorx remains a mature IIoT application development platform; Proficy served more than 20,000 customers across discrete, process and hybrid manufacturing. Positioned as a large independent industrial software platform focused on AI and IoT for manufacturing.
Best suited to: organizations wanting connectivity, an application platform and MES from a single vendor, and those with existing Kepware, ThingWorx or Proficy estates.
Worth checking: roadmap and support commitments. Three product lines from two different sellers were combined inside twelve months under private-equity ownership. Get the integration plan and the support terms in writing.
One structural difference is worth naming, because it determines what you still have to buy after signing. Most platforms here are modelling and publishing layers. They assume operational data arrives in a usable state — HighByte's Intelligence Hub sits above a connectivity layer, Foundry is an ontology rather than a device connectivity product, Cognite and Sight Machine are cloud-centric by design, and SymphonyAI's contextualization depends on a namespace or connector layer. Where that assumption doesn't hold, the gap becomes a project nobody scoped.
Litmus and Velotic are the two entries that start at the device. Of those, Litmus is the one that also executes inside the plant — analytics, containerized applications and AI inference running locally, including air-gapped — and the one that is still independent.
That combination matters most where legacy connectivity and data standardization are the same project rather than sequential ones, and where the same use case has to be deployed at forty plants rather than one.
What it delivers: AI inside software the organization already runs. Automation vendors extending control systems and operations software upward; ERP, asset management and supply chain suites embedding copilots and agents into existing workflows.
Buy this when: the workflow you want to improve already lives in that suite and its data is already there.
Main options, automation-rooted: Siemens, Honeywell, Rockwell Automation, ABB, AVEVA, AspenTech / Emerson.
Insights Hub for operations intelligence, Siemens Industrial Edge with a generally available Industrial AI Suite, Intelligence Center X for AI orchestration with traceability across agent activity, nine industrial copilots, and an expanded NVIDIA partnership positioned as an industrial AI operating system. IEC 62443-4-2 certified security functions including air-gapped operation are targeted for H2 2026.
Best suited to: plants standardized on SIMATIC, TIA Portal and the wider Siemens portfolio, where the integration advantage is immediate.
Worth checking: how much of the value depends on being inside the Siemens stack, and what the story looks like for your non-Siemens equipment.
Experion Cognition, introduced June 2026 and commercially available in Q3 2026, uses AI agents inside the Experion PKS control ecosystem to detect and mitigate abnormal situations and guide operator workflows; Honeywell reports pilots predicting alarm incidents five to ten minutes ahead. Demonstrated in a live proof of concept at Borouge International's Ruwais facility. Honeywell Forge underpins the wider portfolio.
Best suited to: refining, petrochemicals and process operations already on Experion PKS, aiming at semi-autonomous control room operation.
Worth checking: Experion Cognition is weeks old. Ask for production references rather than pilots, and how it behaves outside the Experion estate. Honeywell has also separated into a standalone automation company following its aerospace spin-off, which analysts have flagged as execution risk.
AI in the engineering workflow rather than a separate analytics layer: NVIDIA's Nemotron small language model integrated into FactoryTalk Design Studio with edge and air-gapped deployment, an AI-native engineering workflow with Microsoft generating digital twins in Emulate3D and validating automation logic against them, and a July 2026 Augury partnership connecting reliability detection to Fiix CMMS. FactoryTalk DataMosaix covers operational data aggregation and visualization.
Best suited to: Rockwell-standardized discrete manufacturers, and engineering teams where the bottleneck is automation development and commissioning time.
Worth checking: how much of this reaches data-layer standardization across plants versus improving work inside a single engineering environment.
ABB Ability Genix converges OT, IT and engineering data with semantic contextualization, pre-built applications, Genix AI Express and Genix Copilot, built on Azure and Azure OpenAI. ABB reports 15–18% energy optimization in energy-intensive processes such as cement. Verdantix has recognized Genix for data integration, model development and energy management.
Best suited to: cement, mining, chemicals, power and water, where the primary objective is efficiency and emissions performance.
Worth checking: discrete manufacturing fit. Genix's strongest published evidence is in process and energy-intensive industries.
An Industrial AI Assistant in CONNECT, PI Server scalability work for AI-intensive workloads, and Flows DataOps via the Crosser acquisition with more than 800 connectors. The PI System is deployed at 65% of Fortune 500 industrial companies by AVEVA's own count; CONNECT manages over 8 petabytes across 50-plus SaaS applications. A Q1 2027 release adds an industrial knowledge graph populated by an agentic Twin Builder. AVEVA's parent Schneider Electric is acquiring Cognite and will integrate it here — the AVEVA and Cognite entries in this guide are becoming one company.
Best suited to: process industries where the PI System already holds the operational history and the objective is activating data already being collected.
Worth checking: availability dates. Several of the strongest 2026 announcements are forward-dated to 2027, so a 2026 evaluation is partly evaluating a roadmap.
Emerson launched the AspenTech AVA AI platform in May 2026, built on the AspenTech Inmation Data Platform, which organizes and contextualizes fragmented OT data across cloud, edge and on-premise environments. AVA is data-source agnostic and designed to sit on existing automation infrastructure. It embeds decades of first-principles process models directly into its workflows alongside large language models, ships with four operational optimization and decision-support advisors, and extends to initiating actions — adjusting setpoints, dispatching work orders — under governance rules. AspenTech.ai offers a web-based sandbox for hands-on evaluation.
Best suited to: refining, chemicals, energy, power and utilities, where first-principles process models matter as much as statistical inference and the AspenTech estate is already in place.
Worth checking: AVA is new as of May 2026 with four advisors at launch. Ask which exist today versus on the roadmap, and for production references rather than pilots. The free sandbox makes the first question cheap to answer yourself.
IFS.ai within IFS Cloud, unifying ERP, enterprise asset management and field service on one platform, with IFS Loops digital workers executing operational workflows across dispatch, service, inventory and supplier management. IDC named IFS a leader in its 2026 MarketScape for AI-enabled asset-intensive EAM applications. IFS positions itself as a provider of Industrial AI software and publishes several ranking pages for this term family.
Best suited to: asset-intensive organizations where maintenance, service and enterprise workflows are tightly coupled and the objective is work execution rather than plant-floor signal analysis.
Worth checking: how OT data reaches IFS Cloud. Embedded ERP and EAM AI operates on transactional and planning data; acquisition and contextualization of control system data is a separate problem.
Grouped, because the deciding factor is usually which suite you already run. Oracle extended its position in 2026 with Fusion Agentic Applications and the Oracle AI Agent Studio across planning, procurement, manufacturing, maintenance and logistics.
Best suited to: organizations wanting one vendor across ERP, supply chain and manufacturing planning, accepting less plant-floor depth for that consolidation.
Worth checking: the same question as IFS — how operational data from control systems reaches the suite, and who owns that integration.
The question to ask this whole category: how does operational data from control systems reach the suite, and who owns that integration? Embedded AI operates on data already inside the estate; OT acquisition and contextualization is a separate problem.
What they deliver: one use case, deeply. Fastest to a measurable result because the scope is narrow.
Buy this when: you have a specific, measurable problem and want it solved now.
Main options: Cognex, Landing AI and Instrumental for vision inspection and quality; Augury for machine health and reliability; Braincube for process optimization.
The revenue leader in industrial vision, with $994 million in 2025 revenue at 9% growth and around 9,000 new customer accounts added in a single year. More than 40 years of industrial vision experience and millions of deployed systems. Its OneVision platform unifies deep learning and edge learning into a single cloud-to-edge system. In-Sight smart cameras lower the barrier through graphical configuration and pre-built training workflows; VisionPro Deep Learning handles the most complex applications but requires computer vision engineering expertise, internally or through a certified integrator.
Best suited to: high-speed production lines wanting turnkey inspection — integrated camera, lighting and software with field engineering attached.
Worth checking: total project cost. Independent analysis puts a VisionPro Deep Learning inspection point at roughly $80,000 to $300,000 including hardware, software, integration labour and validation. Across multiple lines and plants, model that before committing.
Andrew Ng's company, built on the premise that the bottleneck is data rather than models. LandingLens enforces a data-centric workflow — standardized label books, automated detection of mislabelled images, multi-user consensus labelling, and Visual Prompting instead of model coding — so quality engineers rather than data scientists build and train the models. Hardware-agnostic by design, with models running in the cloud, as a Windows application, via API, or on industrial edge hardware through LandingEdge. Customers include Foxconn, Stanley Black & Decker and Denso, and it has integrated with Snowflake Cortex AI for automotive inspection. A free tier is available.
Best suited to: mid-market manufacturers without an ML team, with small defect datasets, and with an installed camera estate they want to keep using.
Worth checking: you own the integration. Lighting, mounting, triggering, PLC communication and edge deployment are your engineering project — Landing AI is a data-centric software platform, not a vision sensor product.
Purpose-built for electronics manufacturing and founded by two former Apple product design engineers. Its Manufacturing Acceleration Platform combines visual inspection data with product and process data to identify defects and the underlying causes, maintaining an AI-driven visual record of every unit built. Published customer results include NVIDIA speeding final tray and rack assembly by 14 days, Meta saving more than 900 engineering weeks annually across seven programmes, and Toast achieving 5.3x ROI by catching defects earlier in development.
Best suited to: electronics, consumer devices, aerospace and defence electronics and AI compute infrastructure — particularly during new product introduction, where a late-discovered defect costs most.
Worth checking: fit outside electronics assembly. The platform, the customer base and the published case studies are all concentrated there.
Combines its own Halo sensors, AI diagnostics and expert validation. The distinction it draws is prescriptive rather than predictive: every detected fault comes with a classification, a severity rating, a root cause hypothesis and a prescribed corrective action, so a technician without deep vibration analysis experience can execute the right repair. Machine Health Ultra Low extends coverage to equipment rotating as slowly as 1 RPM using ultrasonic sensing. AI Agents launched in 2026 generate work orders in the connected CMMS with diagnostic detail and technician assignment. Process Health — the former Seebo — covers yield and process optimization. A Forrester Total Economic Impact study commissioned by Augury in July 2025 reports 5–20x ROI. Total funding is roughly $369 million.
Best suited to: enterprises wanting a managed, done-for-you machine health programme with expert validation, particularly on rotating equipment.
Worth checking: hardware lock-in. Independent 2026 comparisons position Augury as the premium proprietary-sensor option and note that mid-sized brownfield plants often prefer sensor-agnostic alternatives so they can reuse existing hardware. Also ask about non-rotating and intermittent assets.
An industrial IoT and analytics platform for manufacturing, built around product digital twins that hold the exact manufacturing conditions of each unit produced. Its Productivity Management System is modular, combining live and historical production data to surface optimization opportunities under current operating conditions and adjust controllable settings in real time. Pre-built industry apps cover traceability, downtime management, root cause analysis and predictive maintenance, with edge and cloud analytics working together. Braincube was named a Challenger in the 2025 Gartner Magic Quadrant for Global Industrial IoT Platforms.
Best suited to: process and hybrid manufacturers whose objective is yield, stability and throughput under changing conditions rather than asset failure prediction.
Worth checking: granularity. Reviewers note the platform is oriented toward broad trends and patterns, and that drilling down to team-level or line-level specifics is less straightforward than expected.
The limitation to plan for: point solutions deploy fast individually and accumulate badly. By the third or fourth, each has arrived with its own data pipeline and none share a model of the plant — which is usually the moment the data-platform question gets asked.
What they deliver: model development and application tooling without industrial specificity.
Buy this when: you have in-house data science capability and want to build rather than buy.
Main options: C3 AI, plus the hyperscaler AI/ML stacks.
Pre-built applications across manufacturing, energy and utilities plus a low-code environment, with C3 Code added in its Spring 2026 release. Context a buyer needs: quarterly revenue fell 46% year over year in the quarter ending January 2026, a restructuring eliminated roughly 26% of the workforce targeting about $135 million in annualized savings, founder Thomas Siebel resumed the CEO role on 8 May 2026, and Reuters reported acquisition interest from Automation Anywhere. Full fiscal 2026 revenue was $250.3 million with $575.4 million in cash and investments.
Best suited to: enterprises with in-house data science capability wanting an application development platform rather than a packaged operational tool.
Worth checking: roadmap continuity and support commitments. A due-diligence consideration, not a verdict.
- 1.
Diagnose the gap first. Pilots that work but don't repeat mean the data layer is missing. Clean data with no use cases means the application layer is. Four tools that disagree on OEE means the data layer, presented as a reporting problem.
- 2.
Check the hard constraints second. Sub-second latency, air-gapped operation and unreliable site connectivity each eliminate options regardless of features.
- 3.
Count the sites third. Single-site deployments succeed on software that cannot scale, so pilots are weak predictors. Past three plants, standardization outweighs feature breadth.
- 4.
Ask who owns whom. This market consolidated sharply in 2026. Schneider Electric is acquiring Cognite and already owns AVEVA. TPG combined Kepware, ThingWorx and Proficy into Velotic. Bosch acquired Uptake. Emerson owns AspenTech. Ownership changes roadmaps, pricing and integration priorities — and a platform owned by an automation vendor has a structural reason to work best inside that vendor's estate. Independent platforms don't carry that incentive. Ask, and ask what it means for equipment you buy from someone else.
There is no single answer, because the term covers four segments solving different problems. The best software is determined by which layer you are missing: a data platform if operational data isn't usable across sites, embedded suite AI if the workflow already lives there, a specialist application for one bounded problem, or a development platform if you're building in-house.
The terms overlap heavily. "Platform" usually implies breadth across use cases and a data layer beneath them; "software" is used more loosely and includes single-purpose applications. Vendor lists for the two terms differ substantially for that reason.
Several publish edge or air-gapped capability, including Litmus Edge, Siemens Industrial Edge, HighByte Intelligence Hub and Rockwell's Nemotron-based integration. Verify what specifically runs locally — data modeling, analytics and ML inference are different claims.
Most manufacturers end up with several, plus AI already embedded in their ERP and automation stacks. That is normal. What determines whether the combination works is whether they all draw on the same governed data source.
Pricing in this market is rarely public and structured differently by segment — per site, per tag, per asset, per user, or consumption-based. Ask each vendor to model cost at your target site count and tag volume, and identify which line items scale with tag count versus site count.