Most Industrial AI evaluations start one step too late.
The shortlist gets assembled from whatever came up in a search for "industrial AI software," which returns data platforms, vision inspection tools, predictive maintenance products, ERP copilots and AI development platforms in a single list. Those solve different problems. Comparing them on the same scorecard produces a winner that may not address the constraint you actually have.
So the first question is not which vendor. It is which category — and that has a determinate answer once you know what is missing.
Category | What it delivers | Buy it when |
Industrial data platforms | Connected, contextualized, governed OT data that AI can run on | Models or use cases exist but the data underneath them isn't usable or consistent |
Embedded AI in automation and enterprise suites | AI inside software you already run — control systems, ERP, asset management | The workflow you want to improve already lives in that suite, and its data is already there |
Specialist applications | One use case, deeply — vision inspection, machine health, process optimization | You have a specific, measurable problem and want it solved fast |
General-purpose AI development platforms | Model development and application tooling without industrial specificity | You have in-house data science capability and want to build rather than buy |
Most manufacturers end up running more than one. That is normal and not a planning failure. What determines whether the combination works is whether they all draw on the same governed source.
Three symptoms map cleanly onto three categories.
The data layer is missing. A model that performs at plant one and cannot be deployed at plant two is not a model problem — the first deployment proves the use case, and the second reveals that tags, context and structures differ at every site.
The application layer is missing. If a governed data foundation is already in place, buying a specialist application or using embedded AI in an existing suite is the faster path.
Connectivity is missing, which is upstream of everything else and usually solved together with the data layer.
A fourth, less comfortable symptom: "We have four AI tools and none of them agree on OEE." That is a data-layer problem presented as a reporting problem. Each tool brought its own pipeline and its own definition.
These are hard constraints, and they eliminate options faster than any feature comparison.
If a decision has to resolve in under a second — a vision reject on a packaging line has under 100ms — it has to run at the edge, and cloud-only software is out regardless of how good it is. If data cannot leave the site for regulatory reasons, air-gapped operation is required, not preferred. If plants have unreliable connectivity, anything assuming a persistent cloud connection will fail intermittently in ways that are hard to diagnose.
Conversely, model training, cross-site pattern detection and enterprise benchmarking belong in the cloud, where scale is available and latency does not matter.
Most production architectures need both, which is a reason to prefer software that supports both over software that is excellent at one.
This is the question that most changes the answer, and the one pilots are structurally unable to test.
Single-site deployments succeed on software that cannot scale. Past roughly three plants, the deciding capability becomes standardization — common data models, template-based rollout, centralized model and version management — rather than feature breadth. A specialist application that is excellent at one plant may require a full reimplementation at the next.
Published reference point for what standardization enables: a Food & Beverage manufacturer reached 95 global sites in 18 weeks using template-based rollout.
Every category carries a form of lock-in, and none of them are free.
Embedded AI in an automation or enterprise suite is the fastest to adopt and the hardest to leave, because the value depends on being inside that estate. Specialist applications are easy to leave individually and hard to leave collectively, because by the fourth one you have four pipelines. Data platforms create lock-in through the data model itself, which is the most portable form but not free either.
The question to ask any vendor: if we replaced you in three years, what would we have to rebuild?
If you are starting now, the order matters more than the vendor.
- 1.
Solve connectivity and context first, at one site, for one asset class. Not a portfolio. One.
- 2.
Prove it at a second site before scaling. The second site is where architecture problems surface. Finding them at site two is cheap.
- 3.
Then add applications. With a governed foundation in place, individual use cases deploy in days rather than months — traditional Industrial AI deployments take 12 to 18 months to reach production, and the majority of that is data work.
- 4.
Add governance before the third use case, not after the tenth. Lineage and ownership are far cheaper to establish on three pipelines than on thirty.
Two common situations.
You bought applications first and they don't scale. This is the majority case and it is recoverable. The applications usually survive; what gets replaced is the integration layer underneath them, which is then reused by everything that follows.
You bought a data platform and nothing has been built on it. Less common and usually a sequencing problem rather than a tooling one — the foundation was built without a use case attached, so nobody owns the outcome. Attach one use case with a named owner and a measurable result before extending the platform further.
Four: industrial data platforms that make OT data usable, AI embedded in automation and enterprise suites, specialist applications for single domains such as vision inspection or machine health, and general-purpose AI development platforms applied to industrial problems.
Diagnose which layer is missing before shortlisting vendors. If pilots work but don't repeat, the data layer is missing. If data is clean but there are no use cases, the application layer is. Then check latency, air-gap and site-count constraints, which eliminate options faster than features do.
A data platform delivers connected, contextualized, governed data. An AI application delivers a prediction, a classification or a recommendation. The application runs on what the platform produces.
For one point solution, often no. By the third or fourth, usually yes — each arrives with its own pipeline, and without a shared foundation they will not agree with each other.
Some can. Local inference, including GPU-accelerated vision and locally hosted small language models, can run inside the plant in fully air-gapped environments. Verify this specifically rather than assuming it, because most software in this market assumes cloud connectivity.
Traditional Industrial AI deployments take 12 to 18 months to reach production. Where a standardized data foundation already exists, individual use cases can deploy in days.