How Heavy Industry Ops Teams Distinguish Process-First AI Vendors From Platform Vendors Wearing a Use-Case Mask

Two structural tells reveal whether a vendor starts with your workflow or its platform.

· 8 min read
Cover illustration for “How Heavy Industry Ops Teams Distinguish Process-First AI Vendors From Platform Vendors Wearing a Use-Case Mask”
AI Vendor Eval · November 9, 2026 · 8 min read · 1,843 words

AI operations transformation for asset-heavy industries can start with one broken process, without buying a whole platform or committing to an enterprise software strategy. The distinction between a process-first AI vendor and a platform vendor wearing a use-case mask lies in the engagement. A process-first engagement begins with the workflow, defines success as operational change, and makes decisions and exceptions traceable in production.

Why the platform-vs-process distinction matters now in heavy industry

An operations team buying AI for a plant or project must choose among vendors whose proposals sound more and more alike. Platform vendors, model providers, orchestration layers, and vertical AI products all say they are purpose-built solutions. The labels offer little help with that decision. “Platform” can mean anything from a single-model API wrapper to a full agentic deployment environment, without explaining who owns the work between installation and operational improvement.

That ambiguity makes vendor mismatch costly in industrial settings. The 2026 Roadmap on Artificial Intelligence and Machine Learning for Smart Manufacturing identifies integration with heterogeneous sensing and control systems, together with trustworthy and reliable operation, as critical deployment challenges. McKinsey’s July 2026 AEC report also says construction productivity grew far less than manufacturing over the same two-decade period, leaving little room to pay for deployments that never change the work.

How platform vendors learn to speak process-first without changing their underlying model

Platform vendors have learned to lead with use cases because buyers demanded an explanation of operational value. Yet that language can leave the underlying sequence untouched, with configuration first, integration later, and value deferred. An environment is the platform vendor’s first deliverable, while a changed workflow is the process-first vendor’s first deliverable.

This is more than a cosmetic change in sales language. A proposal can price the first stage like software setup even though the work calls for data remediation, exception-handling design, and change management across functions. In effect, it sells a phase-two operating model as phase one, before the operational groundwork exists. The buyer then inherits work that the use-case framing made appear included. Even a convincing demo leaves that mismatch in scope in place.

Agentic platforms built as wrappers on foundation models depend on providers that can change pricing, deprecate models, or update capabilities. Without backward compatibility guarantees, the buyer inherits the adaptation burden. Use-case language leaves open who must do that work. The vendor either owns a functioning process through those dependencies or merely supplies access to capabilities.

Where the engagement actually starts — the first structural tell

A process-first vendor asks to watch the workflow run and locate its failures, while a platform-led engagement begins with an IT architecture diagram and a data inventory. The vendor may need both kinds of information in time. Their order shows whether failures in the workflow direct the technical work or wait to be addressed until the technical environment is established.

A process-first entry produces a documented baseline of current decision volume, failure rates, and exception categories before a demo environment opens. That baseline gives the engagement something operational to change. By contrast, a capability map or reference architecture can defer addressing failures in the process to a later use-case prioritization phase. Completion of that phase does not itself produce a production system.

The McKinsey report argues for transforming “domains — end-to-end processes small enough to be redesigned independently but large enough to deliver meaningful business impact”. That framing gives process-first entry a practical boundary. It is neither an enterprise-wide platform program nor a disconnected demonstration.

For operations teams, the practical test is to ask the vendor to describe the first three weeks. If the answer starts with environment provisioning, API credentialing, or data-lake setup, the vendor is starting with a platform. Shadowing a dispatcher, scheduler, or procurement coordinator indicates process entry. The proposed work, rather than the promised use case, supplies the evidence.

Where accountability lives in the vendor’s model — the second structural tell

The statement of work exposes the second discriminator. Success defined as a live process with measurable operational change assigns a different obligation from success defined as a deployed instance with trained users. The former ties delivery to operational performance. The latter can be satisfied while the gap between configured software and a functioning workflow remains the client’s responsibility.

This distinction is independent of the starting point. A vendor can conduct a process audit and still transfer accountability back to the client once configuration begins. The contract therefore needs to be read against the proposed discovery work. Watching a workflow is useful evidence of intent, but it does not establish who owns the result.

When an AI-influenced decision produces a wrong output, someone must own the diagnostic. Platform-led models typically return that work to client IT, while process-first models build exception routing into the workflow design. The resulting process must also leave a record of its decisions and actions, so accountability is supported by evidence rather than an assurance that support will be available.

Deployment logic in procurement and scheduling — where the mask slips under operational pressure

The McKinsey report illustrates field coordination with prefabricated pipe spools that no longer fit after a late engineering change. In its scenario, agents with access to the latest 3D model, procurement records, and construction schedule can draft routing options and check material availability within minutes, reducing work that takes days to minutes. The example is prospective. Its operational premise is nevertheless precise, because the response depends on joining engineering, materials, and scheduling decisions inside the workflow.

A reporting layer alone does not establish that connection. Procurement and scheduling expose the difference when a constraint that designers did not anticipate becomes a primary variable. A configured tool may show the original plan well enough but have no process for dealing with the changed constraint. The diagnostic is what happens next, including who owns the redesign rather than merely identifying another configuration option.

Lead times for generators, switchgear, transformers, and cooling systems extend well beyond typical project planning horizons. A scheduling tool that does not treat power procurement as its primary constraint variable starts from a generic capability rather than the process governing energization readiness. An integrated deployment instead catches a delayed purchase order in the financial system and flags the resulting schedule risk.

For a team evaluating that deployment, the question is direct. “What happens when a constraint we didn’t define at setup becomes the critical path?” A process-first answer explains how the workflow changes and how responsibility follows that change. An answer limited to configuration options leaves the operational consequence unresolved.

What data quality reveals about which vendor type you are actually evaluating

A process-first vendor treats remediation as part of the workflow being rebuilt, while a platform-led vendor treats it as a client prerequisite. Both need usable data, but only one has included the work of getting there in the engagement.

That difference is consequential for heavy industry operations teams. Only a minority of firms rate their own data quality as high, so clean onboarding is not a sound default assumption. The failure pattern begins when a team selects a vendor for demo performance and then discovers that inconsistent production data prevents useful operation. The team absorbs a remediation project that was neither scoped nor priced.

The roadmap identifies effective data management and integration with heterogeneous sensing and control systems as critical industrial AI barriers. That puts getting the data ready within the engineering work. It does not support treating it as an incidental inconvenience outside deployment.

In week-one planning, operations teams should ask what happens when schedule data is incomplete or procurement records are fragmented. A process-first response describes how the engagement handles those conditions. A response limited to the data the client must provide before work can begin reveals a different allocation of responsibility.

How to use the audit trail requirement as a live POC test

For operations teams, the strongest single test is to require the vendor to demonstrate the audit trail from an actual process run. The record must expose the underlying work. It should show what arrived, how the system interpreted it, which rules fired, what actions followed, and what was written to which system.

For regulated or mission-critical workflows, the minimum record begins with a timestamp, decision ID, human user identity, and AI system and model version. These establish the event and the actors involved. Inputs with attribution and the specific policy invoked identify what informed the decision. Plain-language reasoning, the output, and the downstream action then connect interpretation to consequence. Human review status and tamper-evident proof complete the required evidence without substituting a dashboard for the trace.

Operations teams inspect traces to find where a process leaks, while compliance teams use them to demonstrate governance. IT teams need to distinguish a misread input from a rule mismatch, system timeout, or data quality problem. The vendor needs a record of each run to help teams trace those problems. A configured environment alone is not evidence that this production workflow exists.

The vendor should provide an evaluation harness that lets the operations team define its own eval sets, run them on every model or prompt change, and gate deployment on results. The roadmap identifies explainable and trustworthy AI as a non-negotiable requirement in high-stakes industrial environments, making these artifacts engineering deliverables.

How to start AI in heavy industry without a digital roadmap

A correctly structured first engagement identifies one painful process and produces a live changed workflow within weeks. Choosing a platform, drawing up a broader roadmap, and making wider changes all wait until that process is in production. The first deliverable should be a process audit. The vendor watches the work, documents where decisions fail, and proposes what to rebuild first, rather than opening with a capability map or transformation architecture.

A long initial timetable signals that the engagement may have been organized around platform deployment. The McKinsey report’s argument for end-to-end domain redesign supports this bounded approach without reducing it to an isolated use case. The unit of work must be small enough to change independently but complete enough to produce an operational result.

The strongest objection is that platforms can amortize across several use cases, while a specialist addresses one process and leaves broader transformation unresolved. That is a legitimate distinction in scope. However, a live first process creates an operational baseline and institutional confidence for defining the second. A platform must reach production to provide either.

Internal teams know the operational context and may be best placed to handle exceptions, but that responsibility is workable only if the rebuilt workflow clearly identifies exceptions for those teams. The vendor’s obligation is to make ownership executable.

The final evaluation therefore rests on four questions. Does the proposed first deliverable describe a changed workflow? Does success in the statement of work reference an operational outcome? Can the vendor demonstrate a live audit trail in the proof of concept? Did the vendor ask to watch the process run before proposing a solution? A negative answer to any of these questions leaves the process-first claim unsupported by the engagement itself.

Sources

  1. 2026 Roadmap on Artificial Intelligence and Machine Learning for Smart Manufacturing
Filed underAI Vendor Eval

More in AI Vendor Eval