By Alex Housley, Advisor, Solvace
There is a pattern that keeps surfacing across manufacturing organizations that are serious about AI, and it rarely gets the attention it deserves.
The investment case gets approved. The models are selected. The vendor is onboarded. And then, somewhere between the pilot and production rollout, the initiative quietly stalls. Not because the technology failed. Because the environment it was deployed into was never consistent enough to support it.
AI systems are fundamentally pattern-recognition engines. They learn by identifying regularities across large volumes of comparable data. When that data is fragmented, when definitions shift between sites, when processes are executed differently across facilities, when improvement activities are logged in incompatible formats, the model cannot distinguish signal from noise. The output degrades. Recommendations lose relevance. Adoption stalls.
The core issue is one that digital transformation roadmaps consistently underestimate: before you can scale AI, you need to standardize operations. Not as a precondition to satisfy before investment begins, but as a strategic priority to pursue in parallel. KPMG’s 2026 Global Tech Report on Industrial Manufacturing puts it plainly, noting that while 83% of manufacturing executives believe they are building strong AI data foundations, 76% still cite unreliable data as a top AI risk. Confidence and capability are not the same thing.
That gap almost always comes down to standardization. And it operates across three distinct dimensions that organizations consistently underestimate.
Dimension One: Process Standardization
Process standardization is the foundation on which everything else depends. It means aligning the rhythm, structure, and digital capture of core operational routines consistently across every site in a network.
Consider daily tier meetings. In most multi-site manufacturing organizations, these gatherings share a name but diverge significantly in execution. One facility runs a structured 15-minute stand-up with documented escalations logged digitally before the team disperses. Another conducts an informal walkthrough with outcomes written on a whiteboard. A third, under production pressure, skips the structure altogether. From the perspective of an AI system designed to detect operational drift or flag escalation risk, these three facilities are producing fundamentally incomparable data.
The same issue applies to maintenance routines, safety inspections, quality checks, and shift handover protocols. Wherever a process varies in execution across sites, there is a structural gap between what AI requires and what it receives. A predictive maintenance model trained on data from one facility will not generalize reliably to another if the underlying inspection workflows, logging conventions, and fault taxonomies differ.
It is worth being precise about what standardization means here. It does not require eliminating local autonomy or forcing uniform practices across facilities with different regulatory environments, equipment generations, or workforce structures. The objective is to standardize the data layer: the structure, sequence, and capture mechanism of core processes, while allowing local adaptation in everything that does not affect comparability.
Digital execution tools are the practical mechanism for achieving this. A structured checklist on a tablet, a tiered escalation workflow, a standardized shift handover form does not constrain how work is done so much as they ensure that how work was done is recorded consistently. That consistency is the input AI actually needs. Without it, you are not building intelligence. You are generating an elaborate audit trail that no model can usefully learn from.
Dimension Two: Performance Standardization
If process standardization defines how work is captured, performance standardization defines the vocabulary AI uses to interpret it. KPIs are not just management tools. They are the semantic layer through which AI reads operational health, and inconsistency at this layer produces analysis that is not just imprecise but actively misleading.
KPMG notes that data formats and standards vary across different production lines, making integration a technical challenge. At the KPI level, that problem runs deeper still. OEE calculations vary. Downtime classifications include or exclude planned maintenance depending on local convention. Yield is measured at different points in the production process. These discrepancies are often invisible at the human level, because experienced operations leaders have developed intuitions for translating between sites. AI systems have no such intuition. They process definitions literally.
The consequence is that cross-site AI analysis produces outputs that are not just imprecise but structurally wrong. A benchmarking model that ranks site performance will produce incorrect rankings if the underlying metrics are not comparable. A predictive model trained on multi-site data will learn spurious patterns introduced by definitional inconsistency rather than genuine operational signals. More dangerously, it may flag high-performing sites as underperforming, or fail to surface genuine risk, because the data it is reading does not reflect a shared reality.
Resolving this requires three things: a shared taxonomy of operational KPIs with precise, unambiguous definitions applied consistently across all sites and functions; standardized measurement protocols specifying where each metric is captured, at what frequency, and through what system; and standardized visualization ensuring that performance data is presented in a consistent format across every facility and leadership layer.
This work is often politically complex. It surfaces long-standing local conventions and requires governance structures to sustain over time. But it unlocks something genuinely valuable: every site’s data becomes a learning input for every other site. A model that identifies an early quality signal at one facility can surface the same pattern at another weeks before it manifests in production output. That cross-network intelligence is what most organizations are hoping AI will deliver. It requires this foundation to function.
Dimension Three: Improvement Standardization
The third dimension is the least discussed and, in many deployments, the most consequential. It addresses not how work is performed or measured, but how it is improved.
AI-driven continuous improvement depends on one critical input: a structured, searchable history of past improvement activity. For a model to recommend a preventive action, it needs to have learned which actions resolved comparable problems in the past. For it to identify recurring failure patterns, it needs to have analyzed hundreds of previous root cause cycles and found the common threads. That analysis is only possible if those cycles were documented using a consistent taxonomy: shared categories for problem types, root causes, countermeasure types, and outcome metrics.
In most manufacturing organizations, this is not the case. One site runs formal five-why analysis logged in a shared platform. Another uses a local spreadsheet. A third captures outcomes in a free-text comments field. The human knowledge exists. The organizational learning does not, because it has never been encoded in a format a system can read.
Consider a model trained on 500 structured root cause cycles across 20 sites, using consistent logic and taxonomy. That model can surface intelligence that no human analyst could replicate at that scale. It can identify which failure categories are systematically underaddressed at specific facilities. It can detect when a countermeasure that resolved an issue at one site has never been deployed where the same failure mode recurs. It can accelerate organizational learning in ways that compound over time.
But that capability depends entirely on the quality and consistency of what it learns from. Right now, at most organizations, the improvement history is fragmented, inconsistently structured, and largely inaccessible to machine learning systems. The institutional knowledge is there. The infrastructure to make it usable is not.
The practical path forward is to agree on a standard structure for every class of improvement activity: how it is initiated, how the problem is scoped, how root causes are categorized, what interventions are recorded, and how outcomes are measured. Building that structure into the tools teams use daily ensures that compliance is a byproduct of normal workflow rather than an additional administrative layer.
Standardization as Strategic Infrastructure
The argument here is not that AI investment should wait for perfect standardization. Operational environments are never static, and there is no ideal starting point. The argument is that standardization should be treated as strategic infrastructure, not a pre-project checkbox.
Organizations that build consistent process, performance, and improvement standards across their networks are creating a durable advantage. They replicate what works more reliably. They identify failure earlier. They onboard new sites with less friction. And when AI is deployed onto that foundation, its outputs are not just locally informative but globally actionable.
The competitive gap in manufacturing over the next decade will not be determined by who adopted AI earliest. It will be determined by who built the operational foundation that makes AI work at scale. That foundation is, ultimately, a data architecture problem expressed through human behavior. Solve the behavior, and the data follows. Follow the data, and the intelligence becomes real.
Before the next AI investment conversation, the most important question on the table is not what the model can do. It is whether the operational environment is consistent enough for the model to learn anything meaningful from it.
##
ABOUT THE AUTHOR
Alex Housley is a serial technology entrepreneur and AI specialist with over two decades of experience founding and scaling tech companies. He is currently serving as an advisor at Solvace and, he is the founder of Frontier Advisory, which helps businesses leverage AI and build for the rapidly evolving technology landscape. Previously, he founded Seldon, one of the world’s largest open-source MLOps platforms, leading it from inception through Series B and working with major enterprises including Ford, Verizon, and Capital One. A long-standing advocate for open-source AI and responsible innovation, Alex chairs the Open UK AI Advisory Board and has spent a decade advising the UK government on AI policy.
