Industrial AI has a scaling problem. Perhaps we should stop blaming the technology

Imagine a manufacturer testing AI that predicts when a machine is likely to fail. The pilot works, engineers are impressed, and the business case promises fewer breakdowns. Then someone suggests deploying it across five factories, with different equipment, data systems and maintenance routines, and suddenly the algorithm is the straightforward part.

This is becoming a familiar story. IFS and Accenture's 2025 study of 800 manufacturing leaders found that 96% use AI in service delivery, yet only 28% have fully scaled it across their operations. We have become exceptionally good at demonstrating what AI could do, but apparently less accomplished at organising ourselves to benefit from it.

The factory is not a laboratory

The UK's 2026 AI Adoption Plan for Advanced Manufacturing identifies fragmented data, legacy infrastructure, uncertain returns and concerns about safety and assurance as obstacles to deployment. Gartner's 2026 survey of 140 supply-chain leaders found that 56% struggled to integrate AI with existing systems and processes, while 50% reported limited internal expertise.

None of this is surprising when a factory must keep producing while new technology is introduced. Our predictive maintenance system now needs reliable operational data, access to existing equipment, engineers who understand its limits and procedures for what happens when a prediction is wrong. A successful demonstration answers only some of those questions.

Yet we still tend to judge AI one pilot at a time, while the value depends on relationships between technology, people, infrastructure and decisions. The visible achievement is a working model; the less visible challenge is whether the organisation can actually accommodate it.

Making the invisible visible: When everyone does their job, but nothing moves

Our manufacturer now wants to buy the system. Procurement asks about the supplier, cybersecurity examines vulnerabilities, legal reviews liabilities, and engineering needs assurance that production will not grind to a halt. All perfectly sensible questions, although the supplier might reasonably wonder why nobody asked them before the pilot began.

The problem appears when each department works through its own requirements, requests evidence already reviewed elsewhere and reaches a decision without seeing how the pieces fit together. Everyone can be doing their job responsibly while the organisation collectively struggles to get anything done.

Procurement is not necessarily the cause of stalled adoption, but it is where many unresolved assumptions and dependencies become visible. Technical success does not automatically make a supplier enterprise-ready, just as a successful pilot does not prove that a buyer is ready to deploy. Instead of blaming slow purchasing processes, we should reframe procurement as a way to establish the evidence, accountability and operational readiness needed to scale.

Trust needs evidence, not another promise

Industrial buyers need to know what a system was tested on, where its limits lie, who monitors it and what happens if its performance changes. Describing AI as trustworthy is pleasant enough, but it is unlikely to satisfy the factory manager responsible for keeping the machines running.

This is where standards, training and independent assurance can reconnect technical claims with real purchasing decisions.

IEEE 3119-2025, the first standard of its kind focused on AI procurement, establishes a structured process for defining requirements, assessing vendors, contracting and monitoring systems. IEEE's Responsible Procurement of AI training programme puts that standard into practice through tools such as risk registers and supplier evaluation rubrics. Complementing this, IEEE CertifAIEd product certification takes the evidence a step further through independent assessment against defined ethical criteria, including transparency, accountability, privacy and algorithmic bias.

Certification cannot replace industrial safety testing or guarantee performance in a particular factory. But recognised assessment methods and evidence that suppliers can reuse may reduce duplicated work and help buyers distinguish substantiated claims from impressive sales presentations. Surely that would be progress.

From successful pilots to responsible scale

The industrial technology market already understands that connecting systems creates value. In June 2026, Siemens and IFS announced a partnership intended to link engineering intelligence with operational systems across manufacturing and asset management. The organisational equivalent is to connect the decisions that determine whether AI can move into production.

That means revealing dependencies before they become surprises, reframing procurement and assurance as part of deployment design, and reconnecting engineering, suppliers, risk teams and operational leaders early enough to act on what they learn. The purpose is to enable responsible scale, where value is measured through reliable performance and sustained productivity rather than another completed pilot.

Industrial AI will keep getting more capable, but increasingly intelligent machines cannot compensate indefinitely for organisations that struggle to evaluate, buy and operate them. We have invested enormous effort in making machines more intelligent; perhaps the next breakthrough requires making the systems around them a little more intelligent too.