Executive Summary
Manufacturing has the highest ratio of AI pilots to production deployments of any major sector. Deloitte's survey of 600 senior manufacturing executives (2025) finds that 67% have active AI pilots in production-adjacent functions, but only 18% describe their AI program as fully integrated into core manufacturing operations. McKinsey's State of AI 2025 manufacturing findings confirm the same pattern: pilot success rates are high, but the step from pilot to production stalls at the same point in most organizations: where the AI system must connect to legacy operational technology (OT) infrastructure, coordinate with adjacent systems, and operate without a dedicated project team managing the exceptions. The WEF Global Lighthouse Network, which certifies manufacturing facilities for advanced technology integration, identifies system-level coordination: not AI capability: as the primary differentiator between lighthouse facilities and the rest.
Manufacturing leads on AI pilots and lags on production
Manufacturing has the highest ratio of AI pilots to production deployments of any major sector. Deloitte's survey of 600 senior manufacturing executives (2025) finds that 67% have active AI pilots in production-adjacent functions, but only 18% describe their AI program as fully integrated into core manufacturing operations.
McKinsey's State of AI 2025 manufacturing findings confirm the same pattern: pilot success rates are high, but the step from pilot to production stalls at the same point in most organizations: where the AI system must connect to legacy operational technology (OT) infrastructure, coordinate with adjacent systems, and operate without a dedicated project team managing the exceptions. The WEF Global Lighthouse Network, which certifies manufacturing facilities for advanced technology integration, identifies system-level coordination: not AI capability: as the primary differentiator between lighthouse facilities and the rest.
The stall is a D2 coordination gap, not an AI capability gap
D2: Digital Cognitive organization explains the manufacturing AI stall precisely. A pilot operates in isolation: it has its own data source, its own performance definition, and its own team managing exceptions. A production deployment operates in the organization: it must share data with adjacent systems, its performance is measured against business outcomes rather than model metrics, and exceptions are handled by operators who have other responsibilities.
The D2 requirement is that the organization learns through the AI system's operation: that outcomes feed back into the next cycle of decisions. Hindustan Unilever's Tinsukia factory, a WEF Global Lighthouse Network certification holder, demonstrates what this looks like: production variables, quality outcomes, and maintenance signals all feed a shared operational model that continuous improves maintenance scheduling and batch parameters. The difference between Tinsukia and a stalled pilot is not the AI technology. It is the D2 architecture: closed loops, shared data, coordinated decision rights.
Fund only pilots that already have a production architecture
Manufacturing executives approving AI investments should ask one question before signing: does this pilot have a production architecture: a defined path to connecting the AI system to the operational systems it will need to work with at scale? If the answer is "we will figure that out when we get there," the probability of joining the 82% that stay in pilot purgatory is high. The WEF Lighthouse criteria are publicly documented and provide a reference standard for what production-grade AI integration in manufacturing looks like. Using those criteria as a design specification for new AI programs: rather than waiting for a Lighthouse audit: is a practical way to move the pilot-to-production trajectory.
Sector Context: Production Is a Systems Problem
Manufacturing creates a harder environment for AI than a controlled pilot suggests. A model may predict a quality deviation or maintenance event accurately, but production value depends on whether that prediction can enter an operational workflow without disrupting safety, throughput, traceability, or equipment availability. The model therefore sits inside a wider production system made up of OT, enterprise applications, plant procedures, operator judgment, maintenance schedules, and capital assets with long replacement cycles.
That context changes the meaning of “AI readiness.” Readiness is not simply access to models, data scientists, or cloud infrastructure. It is the ability to connect intelligence to production decisions repeatedly and safely. A plant that cannot expose reliable machine data, route a recommendation to the right decision-maker, record the action taken, and capture the resulting outcome has not created a learning loop. It has created an analytical endpoint.
The sector transition is therefore from isolated optimization to cognitive operations. In the legacy model, improvement is organized around individual assets, lines, projects, and functions. In the emerging model, production, quality, maintenance, planning, and supply signals become part of a connected decision architecture. AI becomes valuable when those domains can learn from one another.
Four Forces Behind the Pilot-to-Production Gap
Legacy OT constrains integration. Manufacturing estates contain equipment and control systems built across different generations. A pilot can bypass that complexity with a dedicated data extract. Production cannot. Scaling requires governed interfaces, common semantics, and integration patterns that work without creating fragile point-to-point dependencies.
Operational exceptions expose weak work design. Pilot teams absorb ambiguity manually. At scale, operators must know when to trust a recommendation, when to override it, who owns an exception, and how that decision is recorded. Human-AI decision rights are therefore part of the production architecture, not a change-management activity added after deployment.
Local optimization can damage system performance. A model that improves one machine or process may shift constraints elsewhere. Production AI must be measured against end-to-end outcomes such as throughput, yield, quality, downtime, energy use, and schedule adherence rather than model accuracy alone.
Learning requires outcome capture. A cognitive production system must observe whether an intervention worked. If predictions are logged but maintenance actions, quality outcomes, and operator overrides are not fed back into the system, the organization cannot improve the model or the process that surrounds it.
From AI Pilot to Cognitive Production System
The structural shift is from proving that an algorithm works to designing an operating system in which intelligence can be reused. That requires shared operational data, stable interfaces between OT and IT, explicit decision rights, outcome instrumentation, and governance that survives the pilot team.
D2 is the anchor because the challenge is organizational cognition: sensing, deciding, acting, learning, and adapting. D3 supports it by providing the platform foundations through which data and services can be reused across plants. D4 matters because production deployment must be sequenced around architecture and operational constraints rather than treated as a sequence of disconnected pilots. D5 defines the operator-AI relationship, especially around overrides, escalation, and accountability. D6 can accelerate deployment through reusable blueprints, digital twins, process intelligence, and common integration assets, but acceleration without architecture simply reproduces fragmentation faster.
Opportunities and Risks
The upside of production-grade AI is not limited to automating tasks. Closed learning loops can improve maintenance timing, process stability, quality control, energy efficiency, and planning while allowing knowledge generated in one part of the network to become reusable elsewhere. The compounding effect is the strategic prize.
The risks are equally structural. Poorly integrated AI can create unsafe recommendations, brittle dependencies, untraceable decisions, and new cyber exposure at the OT/IT boundary. Excessive automation can also remove operator judgment from situations where contextual knowledge remains essential. A production architecture must therefore optimize for resilience and accountability as well as speed.
Five Executive Priorities
Require a production architecture before pilot funding. Every proposal should identify target systems, data dependencies, decision owners, exception paths, outcome measures, and the route from local deployment to reusable capability.
Create common OT/IT integration patterns. Reduce repeated engineering by defining governed interfaces, data contracts, identity controls, and observability standards that plants can reuse.
Design the human-AI decision model explicitly. Specify which recommendations are advisory, which can trigger automated action, where human approval is mandatory, and how overrides become learning data.
Measure operational outcomes, not model performance alone. Tie AI performance to throughput, quality, downtime, cost, energy, safety, and other plant outcomes that executives already govern.
Treat every scaled deployment as a reusable asset. Capture architecture patterns, data models, controls, training, and operating procedures so the next plant starts from an established foundation rather than another pilot.
Three signals that will show whether the coordination gap is closing
Three forward indicators: WEF Lighthouse Network additions, McKinsey's 2026 manufacturing findings, and enterprise resource-planning integration rates: will reveal whether manufacturing is finally bridging that gap.
- WEF Global Lighthouse Network 2026 additions and sector pattern: which manufacturing sectors are adding Lighthouse facilities and which are not will indicate where D2 integration capability is being built versus where the pilot purgatory pattern persists.
- McKinsey State of AI 2026 manufacturing edition: the 2025 findings identified the coordination gap; the 2026 edition will show whether manufacturing organizations have changed their program design in response.
- Deloitte manufacturing AI maturity follow-up survey: Deloitte's 2025 600-exec survey baseline will be compared against 2026 data; whether the 18% production integration figure improves will indicate sector-wide progress on the coordination challenge.
Closing Perspective
The central issue is structural rather than technological. The organizations that create durable advantage will be those that turn the capability described in this brief into part of the operating model, with clear ownership, reusable architecture, measurable outcomes, and governance that persists beyond an individual project. The leadership question is therefore not whether to adopt another tool or launch another initiative. It is whether the sector's operating architecture is being redesigned so that each investment strengthens the next one.
