Autonomy vs. Augmentation: Where AI Actually Works in Manufacturing
Every week, we interview practitioners and distill industry podcasts and conferences into what you need to know.
In today's issue:
From Automation to Autonomy: Why Gilead and McCain Trust AI in Operations—But Not for Final Decisions
Why digital transformation investments haven't delivered expected productivity gains (the solution: interoperable data spaces)
A practical, phased digital transformation approach for midsize manufacturers
1. Autonomy vs augmentation: real-world AI deployments in pharma and food manufacturing
UC Irvine’s Center for Digital Transformation Digital Leadership Agenda 2025, Panel: Automation to Autonomy: How to Build Autonomous Supply Chains and Smarter Operations with AI with Jay Agarwal (McCain Foods) and Harmik Begi (Gilead Sciences) (Nov. 13, 2025)
Gilead Sciences and McCain Foods are running autonomous operations in their plants, but with a critical distinction:
McCain runs fully autonomous cold supply chains with zero human intervention. Gilead uses AI for predictive maintenance that cuts unplanned downtime 20-40% and quality control improvements of 15-30% but regulatory compliance means humans still make the final call.
The difference reveals which manufacturing processes can go fully autonomous today (logistics, maintenance, supply chain planning) versus which must remain augmented (quality decisions, regulatory compliance, product release).
TLDR:
Pharma can automate monitoring and recommendations but regulatory compliance requires human decision-making for product release.
Food manufacturing can go fully autonomous in supporting operations like cold supply chain logistics while core production remains human-supervised: McCain runs blacked-out warehouses with low oxygen content and zero human intervention.
The factory of the future requires embedded AI agents that run locally at the edge, not cloud-dependent dashboards because when your plant loses internet connectivity or you build factories next to farms in remote areas, you still need to manufacture product.
The framework many manufacturers miss:
Distinguish between monitoring / recommending AI versus decision-making AI (regulatory posture determines the boundary). Any manufacturing (not just pharma) with safety implications, quality requirements, or regulatory oversight faces the same constraint. AI can predict maintenance needs, recommend process adjustments, optimize supply chains but the final decision to ship product, adjust formulations, or release batches must remain human-supervised until regulatory bodies themselves adopt AI.
McCain Foods runs fully autonomous cold supply chain operations with "low oxygen content and all blacked out" and zero human intervention. The difference? McCain's autonomous operations handle logistics and storage, not final product quality decisions that impact food safety directly.
Deploy AI where it enhances human decision-making first: supply chain, maintenance, and supporting operations offer the highest ROI with lowest risk. Both Gilead and McCain see massive opportunity in supply chain orchestration, integrated demand-supply planning, and predictive maintenance. These supporting operations around core manufacturing deliver autonomous value without triggering regulatory constraints.
Gilead is focused on "end-to-end supply chain orchestration" – connecting demand planning to manufacturing schedules across their network of sites, with AI updating plans automatically based on real-time constraints.
McCain emphasizes predictive maintenance that generates more maintenance notifications, requiring organizational resilience to process increased work orders without adding headcount.
Design for embedded control and disaster recovery – resilience requires local intelligence when connectivity fails. The common architecture is sending everything to the cloud, running models, sending recommendations back to dashboards that operators read. This fails in two ways: operators have too many dashboards, and remote factories with limited connectivity can't afford cloud dependency.
At McCain: "We build our factories next to the farms. Sometimes we have to build the road to get there." When your plant loses internet connectivity, you still need to manufacture. This requires embedded AI agents running locally at the operational technology (OT) level, making decisions at the edge without cloud dependency.
What to do about this:
→ Don’t try to make everything autonomous right away. Map your manufacturing processes into three categories: autonomous now (logistics, supply chain planning, maintenance), augmented now (quality control, process optimization, predictive analytics), and regulatory-blocked (product release, compliance decisions, safety-critical approvals).
→ Start with supply chain orchestration and predictive maintenance pilots that integrate with existing systems (SAP, ERP, MES) rather than building standalone AI applications. Your AI investments should feed an underlying tech stack, not create more data silos.
→ Build organizational resilience to handle increased AI output. Predictive maintenance generates more maintenance notifications; autonomous supply chain planning creates more frequent schedule changes. Your teams need capacity and processes to act on AI recommendations.
2. Why digital transformation investments haven't delivered expected productivity gains (the solution: interoperable data spaces)
Advanced Manufacturing Now podcast, Episode: How Data Spaces and Digital Twins Are Transforming Manufacturing Efficiency (Nov. 13, 2025)
Industry 4.0 promised dramatic productivity gains from digital transformation, but European data shows manufacturing productivity hasn't increased as expected despite massive technology investment.
The reason? Companies build isolated digital experiments on different tech stacks with no interoperability. A supply chain system doesn't talk to the manufacturing execution system. The quality control AI can't access logistics data. Companies optimize individual processes but can't connect systems to create enterprise-wide efficiency.
Data spaces solve this by creating secure, standardized pipelines for sharing data across company boundaries while maintaining sovereignty. You don't upload data to a central platform. You create authenticated connections to access data on partner systems. Data spaces and digital twins finally deliver the promised Industry 4.0 productivity gains by standardizing data, enabling interoperability, and extending your system architecture beyond your company walls.
TLDR:
Industry 4.0 hasn't delivered expected productivity gains because companies build isolated digital experiments with no interoperability (quality control AI, predictive maintenance, supply chain optimization), preventing network effects that drive real transformation.
Data spaces create secure infrastructure for sharing data across company boundaries while maintaining sovereignty. Unlike central platforms where everyone uploads data, data spaces are authenticated pipelines where you access data on partner systems directly, deciding exactly what to share with suppliers, logistics providers, and customers.
Digital twins enable optimization by creating real-time virtual replicas of physical assets but the real value comes when twins across the value chain connect through data spaces, allowing collaborative optimization of production planning.
The framework many manufacturers miss:
Previous industrial revolutions (automation, electrification) showed measurable productivity increases. Industry 4.0 hasn't. Universities and consultancies researching this identify a common pattern: "Companies do digital transformation in forms of different experiments. They start a POC in this department, they build an MVP in another, and develop digital solutions but they don't work together, they're built on different tech stacks, they don't have proper data flows, they don't have proper APIs to interconnect to other systems."
This matters because the value of digital transformation comes from network effects, not isolated improvements.
A quality control AI that detects defects 15-30% better is valuable. But that same AI becomes transformative when it automatically triggers maintenance work orders in your CMMS, adjusts production schedules in your MES, notifies suppliers about material quality issues, and feeds data to your supply chain planning system, all in real-time.
Most manufacturers invest in individual use cases but neglect the underlying infrastructure that creates interoperability.
Your supply chain optimization project uses one tech stack. Your predictive maintenance AI uses another. Your quality control system is separate. None of them share data automatically. The result: you've digitized processes without creating a digital enterprise.
Use data spaces to extend your system architecture beyond company boundaries. The efficiency gains come from collaborative optimization with suppliers and partners, not just internal process improvement.
Think of data spaces as secure, authenticated pipelines. You maintain full sovereignty over your data. But you create specific, policy-controlled connections that let partners access defined datasets. A logistics provider can access your production schedule to optimize delivery timing. A supplier can see your consumption forecasts to plan inventory. A customer can monitor your production status for their order in real-time.
A concrete example: "If I'm missing some logistical data that would improve my processes a lot, I'll ask my logistics provider directly to provide me this data and in this way I get more efficient." And you're not just optimizing your internal processes, you're optimizing the entire value chain.
The competitive advantage comes from participating in ecosystems. When 47 German industrial companies (competitors in real life) collaborate in Factory X to build shared data space infrastructure, they're not giving away competitive secrets. They're creating the underlying infrastructure that lets them individually optimize operations by accessing better data from partners.
Implement digital twins not as isolated simulations, but as connected virtual replicas that enable collaborative optimization across the value chain. Digital twins have "gained a lot of traction in the last decade" but most implementations are isolated. A digital twin of a production line optimizes that line. A digital twin of a logistics network optimizes routing. These deliver value, but limited value.
The transformative use case: digital twins that connect through data spaces. Your production planning twin connects to your supplier's inventory twin, which connects to your logistics provider's delivery twin, which connects to your customer's consumption forecasting twin. Now you can optimize not just your production schedule, but the entire value chain's efficiency simultaneously.
What to do about this:
→ Audit your Industry 4.0 investments and map which systems actually share data automatically. If your quality control AI doesn't automatically trigger maintenance work orders, adjust production schedules, and notify suppliers, you're getting 20% of the potential value.
→ Pilot data space implementation with one supplier or logistics partner where better data sharing delivers obvious efficiency gains. Select a partner where you're missing critical data (delivery timing, inventory levels, quality metrics) and implement secure data sharing.
→ Design digital transformation projects to feed an underlying tech stack rather than creating standalone solutions. Require APIs, standardized data models, and integration architecture for every new system. This approach costs more upfront but delivers compounding value as each new system automatically interoperates with existing infrastructure.
3. A practical, phased digital transformation approach for midsize manufacturers
CIO Lifeline podcast by VCIO Global, Episode: From Legacy to AI-Ready with with Allen Mazelin (Vista Manufacturing) (Nov. 13, 2025)
Vista Manufacturing is a midsize manufacturer that started digital transformation in 2020 with three-quarters of an IT person and a managed services provider doing tactical work. Five years later, they've implemented Dynamics 365 Business Central, made Microsoft Teams their entire internal communication infrastructure (eliminating email internally), modernized file storage, and are now exploring AI in accounting and purchasing.
Vista didn't have a full-time CIO. They hired a virtual CIO to provide strategic direction. The transformation started with executive conviction, proceeded through careful roadmapping, and succeeded because leadership made technology transformation a non-negotiable priority.
TLDR:
Digital transformation requires executive leadership that personally champions technology adoption and mandates usage. Vista's motto became "anything you do internally goes through Teams," eliminating fragmented email communication.
Mid-market manufacturers can access strategic IT guidance through virtual CIO services like coaching to create roadmaps, evaluate technologies, and build internal capabilities.
AI adoption follows the same pattern as previous digital transformation: start with executive education, pilot in high-ROI areas (accounting, purchasing), and focus on augmenting employees who are excited about technology.
Start with executive conviction and strategic roadmapping, not technology selection. Allen Mazelin met his virtual CIO Craig at a fundraiser event. Craig's vision was "helping smaller companies who couldn't have a full-time CIO come in and give strategic direction and help them with their IT needs."
But Mazelin didn't delegate transformation to an IT manager. He personally owned it as a business priority. Whereas most mid-market manufacturers fail digital transformation because executives treat it as an IT project instead of a business transformation.
Leverage virtual CIO services for strategic direction when you can't justify full-time executive IT leadership. Vista used a managed services provider for tactical work: "they were just the tactical piece, we weren't really getting any strategic piece." The virtual CIO model provided strategic direction, coaching, technology evaluation, and roadmapping without requiring full-time executive hire.
This model works for mid-market manufacturers who need strategic IT guidance but can't justify $200K+ for a full-time CIO. The virtual CIO helped Vista create roadmaps, select technologies, evaluate cloud storage options, and coach the small IT team on implementation. But Mazelin made decisions and drove adoption. The virtual CIO provided expertise and options, not execution.
Treat AI adoption as the next phase of ongoing transformation, not a separate initiative. Start with employee education, identify excited early adopters, and pilot in high-value, repetitive-task areas.
Vista recently completed an "Executive AI Journey" with their virtual CIO consultant: monthly training sessions on "here's what AI can do for you, here's how you do the prompts," getting feedback on what people are using it for, why things aren't working.
The next AI initiatives: accounting and purchasing. "They've got a lot of detailed tasks, repetitive stuff" that AI can automate.
What to do about this:
→ If you're a mid-market manufacturer ($10-50M revenue) without a full-time CIO, evaluate virtual CIO services that provide strategic guidance, roadmapping, and coaching for your small IT team.
→ Create a strategic technology roadmap before selecting specific tools. The roadmap should identify business outcomes (better collaboration, eliminate paper, improve customer service) before specifying technologies.
→ Start AI pilots in areas like accounting and purchasing where repetitive, detailed tasks create clear ROI without requiring massive process redesign.
Disclaimer
This newsletter is for informational purposes only and summarizes public sources and podcast discussions at a high level. It is not legal, financial, tax, security, or implementation advice, and it does not endorse any product, vendor, or approach. Manufacturing environments, laws, and technologies change quickly; details may be incomplete or out of date. Always validate requirements, security, data protection, labor, and accessibility implications for your organization, and consult qualified advisors before making decisions or changes. All trademarks and brands are the property of their respective owners.
