Claudio Laterreur, CIO/CDO, Technology and Business C-Suite Executive, Transformation, Turnaround, & Growth, Advisory and Board Member.
Artificial intelligence (AI) has become one of the most compelling opportunities in manufacturing. Applications ranging from predictive maintenance and production optimization to quality improvement, operator assistance and increasingly agentic AI can make factories more responsive and productive.
However, enthusiasm can create its own blind spot. Most manufacturers understand that AI depends on good data, yet many recognize their operational information is not where it needs to be. Machine data, ERP transactions, maintenance records, quality information, logbooks, spreadsheets and operator knowledge frequently remain fragmented, delayed or poorly contextualized. Second-hand data will inevitably produce second-hand AI predictions and/or recommendations.
On the shop floor, an incorrect recommendation can influence quality, safety, production, customers and financial performance. Data readiness is a prerequisite, and it should not be the end of the AI readiness discussion.
The broader question is, “Are manufacturers evaluating the full cost and risk of AI, or primarily the cost of implementing it?” The traditional business case is familiar, i.e., software, implementation and people measured against expected value. As AI moves closer to operational recommendations and decisions, manufacturers should also consider human oversight, governance and revalidation, cybersecurity and privacy exposure, potential liability, and the computing and energy infrastructure supporting AI. These costs may not sit neatly inside the original project budget, yet they remain part of the economics of deploying AI responsibly.
Let’s break down some of the key dependencies to consider.
The cost goes beyond the software.
The most immediate hidden cost may be the humans-in-the-loop. Manufacturing AI cannot always be treated like an administrative copilot because its recommendations can have physical consequences. Should an agent produce hundreds of recommendations per shift and each requires review, approval or exception handling, the organization may create a labor requirement and bottleneck never included in the ROI. Add the expertise required to revalidate models after line changes, address model drift and maintain operating context, and the cost of autonomy can look different from the original business case.
Energy belongs in the equation as well, although it is better treated as an infrastructure consideration than as a direct plant-level charge. The International Energy Agency’s Energy and AI special report estimates data centers consumed about 415 TWh globally in 2024 and projects approximately 945 TWh by 2030 in its base case. Manufacturers may experience the associated costs indirectly through cloud consumption, vendor pricing or infrastructure investments. Across multiple plants and related AI use cases, these upstream costs should be visible when evaluating the broader economics of deployment.
Who owns the decision?
As AI evolves from analysis to recommendation and action, the more important liability question may not be “who owns the exposure?” The manufacturer ultimately remains responsible for what happens in its operation. The question is whether that exposure has been recognized and priced into the AI business case.
An AI recommendation may involve technology providers, models, integrators, data sources, employees and operating systems, but the consequences remain with the manufacturer. A quality issue, equipment failure, missed customer commitment or safety event still occurs within the manufacturer’s operations.
Therefore, decision rights, approval thresholds, auditability and accountability become essential operating controls. Human oversight reduces risk, and it also reduces some of the labor savings promised by autonomous agents. Manufacturers need to understand both sides—what autonomy creates in value and what responsible oversight adds back in cost.
AI also expands the cyber conversation.
Cybersecurity and data privacy introduce another layer of cost and risk. Manufacturing AI becomes more useful as it gains access to machine states, recipes, production orders, maintenance history, operator activity, quality information and proprietary process knowledge. Every additional API, application, plug-in, cloud workload, identity and third-party dependency can expand the environment that must be governed.
IBM’s 2026 breach research found that more than 20% of organizations reported a breach targeting AI models or applications; compromised APIs, applications or plug-ins and cloud misconfigurations affecting AI workloads were each cited in 27% of those incidents. IBM also reported 62% of AI-driven attacks in its study targeted critical-infrastructure sectors.
Data privacy deserves attention as operational AI combines employee, customer and proprietary information, creating questions about access, retention and appropriate use. Leaders need confidence in the integrity, provenance and protection of the data feeding AI. On the shop floor, a compromised recipe or manipulated machine instruction is not merely an IT incident. It can become a safety, quality, production, customer and enterprise-value event.
None of these arguments suggest manufacturers should slow innovation or AI adoption. I would argue the reverse: AI can create substantial value, and manufacturing may ultimately be one of its most consequential applications. The argument is for disciplined progression and a more complete business case.
AI can create substantial value, but value should be measured against the full cost.
A pragmatic approach is straightforward:
1. Start with focused use cases where the data requirements are understood and basic (static versus real time), the operational consequence is bounded, and measurable value can be demonstrated.
2. Establish trusted and contextualized data, and define who owns the recommendation and who authorizes the action.
3. Quantify not only technology and implementation, but also human oversight, revalidation, infrastructure, energy, security, privacy, governance and liability exposure.
4. Expand autonomy as confidence, measurable value and operational trust increase.
Manufacturing has experienced technology cycles where promising solutions accumulated integration, customization, infrastructure and support until the economics looked very different from the original business case. Remember the promises of Industry 4.0? AI should not repeat that pattern.
If manufacturers automate an existing process and fail to account for the data, compute, human oversight, cybersecurity, governance and liability required to sustain it, they may create a more expensive way of achieving essentially the same outcome.
The objective should not be to deploy more AI. It should be to create better operational and business outcomes because of AI. That requires a more complete question before scaling: not simply, “What can AI do for us?” but “What will it truly take and cost for us to trust it?”
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
