The ageing of the manufacturing workforce

Manufacturing has been:  

• Losing talent to software for decades. The boom in enterprise software began with the rise of ERP and digitalization of banking – in the late 80s and early 90s. The rise in software was sped up by the dot com boom at the turn on the century. Spurred by the same forces, consumer software boomed as well. The iPhone launch in 2007 birthed a whole new era of software, with the 2010s boom in social media and e-commerce. As software grew, increasingly bright young analytical mind chose to join undergraduate programs in computer science and software engineering rather than electrical, mechanical, or chemical engineering. Driven by the same incentives, recent graduates with majors in traditional engineering chose to apply their skill to software engineering challenges.  

• Considered an inherently inferior employment proposition. Software has the advantage of low to zero marginal costs that allow companies and teams to grow fast. The software industry affords the young recruit a shorter route to affluence, even wealth. Software clusters tend to be in urban areas – favoured by youth. Oil and chemicals are located often in locations far from the urban hot spots of popular culture.   

• Struggling with a reputation problem. Energy transitions take many decades. Refineries will need rookie plant managers for a long time. But the talent oil and gas is trying to recruit increasingly favours alternative energy sectors (if they were to choose the hardware realm at all). The young are more likely to associate energy and chemicals with negative media reports such as oil spills. This image issue has exacerbated the talent acquisition challenge.  

The industry is exploring options and agentic AI is likely to emerge as the most critical tool 

The ageing workforce problem is interesting in that it’s simultaneously hard to solve and widely acknowledged. Executive leadership, the CIO’s office (often the CDO’s office as well), and the HR function are trying:  

• Organizational measures to retain and promote knowledge sharing. First, it’s hard to expand the pipeline of new talent – the inflow, and therefore industry is trying to slow the outflow. Tenured employees are being nudged towards deferring retirements through incentives, policies, and tweaks to the culture. One charge against the industry is knowledge hoarding and much implicit rather than explicit knowledge. Industry is correcting this through better formal training, mentoring, shadowing and overall – better knowledge management practices.  

• Exploring offshore units for planning, scheduling, and other desk jobs. Some of the largest oil and petrochemical companies are centralizing the parts of maintenance and reliability that can be executed out of lower cost offshore locations. There are TA planners and permit writers working out of Bangalore, India servicing, respectively, turnarounds and control-of-work teams at refineries spread across the globe. In every location, mechanical, electrical, chemical, and instrumentation engineering talent is dwindling. But with the sheer size of India’s talent pool, there’s scope for significant augmentation of refinery-specific talent.  

• Better knowledge management tools. Knowledge in asset-heavy industries is locked in a large corpus of unstructured sources – spanning PDFs, email, spreadsheets, and sometimes even paper. Making artifacts available to an internal search engine is a big step. This basic step of digitization will be major for many plant operations. At the time of writing, the bulk of asset-heavy manufacturing operations is stuck at this level. The next step is to extract entities from unstructured text (past plans, email, spreadsheets, permits, inspection reports) and images (such as pfds, p&IDs and GADs). The more advanced step from that point on is to build a graph linking said entities. At this realm of maturity, there exists an institutional, accessible map of how plant assets are connected in a spatial and logical way; how they interconnect terms of material and energy flows. Ideally, these asset tags only need to be connected graphically to maintenance work streams. After all that has been done, the next step is agents.  

Agents address critical issues that plague asset-heavy industry  

A major aspect of running an industrial plant is planning – for capital projects such as turnarounds. A turnaround is when a plant is taken offline for three to five weeks for a complete overhaul – repair, replacement, and clean-ups. Costing anywhere between a few million and US$100 million, turnarounds (often called TAs or TARs in the downstream oil industry). The stakes are high enough for oil companies to report TAs in annual reports. Naturally, planning is critical for such a major project. Scope estimation is all important. If the team builds too much into the plan, the TA is at the risk of going over budget and stretch beyond the designated date of the plant resuming operations. A typical refinery processes a few hundred thousand barrels of crude a day, the financial impact of even a couple of days of delay is huge.  

Conversely, if scoping is too careful and misses some critical maintenance tasks, and overlooks certain assets, production and safety will likely suffer.  

A careful balance needs to be struck. Turnaround veterans cite scope planning as the top determinant of project success – for a project type where the cost of failure is in millions.  

Planning requires in-depth knowledge of which assets need inspection, repair, or replacement. In addition, the effort estimate for every kind of maintenance activity for each asset type needs to be known. At the scale of a refinery, the number of irreducible asset units that need to be planned for are in the thousands. Every line in a work pack requires knowledge of very high specificity. A particular kind of pump requires a particular kind of inspection and repair work. This requires working at a height, a kind of work that requires a scaffolding of equivalent height, and therefore a specific protocol in accordance with the SoP.  

This process needs to be replicated across many thousands of assets and thousands of workstreams.  

Naturally, planning is a tall order.  

To continue pulling off such complex projects and planning tasks sustainably over the next decade and beyond, a few things must be true.  

First, there must be a mature knowledge management system, mature information lifecycle management policies, and ubiquitous in-context information access.  

Second, knowledge of the plant should be widely distributed among plant personnel, including those who would be in the high noon of their careers in 2035.  

Neither is true for asset heavy industries.  

Which is why the industry needs agents.  

An agentic framework for industrial operations needs to check certain boxes:  

• Tightly defined scope – every agent does one thing and one thing only. Among the agents Maximl has built in support of Turnaround digitalization is a rotary equipment agent, a BOM extractor, a scope maturity assessor, a job/work pack creator, an emergent work predictor. Such specificity increases the likelihood the agent built aligns with a common, well recognized task, and does not hallucinate.  

• An agent greatly minimizes data entry tasks. Building a job list or work pack requires a high degree of plant-specific knowledge. Planners with such depth and expertise are rare. All their time must go into aspects of planning that require lived experience around the plant.  We believe every plan should be informed by all past experiences locked in documents, and the task of scouring for information from the enterprise corpus, and the task of populating the current plan with the information thus retrieved should be offloaded to AI. In 2026, with the current state of AI maturity, there’s no reason why a human (and a very skilled one at that) should have to wade through the initial, non-value-adding 60 percent of it. This frees up bandwidth for the planner to focus on the critical 40 percent – the parts that require hands on experience, discernment, and judgement.  

• Ideally, a set of agents is collectively equivalent to a rookie engineer. If it were the case that young engineering talent was abundant at oil refineries, AI agents (or offshore GCCs) wouldn’t be necessary. The tenured planner would have a team of mechanical, electrical, instrumentation, and chemical engineers to scour the internal corpus – OEM documentation, and regulatory/industry-best-practices literature, and perform the manual data entry tasks. But since such abundance of young talent no longer exists, oil refineries need a library of AI agents.  

• Agents have no autonomy; humans retain control. AI agents are often deployed in white collar environments where the output is digital, and the consequences of failure are limited. This is not the case in industrial operations. AI agents serve as informational aids. Accountability remains with the individual, such as the Turnaround planner.  

• The operating environment needs to be within the enterprise perimeter. The risks, spanning regulatory and operational, are too many. Therefore, all software would (typically) need to be hosted on a private cloud.