Automating junior service work can weaken future expertise

Automating junior work with AI can also remove the work that develops experts

Ask a service team, “who here can diagnose that machine by the sound it makes?” If only two or three names come back, an efficiency decision about junior work becomes a question about future service capacity. Sound-based diagnosis depends on experience that a search tool may struggle to reproduce. Those who have it learned to interpret faults, compare possible causes, and recognize when the usual procedure would fail.

The assignments that build this experience produce two things: completed work and developing judgement. Automating the work leaves employers responsible for developing the person. AI can also speed up training, so employers can redesign routine tasks while preserving opportunities to learn. The decision is which activities will teach juniors the judgement that makes experienced technicians valuable.

Developing that judgement takes longer than an efficiency review. Think of five years spent learning to diagnose, or a recruit becoming hard to replace 20 years later: these illustrate development rather than fixed qualification periods. A warning about savings visible this quarter creating an expertise shortage around 2033 makes the same point; the date is illustrative. Immediate productivity and future competence operate over different periods, and that difference matters more when existing experts are approaching retirement.

Retirement increases the need to develop replacements

US manufacturing already has a relatively small younger workforce behind its experienced employees. Reporting published in June by the Association for Advancing Automation puts the median manufacturing-worker age at 43.9, with roughly a quarter of workers aged 55 or older and fewer than 8% under 25. Replacing an experienced employee therefore requires both a recruit and enough time for that recruit to become competent. That takes suitable work and access to experienced colleagues.

The scale of recruitment adds to the pressure. The Manufacturing Institute estimates that US manufacturing will need 3.8 million new employees between 2024 and 2033, with around two million positions likely to remain vacant. In the same June reporting, the American Welding Society projects a need for 320,500 new welding professionals by 2029, while more than 157,000 are approaching retirement. Employers need more recruits as the people capable of teaching them leave.

This demographic constraint extends across mature industrial markets. The OECD projects a roughly 9% contraction in Germany’s working-age population over the decade covered by its forecast. Japan’s working-age population fell from 87.3 million in 1995 to 73.7 million in 2024, with a further 31% decline projected by 2060. Conditions differ across Europe, the United States, and Japan, but manufacturers in these markets cannot assume a growing pool of potential technicians.

Service leaders have discussed retirement pressure for a decade; AI now adds a decision about how people enter those careers. Employers must decide whether to keep hiring into junior roles built around knowledge that software can reproduce. Retirement reduces available expertise, while fewer entry opportunities can restrict its replacement. Demographics make that choice urgent, but AI’s effect depends on whether companies use it to remove junior work or make that work more productive and useful for learning.

AI changes access to the work where learning begins

Employment patterns show how that choice can affect entry opportunities. In an August 2026 study revision, Stanford Digital Economy Lab examined payroll records covering the period since generative AI arrived. Its researchers found no evidence of broad economy-wide displacement, but they identified a substantial employment gap among workers aged 22 to 25 in occupations most exposed to AI. Experienced workers showed no comparable gap.

For the younger group, Stanford found employment was about 19% below the trajectory implied by less-exposed peers. This figure compares employment paths across groups. It does not describe a fall of that size from the exposed group’s own previous employment level. The gap had widened from roughly 15% in the earlier estimate associated with the July 2025 data vintage.

The gap arose mainly through reduced hiring rather than increased separations, which means employers can narrow entry opportunities without visible rounds of dismissals. Customer support was among the exposed occupations. That makes the finding relevant to service organizations as well as other knowledge-based businesses, where prospective juniors can lose access to the work through which they would otherwise begin developing experience.

How employers use AI matters within those employment patterns. Stanford distinguishes between AI that substitutes for human tasks and AI that complements employees: substitution was associated with falling entry-level employment, while complementary use corresponded to stable or growing employment. Declines were concentrated in roles involving codified knowledge, meaning information and procedures that can be explicitly recorded and reproduced. So deployment design affects access to work.

In industrial service, access to that work can start a longer learning process. A junior may initially contribute by retrieving procedures, handling standard questions, or resolving familiar faults. Repeated exposure can then teach them when a procedure applies, what evidence contradicts it, and when to involve an expert. Reducing hiring into the initial work can also reduce access to these later learning opportunities; this is a workforce-planning risk, distinct from Stanford’s measured employment gap.

AI can also make people entering that work more productive. Boston Consulting Group’s June 2026 field-service executive brief estimates productivity improvements of 20% to 30% from AI-assisted dispatch and troubleshooting in the field. BCG also identifies faster training for junior mechanics among the returns. As a consultancy selling advice on AI and operational change, BCG benefits from companies investing in those activities.

Faster training gives employers a reason to redesign junior roles around AI assistance. BCG’s estimates concern service productivity and training, while Stanford’s findings concern employment patterns under different uses of AI. Each informs a different part of the management decision: how quickly juniors can contribute, and whether they retain access to work. The next question is what they need to experience to learn to diagnose unfamiliar problems.

Make expert judgement visible while people learn

Some of the knowledge needed for diagnosis never reaches a written procedure. Will Healy III of Teradyne Robotics, the parent company of Universal Robots, estimates in the Association for Advancing Automation’s June reporting that a product may depend on two to five undocumented actions. These include recovering from a fault, managing variation, and maintaining quality under production pressure. Teradyne Robotics sells automation systems, so it benefits when manufacturers invest in automation and the knowledge needed to operate it.

To uncover those actions, Healy advises observing failures, urgent orders, and recovery work. Routine production can conceal an experienced operator’s decisions because everything is proceeding as expected. A breakdown reveals what the operator checks, changes, and rules out. An interview away from that situation may fail to bring the same knowledge to mind.

Even when experts explain their work, repetition can make important actions seem too obvious to mention. Richard Meyer, who runs customer service group training at FANUC America, says some plant tasks are understood by only one or two people. After 20 years of repetition, those experts may stop noticing which individual actions affect the outcome. FANUC America supplies industrial automation and training, giving it a commercial interest in employers investing in both equipment and the skills to use it.

Meyer addresses these omissions by making the newcomer responsible for writing the instruction. The newcomer must ask about actions the expert takes for granted, making missing steps easier to detect. Writing also becomes part of learning because the employee has to understand the sequence well enough to explain it. The assignment captures knowledge while developing the person who will use it.

Physical work can require a different way to make knowledge visible. Matt Malchano, vice president of software at Boston Dynamics, describes the knowledge at risk as “hands-on process-oriented knowledge.” He means complex, variable tasks learned through adaptation and direct teaching between workers. Boston Dynamics sells robots, so it benefits from methods that make human work available for robotic learning and execution.

Malchano’s teams capture demonstrations by having operators wear goggles and gloves while a robot mirrors their movements. The demonstrations become training data, recording actions that workers may never have expressed in written instructions. This method captures physical execution, while observation and questioning expose the decisions surrounding it. Together, these methods can give learners access to both actions and reasoning.

Diagnostic capacity determines which service promises remain deliverable

The value of that reasoning becomes clear when a technician resolves a problem that a manual cannot settle. Consider an experienced engineer finding the cause in 20 minutes: the customer is buying the ability to interpret the situation and act. For a manufacturer supporting equipment with a 20-year service life, developing that competence is part of maintaining delivery capacity throughout the equipment’s life.

If diagnostic capacity weakens, customers can feel the effects in several connected ways. First-time fix performance, the share of problems resolved during the initial service intervention, can deteriorate as diagnosis takes longer. Off-hours telephone support can also become less effective when fewer people can interpret incomplete descriptions and decide what to do next. These are service risks to manage, rather than measured outcomes from a particular customer case.

A stopped production line at two in the morning makes the risk tangible. Someone with six months of experience and search access may find the relevant procedure quickly, yet struggle to judge whether it fits the fault. Someone drawing on six years of remembered cases may recognize a misleading symptom or an unusual combination of conditions. Search retrieves information; accumulated experience helps the technician decide how to use it.

Customers can carry those experiences into renewal discussions, especially when they judge a contract through the last three visits. Uptime guarantees, response windows, and availability contracts depend on people capable of delivering the promised service. Staffing and training decisions therefore affect the risk attached to those commitments. Developing diagnostic judgement belongs in service-delivery planning and decisions about contractual exposure.

Build learning into knowledge capture and AI assistance

Managing that exposure requires five connected practices: continuing knowledge capture, deliberate copilot design, revised junior roles, formal training routes, and flexible access to veterans. Each meets a different need, from making expert decisions visible to ensuring someone is present to learn them. The practices need owners and time within everyday work so they continue as equipment, employees, and service demands change.

Start with knowledge capture as a continuing learning activity. Manufacturers have already tried mentoring schemes, video collections, and knowledge bases, and some repositories stopped receiving updates. As equipment, procedures, and operating conditions change, an unattended collection becomes less useful. Assign people to observe important work, record what changed, and turn the result into material someone uses.

Handling exceptions gives that assignment a practical focus. Combine Healy’s emphasis on failures and recoveries, Meyer’s newcomer-written instructions, and Malchano’s use of demonstrations where actions are difficult to describe. The junior should be able to explain what evidence caused the expert to change direction. For the machine in the opening question, that means learning what the sound suggests, which checks follow, and what would overturn the diagnosis.

AI assistance should create further opportunities to practise that reasoning. A technician copilot, an AI assistant used during work, can present possible diagnostic routes for a junior to assess. The junior then considers alternatives, chooses a check, and explains why the result supports or weakens a proposed cause. An experienced colleague can review both the decision and the reasoning behind it.

A finished recommendation can still be valuable when speed and standardization are the immediate priorities. The training decision is where the technician also gets practice evaluating how a diagnosis was reached. Over a five-year horizon, repeated assessment of alternatives can support growing independence, while repeated acceptance of answers can encourage reliance on the tool. This gives employers a reason to build learning opportunities into each workflow.

Those opportunities need explicit boundaries and responsibilities. Some tasks can be completed automatically; others can require a junior to review a proposed diagnosis with an experienced colleague. Managers should identify where independent reasoning is practised and who assesses it. Configuring the copilot then becomes part of training design, with implications extending across a ten-year workforce-planning horizon.

Keep entry routes open and give them structure

Training design needs positions in which inexperienced employees can contribute. BCG’s identification of faster training gives employers a return to consider alongside time saved on dispatch and troubleshooting. A service operation can automate standard information retrieval and direct more junior attention toward interpreting symptoms, reviewing recovery work, and handling progressively harder cases. The role changes as the employee becomes productive in a different way.

A redesigned role also addresses the access problem visible in Stanford’s employment findings. More intellectually demanding, cleaner manufacturing work may attract younger applicants, but those applicants still need a route into it. Employers should examine whether an efficiency plan leaves genuine positions where inexperienced people can contribute and develop. That connects hiring decisions to the work through which future diagnostic capability will grow.

Entry positions become more useful when they sit within formal development routes. FANUC runs certification, apprenticeship, and training academy programs, providing an example of structured technician development. These programs create an organized route into competence, while Meyer’s instruction-writing method makes a specific assignment teachable. An employer can combine an overall development program with learning activities built into everyday work.

Establishing that route before recruitment pressure peaks gives the activities time to work. A training arrangement needs participants, instruction, supervised practice, and assignments in which people can apply what they learn. Waiting until an experienced technician leaves compresses those activities into a period when service capacity is already under pressure. Earlier preparation gives managers more room to plan progression and allocate supervision.

Retain expert access and audit who is developing

Formal development still depends on access to people who can teach difficult judgement. Part-time mentoring, remote consultation, and assignments focused on difficult cases allow experienced engineers to contribute without retaining their former full-time workload. Discussing these arrangements before complete retirement gives both sides time to define the work. Its purpose should include developing other employees while keeping difficult-to-replace expertise available.

Consider a retired engineer available through remote video access two days a week. Connecting that engineer to a live problem and the technician learning to solve it can reveal which symptom changed the diagnosis or why a seemingly reasonable test would be unhelpful. Recording and reviewing the reasoning can extend learning beyond the immediate call. The arrangement needs time for explanation as well as resolution.

Healy’s assessment is that serious efforts to preserve knowledge create value, while neglect allows that value to disappear. Managers can make the effort observable by assigning difficult-case experience, naming who reviews the learner’s reasoning, and tracking how responsibility progresses. These records give an operations review something more useful than a count of completed training sessions. They show who is gaining experience with decisions the service organization needs people to make.

At that review, ask how many employees can diagnose through judgement and what their average age is. Then ask how many people the organization hired last year into roles where that judgement develops. Request the actual roles, learning activities, and review responsibilities behind the headcount. A small expert group combined with zero hiring into judgement-building work warrants investigating who will acquire the competence needed to deliver existing service commitments.

Recap

Automating junior work can improve productivity without weakening the next generation of technicians, but that outcome requires deliberate choices. The work removed from a role may also be where someone learns to interpret symptoms, test a diagnosis, and recognize when a procedure does not apply. Efficiency plans need to account for how that experience will be replaced.

For business leaders, this makes AI deployment, workforce development, and service commitments connected decisions. A productivity case should identify which tasks will disappear, where juniors will practise diagnostic reasoning, and who will supervise their progress. Retirement planning should establish how long experienced technicians will remain available to teach, not simply when their positions become vacant.

Measure the resulting capability alongside time saved. Track who can handle increasingly difficult cases, explain their decisions, and work with less supervision. Those measures help establish whether AI is building service capacity or leaving the organization more dependent on a shrinking group of experts.

The objective is not to preserve every junior task. It is to preserve a credible route from entry-level contribution to independent judgement. Companies that maintain that route can use AI to improve service today while developing the people needed to deliver their commitments tomorrow.