Experienced operators are retiring, and much of what they know was never written down. According to McKinsey (2025), the share of US manufacturing workers over 55 more than doubled in twenty years, and retirement rates have surged. In a process plant, that knowledge covers how a unit behaves during a start-up, which early signs come before an upset, and why a procedure is written the way it is. When those people leave, the plant relearns it slowly, often after an upset.
AI can help keep that expertise in the plant, provided it is captured in a form operators use on shift. The four steps below set out a practical order of work: assess the gaps, choose and integrate the tools, design the training, then measure and improve.
Step 1: Assess knowledge gaps and data readiness
Before choosing any technology, define which expertise matters most and where it is at risk. A structured knowledge assessment maps who knows what today, how that knowledge moves between shifts, and where it is thinning as specialists near retirement. Every undocumented technique that leaves with a retiring operator widens the skills gap.
Start by sketching the current knowledge pathways: shift logs, informal mentoring and engineering archives. Trace each one to the subject-matter experts who still hold the tacit pieces. These include troubleshooting intuition, the small moves that keep a unit on target, upset playbooks, and the history behind critical procedures. Some plants find that a single operator holds a unit's entire start-up procedure.
At the same time, check data readiness. Review sensor coverage, historian depth and tag metadata so that the data is clean enough for models to learn from. Poorly labeled tags and missing calibration records undermine every later analytics effort.
Common pitfalls at this stage:
- overlooking the informal networks between operators and shifts
- relying on manuals alone
- skipping data cleansing
- leaving front-line operators out of the assessment
Set a baseline before moving on. Practical measures include the share of high-risk tasks that are undocumented, the time taken to complete a knowledge audit, historian data-quality scores, and a red, yellow or green risk rating for each knowledge area.
Step 2: Select and integrate the tools
With the gaps mapped, choose technology that captures tacit expertise and puts it in front of operators when they need it. These criteria hold up in practice:
- Learning tools over static documents. Prefer tools that learn from historian data, shift logs and operator actions to documents that have to be updated by hand.
- Natural-language access. Conversational search lets an operator ask a question such as "Why is my compressor surging?" and get an answer tied to the plant's own context.
- Decision support. Setpoint recommendations guide the next move, and anomaly detection flags developing problems before they escalate.
- Closed-loop control where the unit is well modeled. Once a model has captured how a unit responds, closed-loop control applies that knowledge every minute and on every shift, inside limits set by plant engineers.
- Integration. On existing plants, connect historian, DCS and maintenance management (CMMS) data without disturbing the control strategy. On new plants, design the data pipelines in from the start.
- Selection factors. Weigh scalability across units, deployment flexibility, cybersecurity compliance and fit with existing systems.
- Business focus. A simple evaluation matrix keeps the discussion on operating value rather than feature lists.
Start on one high-value unit, measure results with metrics such as mean time to repair, then extend to other units. Expect challenges around data integration, operator trust and security requirements. Handle them through transparent governance and models whose recommendations operators can inspect.
Step 3: Design AI-enhanced training and knowledge sharing
The tools pay off when front-line teams use them as a matter of habit. Static PDFs and classroom lectures cannot keep pace with high-hazard, high-throughput units, so training has to develop alongside the workforce and surface the right insight at the moment it is needed.
A simulator built on a model of the plant's own operating history lets new operators practice start-ups, rate changes and upset recovery without touching the DCS, and each drill reflects how the actual unit has responded. Short refresher modules on routine tasks, such as pump alignment or catalyst change-out, reinforce that operator training from day to day.
Capturing the expertise that feeds these tools takes a deliberate workflow:
- Hold structured elicitation sessions while veterans are still on site.
- Record video walkthroughs and annotate the key decision points.
- Tag each knowledge asset to the historian trends or setpoint changes it explains.
- Break complex jobs into steps through detailed task analysis.
- Use those tags so the system can offer relevant guidance in context.
- Keep all captured expertise in a version-controlled knowledge base.
- Have peers review content for accuracy before release.
Organizational factors matter as much as the tools. Reward engineers who document a clever workaround as much as those who close a production gap. Visible sponsorship from plant leadership signals that upskilling is part of the job. Present the tools as support for operator judgment, ask operators for feedback, and measure progress.
Protect the content as well. Encrypt recordings, restrict access by role, and scrub sensitive process parameters, in line with control system cybersecurity practice. Use analytics to tailor learning paths and to keep the knowledge base current.
Step 4: Monitor, measure and improve
Once the program is live, keep it relevant by tying daily performance back to the expertise held in your systems. Start with one KPI dashboard that everyone can see, from the board operator to senior leadership. It should combine three kinds of measure:
- operating continuity, such as uptime and error rates
- learning, such as training completions and assessment scores
- knowledge-base health, such as article views, search success rate and content freshness
Together they show where captured knowledge is improving performance and where it is still leaking away.
To keep improving:
- Retrain and validate models regularly, and update them as sensors are added and process constraints change.
- Collect user feedback through in-app prompts and regular pulse surveys.
- Use analytics to find the most-used procedures and the questions operators ask most often.
- Hold an annual knowledge-transfer review, run by a cross-functional committee that keeps the program aligned with business goals.
- Feed findings into the plant's continuous improvement program so that insights become funded projects.
- When metrics stall, audit data quality, run stakeholder workshops, retrain models and revisit incentives.
It also helps to benchmark maturity along the way. At the reactive end, a plant tracks lagging indicators after incidents. Further along, captured knowledge sits in models that recommend moves to operators. At the autonomous end, the critical decisions run in closed loop, inside limits set by plant engineers, and the plant's expertise is applied on every shift.
How the Imubit Platform applies
Competency is one of the five areas where the gap between the plan and the plant shows up, alongside safety, reliability, margin and compliance. The Imubit Platform applies Map, Model and Control across refining, chemicals and polymers, cement, and mining and minerals. Map traces each gap to the recurring decisions behind it. In Model, process engineers build Deep Learning Process Control® (DLPC®) models of how each unit responds, learned from the plant's own history, in DLPC® Builder with no Python and no machine-learning background. That model holds part of what experienced operators learn over years at the board, and DLPC® Sim App lets the team see the result of a change before they make it. Control takes the critical decisions to closed loop on top of the existing APC and DCS, with operators able to take over at any time. See how the platform works.
