mh.MARKUS HIRSCHCONSULTING & BUSINESS DEVELOPMENT
INDUSTRIAL AI
Home/AI in Practice/Production

AI IN PRACTICE / PRODUCTION

Production

Manage variants. Make operations more predictable.

Prepare a controlled restart after a variant change Build a practical sequence for a bottleneck resource

Illustrative process scenarios — not customer references. Feasibility, benefits and approvals need to be established for the specific organisation.

01

ILLUSTRATIVE SCENARIO

Prepare a controlled restart after a variant change

A released product change affects tools, operations and inspections. An assistant brings the relevant changes together in a traceable restart pack.

UNDERSTAND THE AI

What happens behind the scenes.

TECHNICAL INSIGHT
POSSIBLE SOLUTION DESIGN

Generative AI · Retrieval-augmented generation · Document comparison

What does the AI actually do?

AI identifies operations, tools and inspection items described in change documents, matches differently worded information and drafts a justified restart checklist.

How the solution works

  1. 01
    Retrieve the right records

    Part number, revision, validity date and permissions limit retrieval from PLM and ERP. Keyword and semantic search find relevant passages despite different terminology.

  2. 02
    Interpret the changes

    A large language model (LLM) compares retrieved text and extracts changes into defined fields. Fixed rules check release status and revision relationships; ambiguous matches remain unresolved.

  3. 03
    Present an evidenced proposal

    Each proposed change links to its source and affected check. Specialists receive a reviewable checklist based on the retrieved evidence.

What value can it deliver?

Less document searching for manufacturing engineering, earlier visibility of conflicts and missing approvals, and potentially less first-batch rework. Compare actual changes to establish the effect.

A concrete example

Illustration: a material change is linked to a revised tool recommendation. The checklist cites both sources and flags the outstanding first-piece approval.

Technology, setup & operation

A pilot can usually use a pretrained LLM and a maintained search index. Retrieval-augmented generation (RAG) supplies relevant company records with each question. New revisions require index updates; training a proprietary model is not inherently required.

Illustrative solution design. The final architecture depends on data, systems and pilot evaluation.

Starting situation

A precision-component manufacturer introduces a new material variant. The drawing, tool list and work instruction were revised at different times. Before the first batch, production needs to establish which versions belong together and which approvals are still missing.

How the workflow could look

  1. Match the changes

    The assistant links the change notice, drawing revision, routing and tool version using the part number and effective date. Conflicting information becomes an explicit open item.

  2. Prepare the restart pack

    It drafts the changed operations, required tools and first-piece inspections. Each statement links to the relevant source passage.

  3. Record approval

    Manufacturing engineering and quality review the pack. Only their confirmed version is made available at the affected workstation.

Illustrative work product

A restart checklist for a specific part and revision, listing confirmed documents, outstanding approvals, owners and linked evidence.

02

ILLUSTRATIVE SCENARIO

Build a practical sequence for a bottleneck resource

Orders compete for the same machine. Runtime forecasts and optimisation under fixed constraints make alternative schedules comparable.

UNDERSTAND THE AI

What happens behind the scenes.

TECHNICAL INSIGHT
POSSIBLE SOLUTION DESIGN

Predictive AI · Regression model · Mathematical optimisation

What does the AI actually do?

AI estimates processing and changeover durations from completed jobs, learning how material, batch size, coating and the previous batch relate to actual runtime.

How the solution works

  1. 01
    Learn from actuals

    Completed orders supply input features and measured durations. A regression model, such as gradient boosting, learns an estimator; later orders form a separate evaluation set.

  2. 02
    Calculate sequences

    An optimiser combines predicted durations with calendars, capacity and approved combinations. It searches for feasible sequences against an agreed objective, such as lower tardiness within a setup budget.

  3. 03
    Explain alternatives

    The interface shows dates, setup effort and trade-offs for each option. Planners can adjust priorities and recalculate before approving transfer to MES.

What value can it deliver?

More realistic durations can reduce scheduling errors. Comparing alternatives helps avoid unnecessary changeovers and exposes the consequences of urgent jobs. Forecast error and actual setup time provide measurable evidence.

A concrete example

Illustration: one sequence avoids a cleaning changeover but delays a customer order. Both options use the same constraints, allowing planners to assess the explicit trade-off.

Technology, setup & operation

The forecasting model needs operational training data and comparison against simple baselines. The optimiser is a separate mathematical component, for example constraint programming. No LLM is required for schedule calculation. Changed operations require renewed checks of model accuracy and data distribution.

Illustrative solution design. The final architecture depends on data, systems and pilot evaluation.

Starting situation

In a coating plant, many orders share one oven. Material compatibility, batch size and cleaning requirements determine a workable sequence. An urgent order can trigger extra changeovers and put several committed dates at risk.

How the workflow could look

  1. Estimate effort

    A forecasting model estimates processing and changeover times from comparable jobs. Missing data and unusual orders are flagged.

  2. Calculate alternatives

    An optimiser respects approved material combinations, capacity and delivery targets. It compares the current schedule with alternative sequences.

  3. Review the impact

    Planners see expected dates, setup effort and at-risk orders for each option. A confirmed option is transferred to the planning system in a controlled step.

Illustrative work product

Alternative schedules with explicit trade-offs: fewer changeovers, less delay or a need for additional capacity.

CONTINUE EXPLORING

More processes.
More perspectives.

Quality managementComplaints managementCompliance & reportsLogisticsSupply chainService & self-service
All ten scenarios

FROM SCENARIO TO A PRACTICAL NEXT STEP

Which process deserves
a closer look?

Together, we can clarify the current process, available evidence and a useful pilot. My contribution: structure, priorities and alignment between business teams, IT and implementation partners.

Discuss a use case