Gains:
- Ability to accurately define and calculate production KPIs such as OEE, cycle time and WIP
- Ability to configure dashboard design, discrete event simulation and scenario analysis with AI
- Ability to verify simulation and dashboard outputs with input assumptions and real data
Industrial engineers make decisions based on measurement, not intuition. Properly defined KPIs (Key Performance Indicators) keep the pulse of a production system; Simulation, on the other hand, asks "what if?" without stopping the real system. answers the question with confidence. AI is an accelerator in both: it drafts KPI definition and dashboard design, builds the skeleton of the simulation model, interprets scenario comparisons. However, both KPIs and simulation are susceptible to the traps of “wrong definition” and “wrong assumption”. In this unit, we will cover production KPIs, dashboard design and discrete event simulation with the discipline of verification.
Defining the Right KPI Correctly
An incorrectly calculated KPI is more dangerous than no KPI at all because it gives false confidence. The most common production KPIs:
KPI
Description
Attention
OEE
Usability × Performance × Quality
The three components must be calculated separately
Cycle time
Time between pieces
mixing with tact
Delivery time (lead time)
Order → delivery
different from cycle
WIP
semi-finished product in process
bound by Little's law
Waste/scrap ratio
Incorrect / total
Reserve rework
On-time delivery (OTD)
On time / total order
Define partial delivery
OEE example: If Availability is 90%, Performance is 95%, Quality is 98%:
OEE = 0.90 × 0.95 × 0.98 = 0.8379 ≈ 83.8%
A common mistake is to add these three numbers and average them; OEE is multiplication, so each component compounds the result.
Little's Law establishes the WIP-throughput-flow time relationship and is very powerful:
WIP = Throughput × Flow time Example: there are an average of 30 pieces (WIP) on the line, if the output is 6 pieces/hour Flow time = WIP / Output = 30 / 6 = 5 hours
Tip: When you have AI calculate a KPI, also print its definition. "What is OEE, with what formula did you calculate it?" ask. AI sometimes gives only availability instead of OEE or skips the quality component; Don't trust the number without seeing the definition.
KPI Dashboard Design
A good dashboard doesn't show a lot of numbers; Answers the right questions. AI is helpful in outlining which KPIs make sense for which audience and how to group them.
Role: You are an industrial engineer with experience in production analytics. Task: Design an admin dashboard for a CNC shop. Specify:1. Up to 6 KPIs (definition and formula for each).2. Target value proposition and “good/attention/bad” thresholds for each KPI.3. Which KPI will be displayed on which chart (trend, indicator, chart).4. 3 management questions the dashboard answers.Rule: Avoid KPI inflation; Explain that each indicator serves a decision. Don't give a made-up target value, say "it should be adjusted according to the sector".
Weak Prompt / Strong Prompt
Weak prompt:
Make me a production board.
If there is no target audience, data source and purpose, AI will dump a general list; No decision is made.
Powerful prompt:
Design a board for the shift supervisor, which will be read in 5 minutes every morning. Purpose: where was the loss in yesterday's shift, what should be paid attention to today. Maximum 5 KPIs, each with definition + formula + threshold. Hourly production, downtime and wastage records from my data source, MES. Avoid excessive detail.
The second prompt clarifies the user, the moment of use, and the data source; The result is a board that is actually used.
Discrete Event Simulation (DES)
Simulation is a way to experiment without disrupting the real system: what happens to the queue when you add a new machine, how is output affected when a shift changes? Discrete event simulation (DES) models the system as a sequence of “events” (part arrival, process finish). Basic inputs: inter-arrival time distribution, service time distribution, number of resources, queue discipline.
Role: You are an industrial engineer experienced in simulation.Task: Write a framework to model the following system with Python SimPy:- Parts arrive every 4 minutes on average (exponential distribution)- Single machine, processing time average 3.5 minutes (exponential)- Queue FIFOWhat I need to measure as output: average waiting, queue length, machine utilization rate.Explain the code with comments; write the distribution assumptions at the beginning.Also: what analytical check can I do to verify the result (e.g. comparison with M/M/1 tail formulas)?
The last line here is critical: a good simulation is compared to an analytical check if possible. The simple system above is an M/M/1 queue; The simulation result should hold with theoretical values such as the utilization rate ρ = 3.5/4 = 0.875 and the average number in the system L = ρ/(1−ρ) = 0.875/0.125 = 7.
Caution: Simulation output is only as good as the input assumptions ("garbage in, garbage out"). AI sometimes assumes the distribution without verifying it or presents the result of a single run as the hard truth. Run multiple runs (replications), calculate confidence intervals, and verify input distributions with real data.
Mini Case: Simulation with False Hypotheses
On an assembly line, management thinks adding a second packing station will kill the queue. Industrial engineer Deniz builds a DES model with AI. The first result says "the second station reduces the queue by 70%". But when Deniz checks the input distribution, he sees that the real arrival data is not exponential, but has an "explosive" structure that concentrates during the lunch break. When we run the model again with the correct distribution, it turns out that the second station does not solve the real bottleneck, and the problem is in the break planning. Simulation saves recommending an investment unnecessarily. Lesson: AI built the model quickly but suggested the wrong decision with the wrong assumption; It was the person who put the correct data.
Common Mistakes
- Misdefining the KPI: Adding up OEE components or confusing cycle and lead time.
- KPI inflation: Putting dozens of indicators on the dashboard that do not serve the decision.
- Relying on one run: Running the simulation once and assuming the result is certain; not replication.
- Not validating the input distribution: Not comparing the default (exponential, etc.) distribution with the real data.
- Bypassing analytical checking: Not cross-checking the simulation with the theoretical queuing formula in simple systems.
In summary
- The power of KPI comes from correct definition; OEE is the product, Little's Law connects WIP-flow-output.
- A good dashboard answers the right questions, not a lot of numbers; The user, instance, and data source must be clear.
- DES tries scenarios without breaking the real system; The input distribution determines the quality of the result.
- Validate the simulation by replication (confidence interval) and, if possible, by analytical formula.
- AI builds model and board quickly; It is the person who puts the assumptions correctly and interprets the result.
Application task
Choose a production/service system. First you have the AI design a shift dashboard with 5 KPIs or less; Have the definition and formula of each KPI shown and verify that the three components are multiplied if OEE is present. Then have the AI print a SimPy skeleton for a simple queuing system (single server) and print the input distribution assumptions at the top. Critical verification: consider the system as an M/M/1 queue and manually calculate the utilization rate ρ and the average number L in the system, compare with the simulation result. Finally, suggest a scenario change (adding a second server) and criticize the AI's result by questioning the realism of the input distribution.