What a Sensitivity Analysis Should Show Before Approval
A good sensitivity analysis shows whether the recommended option still wins when the assumptions behind it turn out worse than planned. In a deterministic factory model you do this by running variants with different fixed assumptions: a cycle time 10 or 20 percent longer at one station, a later order release, a shorter shift, a demand peak. Before approval the pack should show, for each step, throughput, on-time delivery, WIP, lead time and where the bottleneck moves, plus the break-even value at which the recommended option stops winning. If it cannot say which assumption breaks the plan first and who owns that assumption, it is not ready for a signature.
Why factory plans break after approval
Most plans that fail on the floor looked sound in the meeting: cycle times from the routing sheet or a supplier quote, a steady mix, the full shift pattern, orders released on time. Then the line runs. A new station is slower than quoted. Orders reach the line a day late. A peak month arrives earlier than planned.
The usual reaction is to blame execution. Often the approved result depended on one or two assumptions nobody tested. The plan held only if every assumption came true.
A sensitivity analysis finds those assumptions before the money is spent and checks whether the result can carry a decision.
Base run, variants and sensitivity
Three things often get mixed up in approval packs.
- A base run uses one set of assumptions and gives one answer. It is useful while you build the model and too thin for approval.
- A variant is a different option: another layout, a robot cell instead of a manual station, an extra shift. Variants are compared side by side on a copy of the base model.
- A sensitivity analysis keeps the option and changes one assumption you are not sure about, such as the cycle time of a new cell, the day orders are released or demand for one product family. It shows how the result and the ranking respond.
Variants tell you which option is better on your assumptions. Sensitivity tells you whether the choice depends on an assumption that might be wrong.
Which assumptions to move, and how far
Pick the assumptions the decision leans on hardest, not every parameter in the model. These four cover most capital decisions on a line or a layout:
| Assumption | Typical steps to run | Why it matters |
|---|---|---|
| Cycle time at one station | Base, +10 percent, +20 percent | Supplier quotes and ramp-up are often optimistic |
| Order release | On time, one shift later, one day later | Late release squeezes the time left before due dates |
| Shift length or working days | Planned pattern, one shorter shift, one day fewer | Shows how much slack the plan really has |
| Demand | Planned mix, peak week for one product family | A peak changes which station limits the line |
Choose the steps from what the plant has seen, not from a fixed rule. If a similar station ran 15 percent slower than quoted during its first months, test at least that far, and say where the number came from.
Lost time you expect for reasons outside the model, for example from stoppages, can be tested the same way: as a fixed assumption, such as a shorter effective shift or a slower cycle at the station concerned.
What a sensitivity pack should show
A weak pack moves many parameters and shows a colorful chart with no link to the decision. A useful pack shows six things.
| What the pack shows | Question it answers | Weak version to avoid |
|---|---|---|
| Ranked assumptions with owners | Which assumptions matter, and who is accountable for each? | Fifty parameters, no names |
| Tested steps vs plant history | Did we test as far as the plant has actually moved? | The same small step on everything |
| Throughput, on-time delivery, WIP and lead time per step | How does each result respond, not only output? | Only the total output figure |
| Bottleneck shift | Does the result just get worse, or does the bottleneck move? | No mention of where the line is limited |
| Combined case | What happens when two assumptions turn worse together? | One assumption changed at a time only |
| Break-even point | At what value of the assumption does another option become better? | Ranking shown only for the base case |
The break-even point is the most useful line for approval. Here is how a one-factor-at-a-time sweep finds it. The numbers below are illustrative, not from a real model. Option A is a robot cell, option B a second manual station. Only the cycle time of the robot cell changes, so option B stays the same in every run.
| Robot cell cycle time (illustrative) | Option A output per week | Option A on-time delivery | Bottleneck in option A | Option B output per week |
|---|---|---|---|---|
| Base (supplier quote) | 1,040 | 99% | Packing | 980 |
| +10 percent | 1,010 | 97% | Robot cell | 980 |
| +15 percent | 980 | 95% | Robot cell | 980 |
| +20 percent | 950 | 92% | Robot cell | 980 |
In this illustrative case, option A wins as long as the robot cell runs no more than about 15 percent slower than quoted. At +10 percent the bottleneck has already moved from packing to the cell, which is a warning the output figure alone would hide. The +15 percent run was added after the first sweep, to narrow the step where the ranking flips. Now the approver knows exactly what to check with the supplier, and who should check it.
See how a ranking holds when an assumption changes
Think of one assumption your team argues about, such as the cycle time of a new cell. The demo shows how variants with different fixed assumptions are compared and where the preferred option stops winning.
Move assumptions together, not only one at a time
Bad weeks rarely bring one problem. A late order release often comes with a rush order. A new product brings a slower cycle and a demand peak in the same month. One-factor-at-a-time runs show which assumption matters most, but they miss the cases that cost the most.
Add at least one combined case built from something the plant has lived through. For example: a peak week for the high-variant family, the new station at +20 percent cycle time and one shift shorter on Friday. Run every option through it.
Also watch where the constraint goes. When one assumption moves, the bottleneck often moves to another station or to transport. A result that only says "output drops" hides that the forklift is now the constraint, and the fix is different. For running every option through the same stress cases, read How to Compare Resilience, Not Just Throughput.
Repeatable results, not probabilities
The simulation in DBR77 Digital Twin is deterministic. Times are fixed nominal values, and the same input always gives the same result. That has two consequences for a sensitivity pack.
First, every difference between two runs comes from the assumption you changed, not from noise. You do not need many repetitions to see an effect, and anyone can rerun a case and get the same numbers.
Second, the results are not probabilistic. The model does not draw random values, so it gives no confidence intervals and no probability that the plan holds. The range in the pack is the range of assumptions you chose to test. That is why the steps, the reason for each step and the owner of each assumption belong in the pack. See Averages vs simulation over time.
When is a result strong enough to act on?
A model can always be refined, so a request for more runs can postpone the decision for months. Five checks tell you when to stop.
- The option set is closed. The committee chooses between named options, and no new idea enters during the meeting.
- Every option faced the same tests. Same assumption steps, same combined case, same simulated period.
- The ranking is stable. The recommended option still wins, or loses only slightly, across the tested steps. Where it flips, the break-even point is known.
- The top assumptions have owners. Someone in engineering, operations or procurement stands behind each top assumption and the steps it was tested over.
- There is a trigger to rerun. You know which change, such as a new supplier quote or a different product mix, forces a new run before the next payment.
If all five pass, the result is ready for commitment. If not, it is still exploration, and nobody should sign a purchase order on it.
An estimated assumption can support a decision if it was tested over wide steps and the recommended option still wins at the bad end. What sensitivity cannot fix is a wrong model boundary. If the real constraint, such as a shared warehouse or a paint shop outside the model, is not in the model, no amount of testing will show it.
How to present sensitivity to the approver
Three plain sentences serve the approver better than a technical chart. For the illustrative robot cell, they could read:
- "Two assumptions matter: the cycle time of the new cell and demand for family C."
- "If the cell runs more than about 15 percent slower than quoted, the manual station gives more output and we start to miss the weekly delivery target."
- "Procurement confirms the cycle time in the supplier test before the second payment."
The first names the drivers, the second gives the break-even point and what breaks, and the third says who checks it and when. Where this fits in the comparison of several investment options is described in How to Compare CAPEX Options When Every Scenario Looks Plausible, and how to put it on one page for the board in How to Present Simulation Results to the Board.
FAQ
What is sensitivity analysis in factory simulation?
It is a set of runs in which one assumption, such as a cycle time, the order release, the shift length or demand, is changed in fixed steps while everything else stays the same. It shows how throughput, on-time delivery, WIP and lead time change, where the bottleneck moves and whether the recommended option stays the best one.
How many assumptions should a sensitivity analysis cover?
Usually the three to six the decision leans on hardest. Pick them with engineering, operations and finance together, and give each one an owner.
Does a deterministic model give probabilities or confidence ranges?
No. The same input always gives the same result, so there is no random spread to measure. The range in a sensitivity pack comes from the assumptions you chose to test, and each step should say why it was chosen.
What if the ranking flips within the tested range?
Then the decision depends on that assumption. Either reduce the uncertainty first, for example with a supplier test or a time study, or choose the option that stays acceptable on both sides of the break-even point.
Conclusion
Before the next approval meeting, ask the project team for one line: the break-even point of the recommended option and the name of the person who owns that assumption. If they cannot give you both, the analysis needs another round before anyone signs.
Not sure your result is strong enough?
Which assumption does your current business case lean on hardest? Send it to us with the decision it supports, and we will suggest the steps it should be tested over before approval.
Sources
- DBR77 Digital Twin, Averages vs simulation over time
Want to test a decision from your plant?
Describe the decision to a DBR77 expert and we will show how DBR77 Digital Twin compares the options on output, waiting time and transport load.
Questions
Have a question about this topic?
Tell us what it is about. We will reply within one business day.
Digital Twin newsletter
Short notes on testing layout and flow decisions on a model, and news from DBR77 Digital Twin. You can unsubscribe at any time.