Bottleneck analysis

What is a bottleneck?

A bottleneck is a step in your process where work piles up or sits waiting. The rest of the process can run smoothly, but if there is one place where cases bunch up, that single step sets the total throughput time. The idea comes from the Theory of Constraints by Eliyahu Goldratt (The Goal, 1984): every system has one weakest link that determines the end result.

Think of it as a bottle's neck. No matter how wide the rest of the bottle is, everything has to go through the same narrow opening. In a business, that neck sits at a person who has to approve, a system that answers slowly, or a decision point where someone hesitates.

If you remove the current bottleneck, the problem almost always shifts to the next weakest link. That is why bottleneck analysis is not a one-off exercise but a cycle: identify the constraint, get the most out of it, align the rest of the process to it, add capacity where needed, and start again.

How do you find bottlenecks in a process?

In the past you did this with interviews and a stopwatch. Today the answer sits in your systems. Every ERP, CRM or ticketing system records events: who did what, at what moment. Pull them together in an event log and a process mining tool can work out where the time goes.

The strongest signals are:

• Long wait time between two activities. The processing does not take long; the pause before it does. An invoice sits three days with the approver, the actual click takes thirty seconds.

• Big queue at one resource. One person, team or system always has more work queued up than the others.

• Repeated hand-offs. The case bounces back and forth between two teams before it moves on. Every hand-off adds wait time.

• High variance on one step. Some cases fly through, others stall for weeks. That points to a decision point that is not standardised.

In Power Automate Process Mining there is a Time Analysis view that shows this: average and total duration per activity and per transition. In Celonis it is called a performance or throughput view. The split between processing time (someone is actively working) and wait time (the case sits idle) matters a lot here. In most office processes, 80 to 90 percent of throughput time is pure wait time. The fastest wins therefore usually come not from working faster, but from waiting less.

Types of bottleneck

Resource bottleneck
One person, team or machine does not have enough capacity for the volume. The classic example is a manager who has to approve every purchase. Fix: split authority by amount, add approvers, or automate below a threshold.

Decision-point bottleneck
A step where someone has to choose between several paths and the rules are unclear. Employees hesitate, ask for advice, pass the case back and forth. Fix: document clear rules, build in decision trees, or move the decision.

System bottleneck
An IT system that responds slowly, or an external API that only calls back after minutes. Typical in integrations between SAP and a credit check or VAT validation. Fix: go asynchronous, add a cache, or scale the system up.

Hand-off bottleneck
The case has to move from team A to team B and stalls in the hand-over. Neither side feels like the owner. Fix: one owner per case, clear escalation paths, or remove hand-offs.

What do you do about it?

1. Prioritise by throughput-time impact. Multiply wait time by the number of cases per year. A step that costs ten minutes on ten thousand cases outweighs a step of one hour on two hundred cases.

2. Automate where you can write out the rules. Routine approvals and validations belong in a workflow engine or a Power Automate flow. The processing time stays the same, but the wait time disappears. This only works if the rules are clear.

3. Redesign where the fundamentals are off. Sometimes the problem is in the process itself: too many approval layers, unnecessary hand-offs, duplicate work between teams. Clicking faster will not help. You cut steps or redraw the process in BPMN before you reach for tooling.

Measure again after every change. A bottleneck you remove almost certainly shifts to the next weakest link. That is not failure; that is how processes work.

Worked example: purchase invoice approval

A small business notices that suppliers complain about late payments. The event log shows: invoices arrive on day 0, are booked on day 1, and only approved on day 12. Payment follows two days later. Total throughput time: fourteen days. Effective processing time across the whole journey: roughly twenty minutes.

The process mining tool shows that 90 percent of cases sit idle for eleven days at one step: awaiting owner approval. That is a classic resource bottleneck. The owner gets twenty invoices on their desk every day and approves them in batch, once a week.

The fix has three parts. Invoices under 500 euro go straight to payment if the supplier is known and the purchase order matches. Invoices between 500 and 5,000 euro go to the department manager. Only above 5,000 euro does the owner still step in. After two months the average throughput time drops to three days. The bottleneck is not gone; it has shifted to the department managers, where capacity is larger and decisions happen faster. That is the whole point of bottleneck analysis: not making the process perfect, but moving the constraint to a place where it no longer hurts.