“Deliveries arrive late, so we need to hire more drivers” is a proposed solution, not a diagnosis. Orders might leave the kitchen late, packing might be overloaded, promised windows might be unrealistic, or timestamps might be wrong. Investigation means testing competing explanations, not collecting arguments to defend the first guess.

This article starts with a problem already prioritized. If you do not know which issue warrants attention, start with the business health diagnosis. If you need to record an individual complaint and offer the customer a remedy, follow the operational complaints and recovery process. Here we investigate system-wide patterns. ASQ distinguishes tools that generate hypotheses, including fishbone diagrams, from tools that organize frequencies and relationships [R01][R02][R03]. No diagram, by itself, turns correlation into causation.

1. State the problem using verifiable facts and a clear scope

Use a sentence that can be checked without debating motives: “Of 240 fulfilled orders with a promised delivery time, 60 arrived after that time during two weeks; lateness was concentrated around lunch.” This is a fully hypothetical example, not evidence from a real customer.

Specify what happened, when, where, number of events, denominator, segment, and severity. Define promised time, actual arrival, when preparation started, and when an order left the facility. Without event definitions, “late” can mean several incompatible things. If fifteen orders were canceled before any delivery promise was accepted, do not casually include them in the denominator of 240 fulfilled orders.

“Delivery is bad” is an unhelpful statement. “60/240 = 25% of fulfilled orders missed the agreed window, particularly during a specific period; the cause is unknown” can be investigated. The percentage is an outcome, not an assignment of personal blame.

2. Reconstruct the actual workflow and event timeline

Follow the sequence order received → capacity confirmed → kitchen → packing → assignment → dispatch → route → receipt and capture times at each relevant transition. Reuse the existing operating and logistics flow map if it still reflects reality. Compare what the procedure expects with what actually happened; a documented procedure is not necessarily how work is performed.

For each late order, retain order_id, promised window, verified capacity status, stage timestamps, last-minute changes, recorded incident, and review date. Service appointments and software products have different steps: for a website, those might be page load, login, form submission, and completion. Do not collect unnecessary personal names or phone numbers when an internal identifier can support the analysis.

Illustrative timeline: 12:00 order confirmed with a 12:45 promise; 12:12 preparation begins; 12:34 kitchen finishes; 12:46 packing finishes; 12:51 departure; 13:04 arrival. The delay relative to the promise is 19 minutes. Packing took 12 minutes, but that does not prove that faster packing alone would solve the issue. Waiting, shared resources, and routing must also be understood.

3. Build competing hypotheses with a fishbone diagram

A fishbone or cause-and-effect diagram groups several possible causes into categories. Its purpose is to widen and structure investigation, not declare a winner [R02]. For lunch delivery, adapt categories to the work:

Branch Candidate explanation Evidence to gather
Method Orders are batched without considering promised time Actual preparation and dispatch sequence
People and coordination Packing-to-driver handoff has no clear owner Handoff records and non-accusatory observations
Materials and equipment Packaging is missing or defective Inventory, interruptions, defects, and lots
Measurement The clock or meaning of “delivered” changed Time zone, event definitions, and logs
Demand and environment Simultaneous orders or unusually slow routes Requests by time slot and verifiable conditions
Commercial commitment Time slots were offered without checking capacity Confirmations and hour-by-hour availability

Manufacturing's classic “six Ms” need not be copied rigidly into every service business. Where a candidate explanation raises a serious food-safety or compliance concern, institute appropriate safeguards immediately while other hypotheses remain under investigation.

4. Use the five whys to dig deeper while keeping alternative branches

The five whys method repeatedly asks what explains an outcome. ASQ notes that fewer or more than five questions may be required and that the method can be paired with brainstorming or fishbone diagrams [R01]. Do not force a single chain or treat unverified answers as facts.

One hypothetical candidate chain could be:

  1. Why did an order depart late? Packing finished after the intended dispatch time.
  2. Why was packing late? Many orders entered the same time slot simultaneously.
  3. Why did those orders cluster? More were confirmed than the verified slot capacity allowed.
  4. Why were they confirmed? The sales screen displayed daily availability rather than hourly capacity.
  5. Why was that restriction missing? Booking rules had not incorporated the measured time-slot bottleneck.

Every answer needs evidence. If the second explanation cannot be verified, writing three more questions does not rescue the chain. Other branches may involve packaging shortages or routing. Causes may be multiple, interacting, or still unknown.

“An employee forgot to notify the driver” should not automatically be the final answer. Examine whether the signal existed, whether it was visible, whether there was enough time to act, and whether the process provided a safe alternative. Investigating system conditions is not the same as claiming nobody is accountable; it means avoiding a blame-based conclusion from a single incident.

5. Collect comparable observations and stratify the events

Before drawing charts, define a check sheet: observable event, classification, recorder, collection period, and coding rule. ASQ recommends operationally defining the problem and briefly testing the collection form so observations can be recorded consistently [R04]. Do not write “cause: kitchen” when all you observed was “the order left the kitchen later than expected.”

Stratification separates shifts, routes, locations, and order types that an overall average may hide [R06]. In the same hypothetical two-week case, suppose 100 committed deliveries occurred around lunch and 140 in other windows. Of the 60 late orders, 38 were at lunch and 22 outside it. The late-order rate is therefore 38/100 = 38% at lunch versus 22/140 = 15.71% elsewhere. This pattern tells us where to investigate; it does not establish that time of day directly caused the delays.

Distinguish events from opportunities for error. If a route has five deliveries, those orders may share one upstream delay; that matters when interpreting independence. Sixty observations are not sixty independent randomized experiments.

6. Use Pareto analysis to direct attention, not invent causation

A Pareto chart orders categories by frequency, time, cost, or another consistently measured quantity to help focus investigation [R03]. Suppose that, for the sixty late deliveries, we classify the first step where the process fell behind plan, using one mutually exclusive rule:

First stage with observed deviation Orders Share of 60 late deliveries
Kitchen 28 46.67%
Packing 18 30.00%
Assignment/dispatch 9 15.00%
Route 5 8.33%
Total 60 100.00%

The first two categories represent 46/60 = 76.67% of recorded first deviations. That does not establish that the kitchen and packing caused 76.67% of the problem. It records where the delay was first noticed under the chosen coding rule. An unrealistic sales promise might create delays visible at several stages. The 80/20 relationship is not a law that every Pareto chart must satisfy.

If one event can belong to several categories, choose a single-label system for mutually exclusive totals or a multi-label system whose percentages must not be added as exclusive shares. Document the coding rule so another analyst can reproduce the chart.

7. Separate routine variation, unusual signals, and misleading relationships

Daily performance naturally fluctuates; a single change does not necessarily indicate a new mechanism. With suitable processes and enough historical data, control charts can distinguish common variation from signals suggesting special causes through calculated limits [R05]. Do not draw statistical control limits by adding and subtracting 10% from a target. With only a few days, start with an annotated time-series chart and be explicit about uncertainty.

Correlation means two things move together; it does not prove one produced the other. Rising orders and rising delays could reflect overloaded production, longer routes, more complex orders, missing supplies, or a changed promise. Ask which observations each explanation predicts, what might contradict it, and what other conditions stayed constant.

A contributing factor may increase the likelihood or severity of an event without explaining every occurrence. Naming one “root cause” can be misleading in a system with multiple dependent stages. Consider recording “supported factor,” “ruled-out factor,” “unverified factor,” and “evidence scope.”

8. Verify hypotheses with new evidence and safe checks

Prioritize explanations by plausibility, potential impact, cost of checking, and consequences of being wrong. Seek checks that distinguish rival theories. If packing overload is suspected, observe packing queues and time by slot; if routing is suspected, compare departure, route, and arrival times without needlessly tracking individual workers.

As a hypothetical, non-conclusive check, observe twelve consecutive lunch orders, capturing confirmation, kitchen completion, packing completion, and dispatch. Eight depart later than their internally planned departure time, and in six of those eight the delay is already visible before packing finishes. That weakens the theory that traffic alone explains the problem, but it does not yet explain why the kitchen or packing stages ran late. Further observations on days with different order mixes may be needed.

Do not intentionally degrade safety or service quality to “prove” a problem. A controlled intervention with test groups, safeguards, and side-effect measurement belongs to operational improvement experiments. This stage is about examining mechanisms through logs and safe observations.

9. Separate immediate containment, individual correction, and prevention

These three actions answer different questions:

  • Immediate containment: stop promising nonexistent slots, inform affected customers, and offer acceptable alternatives while the investigation continues.
  • Individual correction: complete, reschedule, refund, or otherwise remedy an affected order as appropriate under agreements and applicable obligations; record it through customer support.
  • System prevention: if evidence supports an hourly-capacity issue, introduce slot controls and verify their continuing operation. Assign an owner, procedure, date, and follow-up rather than adopting a change because it fits a plausible story.

Hiring another driver might do nothing if orders leave the kitchen too late. But do not rule out transport capacity on principle: reconsider it if the data reveal a verified delivery bottleneck. The choice depends on which hypotheses survive scrutiny and the expected consequences of acting.

10. Record a traceable investigation and decide the next step

A practical case file includes ID and date; measurable problem; timeline; process map; competing hypotheses; favorable and contrary evidence; missing data; checks; risk; containment; owner; provisional decision; and conditions for reopening. A five-whys chain or fishbone sketch can be attached, but neither replaces the case file.

A calibrated closing statement could read: “We observed 60/240 late deliveries, with a higher late rate at lunch; 46 of the 60 first showed a deviation in the kitchen or packing. Time-slot load and capacity should be checked next. No single root cause has been demonstrated.” Such wording can accommodate better evidence without rewriting history.

When evidence supports a mechanism strongly enough and a change can be made safely, move to the appropriate improvement cycle or test. If a critical hazard appears, apply proportionate protective measures before the investigation paperwork is complete.

Templates and outcome

Keep an event check sheet, Pareto table with its coding rules, causal hypothesis matrix, and investigation record. The accompanying CSV files are reusable educational examples, not actual staff or customer records. A small business can complete them on paper or in one spreadsheet; a regulated operation needs controls proportionate to its risks.

By the end, you do not need to claim to have found a final “root cause.” You need a clear problem, plausible competing explanations, evidence that can distinguish them, and a proportional next decision. Good investigation is willing to change its conclusion when better observations appear.

References