The “travel” line of an industrial equipment distributor swelled every quarter, and nobody could explain why one ticket costs $180 and the next $520. The answer was already sitting in the booking data. It just had to be processed seriously.
This engagement is presented without naming the company, at its request. It is an industrial equipment distributor of about forty employees, with six reps constantly on the road: client visits, trade shows, installation follow-ups, on both sides of the border. Every trip was booked at the discretion of the person leaving, with no rule and no benchmark.
Three questions before touching the data
Before opening a single file, we wrote the business questions down in black and white. What really sets the price of a ticket? Which factors does the company actually control? And can we draw from it a policy that we can defend in front of management?
It seems obvious, and yet it is the step most often skipped in data projects. People dive into spreadsheets, find amusing correlations, and six weeks later nobody knows which decision they serve. The three questions acted as a filter: anything that didn’t help answer them was set aside.
From eight million rows to 163,000 trips
The starting point was raw: more than eight million itineraries. We sampled, cleaned and documented every decision. Outlier fares were dropped, unusable variables were removed, and each removal was justified in writing. The result is a reliable, reproducible set of 163,000 trips, which matters more than its size: anyone can redo the cleaning and land on the same numbers.
No personal data was used. The analysis looks at the characteristics of itineraries, never at the people who travel.
A gap you can see by eye is not a real gap
When you plot price curves, you always see differences. Many don’t survive a test. So every observed gap was confirmed by a statistical test before it was kept in the analysis. What didn’t pass the test left the conversation, even when the story was a good one.
Two families of models were then pitted against each other on the same split of the data, and judged on data they had never seen. The model we kept is not the most accurate. It is the most explainable. The accuracy gap was negligible, and a rule you can’t justify in one sentence won’t be applied by the reps.
“The most accurate model was not the right choice. A rule that changes a team’s habits has to be explainable in one sentence and defensible in a meeting. Otherwise nobody applies it.”
What the data showed
Cost levers were ranked on two criteria: their impact and how much control the company has over them. Only one came out as significant, free and entirely in its hands: booking ahead, within a window of 22 to 45 days. Depending on fare class, the gap reaches $151 per ticket. It was the first lever turned into policy.
The other result is less comfortable. A saving taken for granted for years does not hold up against the data: layovers don’t lower the price. They did, however, cost the reps time, since they were stretching their days for nothing. The company stopped requiring them.
What people tend to remember from data projects is what they confirm. This one is mostly worth it for what it took away: a rule that cost time and fatigue, applied in good faith for years on the strength of an impression.
A tool rather than a report
The engagement ended with four purchasing rules and an internal application: the rep enters an itinerary, and the application shows the right price before booking. The recommendation lives in a tool they open at the moment of decision, not in a PDF read once in a meeting and forgotten. An expense becomes a decision.
What we take from it for other expense lines
This is not a project reserved for plane tickets. Whenever an expense line grows without explanation, the same sequence applies: write the questions, clean while documenting, keep only what passes a test, prefer the model you can explain, and deliver the conclusion in the tool people already use. Recurring purchases, subcontracting, freight costs, replacement parts: each of these lines often hides a free, controllable lever and one or two received ideas.
The full engagement, with its method and results, is detailed in the case study →
An expense line that keeps growing?
If you have booking, order or billing data and few answers, a well-framed analysis is often worth more than a new tool. We start with your questions.
What this article argues, we put into production for a client. The case walks through what was built, how, and what changed.
