# Why I stopped trying to calculate what change orders cost

*A derivative of the research brief “What Counts as a Change Order?” Built from public federal contract data.*

I started this project with a simple question. What do change orders actually cost a construction project? It is one of those numbers everyone in the industry seems to half-know. You hear that change orders add eight to fourteen percent to a job, and the figure gets repeated in sales decks and trade articles until it feels like settled fact.

I wanted to check it against real data. The federal government publishes every contract action it takes, with a code, a date, a description, and a dollar amount. Eleven years of General Services Administration construction contracts sat there in public, waiting to be counted. It looked like a clean path to a benchmark.

Then I read the records and the benchmark quickly fell apart.

## The number was easy

Pulling the figure was straightforward. Across 1,481 completed, firm-fixed-price GSA construction contracts that each began with at least a million dollars committed, the actions carrying an official change-order code added up to 3.24 percent of the money originally committed. Reproducible, defensible, done.

Then I looked at what those coded actions described. I pulled a random sample of eighty change-order records and read the descriptions one by one. Thirty-nine of them clearly changed the work the contractor had to do. Twenty-one changed only the schedule. The rest were a grab bag: administrative corrections, accounting adjustments, contract options being exercised, a few descriptions too thin to classify at all, and one record that had nothing obvious to do with construction.

One action worth 10.6 million dollars carried a change-order code but mostly exercised construction options that had been priced into the original contract. Options are not changes. They were part of the deal from the start.

So the 3.24 percent was real, and it was also close to meaningless as an answer to my original question. It totaled everything filed under one label, and the label encompassed at least seven different kinds of business activity.

## The definition decides the answer

Here is where it got interesting. I widened the definition to include a neighboring code, supplemental agreements for work within the original scope. The figure jumped from 3.24 percent to 17.07 percent. A five-fold swing, from the same contracts, over the same eleven years.

At first that looked like a plausible range. A cautious low, a generous high, the truth somewhere in between. Reading the added records killed that idea too. Nine of the twenty largest were options worth 658.1 million dollars between them. Five more were paperwork and contract clauses. The higher number was not capturing more change orders. It was capturing more things that were never change orders in the first place.

At that moment, the question stopped being *how much do change orders cost?* and became *does the category even mean one consistent thing?* Based on the public record, it does not. What a change order is depends entirely on which codes you decide to count, and the codes were never designed to answer the question I was asking. I never did settle whether eight to fourteen percent is right. What I found is that the question cannot be answered from the codes alone.

## The gap is the story

The interesting finding was hiding inside the failure of the original one. A single event on a jobsite, a wall opened to reveal conduit where the drawings showed nothing, travels through a photo log, an RFI, a contract modification, an invoice, and finally a line in an executive report. At each step the label can change. The management report at the end inherits whatever label survived the trip, and it looks far more certain than the records underneath it actually are.

That gap between what happens in the field and what reaches the dashboard is an operating problem, and it is more useful than the benchmark I set out to find. A precise-looking percentage that combines scope changes, schedule extensions, and options tells an executive very little. A record that keeps those things separate from the moment a change is discovered tells them a great deal.

## What I would tell anyone starting from a benchmark

If you are trying to measure something from coded data, read a sample of the raw records before you trust the total. The code tells you how a transaction entered a system. It does not tell you what happened on the ground. Those are different facts, and the distance between them is where most confident-sounding statistics go quietly wrong.

---

*The full ten-page brief, the analysis code, and the dataset are available with the case study.*