This audio is automatically generated. Please let us know if you have any comments.
Elliot Powell is the director of data and AI at Seattle-based heavy marine and civil contractor Pacific Pile & Marine. The opinions are the author’s own.
When we moved Pacific Pile & Marine to an enterprise-wide AI platform, I expected the hardest part to be the technology: choosing the model, connecting it to our systems, ensuring system security. These turned out to be the parts that were planned.
The hard part came once people started using it for real work and it didn’t seem like an AI problem at all. Instead, it was a simple question: What was the current revision of our presentation? It turns out that the AI couldn’t confidently answer that question, since we didn’t know the answer ourselves.
The model was not wrong. He was reading exactly what we had given him: the same document saved in various places, under various names, with no say in which copy really mattered.
For years, people had filled this memory gap. The superintendent knew which folder was the real one. The estimator knew that “final_v3_REV” exceeded “final_FINAL”. The AI doesn’t have that tribal knowledge, so all the inconsistencies we’d been quietly working on came out.
This is the lesson I would give any contractor starting down this path. AI will not solve your data problems. Instead, it will expose them quickly and for all to see.
The problem was never the model
In a recent Construction Dive piece, two Palantir authors argued as much Contractors’ problem with AI isn’t the software: “The problem is that no tool understands the business like the people who run it.”
Their answer is an ontology, a digital model of the business that ensures that the field and the back office speak the same shared language. I agree with this takeaway.
From the field, however, this shared language starts much smaller than a business ontology. It starts with what a folder is called, what’s in it, and what a file is called when someone pulls it from a workplace trailer at the end of a long shift.
If two project teams submit the same type of document in different places with different names, no platform will make those two jobs comparable, ontologically or otherwise. For us, unifying the language meant agreeing on nouns before applying it to anything.
We made this the first principle of our adoption framework. We asked all teams to get a few things right before building an AI workflow. Data comes first, ahead of how you delegate it to the tool, how you describe the task, and how you check what it returns.
Standardize a job, not the whole company
Our instinct was to write a company-wide folder standard and deploy it. We didn’t do it. Instead, we assigned a project manager to pilot a standard structure and naming convention in a live project. We’re treating it as a field test rather than a policy.
This choice matters. A standard naming convention written in the office is judged by whether it looks tidy. A standard convention used in a workplace is judging whether someone in the field can find what they need without calling the office.
A field test tests what a conference room can’t: whether a name is short enough to write on a phone, whether a folder makes sense for the crew and for accounting, and whether two categories overlap in a way that no one noticed on paper.
It also changes who owns the standard. When a PM is the one that proves it works, the standard stops being an IT mandate and becomes the way the project is run. This is the version most likely adopted by other PMs.
The second half of the job is deciding what counts as authorized. Working project files are messy by nature and should be. Material that an AI tool treats as truth needs a person to review it before making a final decision.
The story of our past project is a good example. We need comparable jobs quickly for prequalifications, so instead of letting a tool dig through years of closing folders, we’re creating a consistent summary for each completed project. The AI part is easy. The real project is to agree on what each summary should contain.
Adaptability trumps technical skill
Anthony Chiaradonna, Chief Information Officer at Consigli Constructionsaid it well at Construction Dive earlier this year: “the biggest skills shift we’ve seen within the industry isn’t on the technical side. It’s how teams are able to learn, adapt and manage change in real time without compromising quality, safety or schedule.”
This matches what I’ve seen, with a specific twist on the data. The people who get the most out of a launch aren’t necessarily the most technical. They are the ones willing to change a habit they have had for most of their career: how and where they save a file. This is a more difficult task than learning how to write a good prompt.
That’s also why our training, which includes a mandatory core course, a rapid shared library and open sessions, must cover where things are going and not just what to ask.
It also changes how we respond to resistance. When someone says AI “doesn’t work,” the useful follow-up is to ask why. The answer is usually worth hearing, because AI tools can only find what’s already there, and what’s already there is usually the problem.
I would tell other contractors that:
- Treat your first few months of using AI as a data audit. Each bad response points to a file, folder, or name problem.
- Agree on the nouns you want to use before choosing the platform. Folder names, document types, and file naming rules are the language your AI will speak.
- Pilot your naming convention in a live job with a PM who owns it.
- It decides what is authorized and puts a person between the working files and the material that AI treats as truth.
- Train and hire for adaptability. The skill you’ll need most is a willingness to change the way work is presented.
The model we ultimately choose will be replaced, probably sooner than any of us expect. The folder standard, the naming convention, and the habit of deciding what is authorized will last longer. This is the part of an AI release that no one demos, and it’s the part that decides if the release works.
