How Gilbane used an AI tool to track 21,000 documents


This audio is auto-generated. Please let us know if you have feedback.

Keeping track of every detail in every contract a builder signs before stepping onto a jobsite can be a Sisyphean task. You can roll that metaphorical boulder up that mountain, but something, some detail, has the chance to bring progress tumbling down. Gilbane Building Co.’s Andrew Roy knows this well.

Roy, a lead superintendent, oversaw the Baird Center expansion project, a $456 million renovation of the Milwaukee convention venue, as part of a joint venture with Fond du Lac, Wisconsin-based C.D. Smith. 

Together, Roy and his team added approximately 300,000 square feet to the existing structure — which originally opened in 1998 — including ​​24 new meeting rooms, more than 400 indoor parking spaces, six loading docks and an executive kitchen, according to the project’s information page. The group broke ground in 2021 and fast-tracked completion, leading to a May 2024 finish. 

Thousands of documents, one chatbot

But getting there wasn’t easy. The job’s specifications and contracts added up to around 21,000 discrete documents, an issue that has become increasingly common on today’s more complex builds. 

“It is humanly impossible to be completely familiar with every single document on the project and every single change or conversation that’s going on,” Roy said.

This is where New York City-based Trunk Tools came into play. The company makes an artificial intelligence-based tool that contractors can use to track a project’s documents and contracts to get immediate answers without leaving the jobsite.

In order to navigate the sea of data contained in the Baird Center’s files, Roy’s team uploaded all their documents into the platform, including drawings, RFIs, contracts and change orders. 

Once they did, the large language model behind the tool — a chat-like feature known as TrunkText, which builders can access on mobile devices or a computer — was able to answer questions, respond to queries and reduce the time it took to find specific details when working through an issue.

“We kind of stumbled upon it, and then we started to use it for coordination between documents, so coordination between the door hardware schedule and the electrical drawings, or the low-voltage drawings, or the life safety drawings,” Roy said. “That’s where it became extremely powerful for us.”

TrunkText is but one construction-oriented AI offering in a quickly expanding field — other examples include DocumentCrunch, which can search, evaluate and mark up contracts for users based on questions and jobsite risks, and Togal.AI, which uses deep machine learning to help estimators with accuracy.

Gilbane started using TrunkText on the back third of the Baird Center project’s lifespan, piloting it in January 2024. Roy’s team leveraged it on interior finishes, the interior buildout, and a lot of the exterior sitework and exterior enclosure, Roy said. Gilbane declined to share the cost of the service.

Avoiding rework

Roy pointed to a question the team had about a fireplace in a feature wall with a large exhaust system covered with acoustic plaster. The material for the finish took six months to acquire from Europe. Roy and the mechanical contractor performed an inspection and noticed an anomaly in the ductwork.

They had a question — did the seams need to be sealed? A mistake would mean costly rework, expensive equipment rental fees and lost time. Typically, answering this kind of question would involve a time-consuming email chain between the mechanical contractor, the design team and Roy, which would take hours, if not days. 

Instead, they asked TrunkText.

“Sure enough, within 20 seconds, I had five, six different documents pulled up in front of my face where I was and a text response saying, yes, the seams need to be sealed on this ductwork in order to create a proper vacuum,” Roy said. “If seams are not sealed, warranty will not be maintained, and the system may not function properly.”

That’s just one example of hundreds of queries workers submitted over the course of the project. 



Source link

About The Author

Scroll to Top