Artificial Intelligence (AI) is a concept that is widely
discussed, but not many people have seen it working behind-the-curtains. As a
seasoned AI developer, the author routinely designs construction dispute AI agents
that specialize in using large language models to ingest, analyze, and extract
data from construction documents. In this article, the author will provide a
behind-the-curtains look into the current breakthroughs and struggles of generative
AI, which is an advancement that will continually reshape the construction industry
for years to come.
Generative AI – The Current Breakthrough
The current breakthrough, generative AI, was recently unleashed
by significant advances in the use of mathematical transformers. These
breakthrough transformers are essentially giant calculators for words, which
are designed to “predict” the next letter, word, sentence, or paragraph based on
some observed pattern.
In the generative AI developer space, we have spent the last
two years learning and optimizing how to request and receive information from
these large language model calculators. There are arguably two major large language models
in existence today with several others contending for the third place. These
two primary large language models can be considered the “Microsoft vs. Apple”
of the Generative AI landscape. While there are others, these big two, OpenAI and Anthropic, dominate. Each
model has taken years to build and train, and each model strives to ingest and
compute all written human knowledge in the English language. These primary large
language models have a semi open-source philosophy that allows AI developers to
“call” the models with computer code and then receive their calculated generative
AI reply.
As developers, we pay a small fee for every single “call” to
the model. This fee is charged to our developer account based on the amount of
words we send, and the amount of words received back from each call. As an
example, in a recent project extracting data from daily reports, over 15,000
“calls” to the big two large language models were made over a few hours. Here
is where we begin to peel back the curtain on the struggles of generative AI development
for construction. There is a maximum length of a question we can ask during
each “call” to the models. For example, in this example I could not ask a question longer than
about +/- 145 pages of text. This means we cannot simply ask AI to reply with
all issues from a large construction document set. We are also limited to about
105,000 generative AI response words per minute, and thus, we cannot just ask
thousands of documents a single question all at once.
Generative AI developers must code programs that work within
the constraints of the current large language model call limitations along with
devising condensed vector databases to store document data. The industry is
working on increasing the limits allowed while also maintaining reasonable
compute times for everyone to be able to use it at scale.
AI Agents and Their Applications in the Construction
Industry and Claims Resolution
To space out calls to the generative AI models and be within
the allowable limits, developers have created the concept of AI agents.
Each AI agent is programmed to be a specialist at a specific task, and it uses
a series of cascading instructions to perform a small portion of a much larger directive
over time. For example, a Data Extraction Agent will read a given document page
by page and extract specific facts from each page. The extracted data will then
be handed to the next AI agent that specializes in analyzing construction dispute
facts related to drilled shafts. Our programmed use of multiple specialized AI agents
allows for the calls to the large language model to be spread over thousands of
small pieces and then combined into what looks like a single response. This new
AI agent process allows for a full AI task to be performed over the entire
document set using multiple agents.
Below is a current list of generative AI agents that are being developed
for construction dispute resolution:
1)
Deposition Review and Prep Agents
a. Reads
each e-mail and identifies main issues and conflicts
b. Recommends
people to depose and specific questions to ask about specific emails
c. Prepares
legal firm for potential deposition scenarios and topics
2)
Mock Trial Agents
a. Identifies strengths and
weaknesses in legal cases using a three-panel judge scenario
b. Provides legal brief and
legal management plan to increase the odds of a favorable mock ruling from each
judge next time
3)
Timeline Agents
a. Builds event timelines
b. Converts years of daily
reports into work performed each day, by location
4)
Data Extraction Agents
a. Extracts-the-facts from
repetitive document sets (Example: review each of the 16,000 daily reports and
extract any facts that had to do with dewatering that day)
5)
Contract Risk Matrix Agents
a. Processes the Prime contract,
specifications requirements, special provisions requirements, and plan notes
b. Evaluates contractual risk
allocated to each party by the contract documents above
c. Prepares contract risk
register
6)
Project Notice Agents
a. Processes the prime contract,
specifications requirements, and special provision requirements
b. Identifies required notice
letters and triggering compensable events listed in the contract
c. Drafts example project notice
letters using anticipated real-world scenarios likely to happen on the project
7)
Project Issue Agents
a. Reads each letter and
identifies main project issue groups
b. Creates issue file (all
letters organized by main and sub issue groups)
8)
Change Order Cost Estimate Agents
a. Creates full cost estimates for specific scopes
of work (Noted to be getting much better over the last 60-days with new LLM
model releases)
9)
Project Scheduling Agents
a. Creates full project schedules for specific
scopes of work (Noted to be getting much better over the last 60-days with new LLM
model releases)
Concluding Remarks: Future of Generative AI
Today's professionals need to understand both the
capabilities and the limitations of generative AI. The construction industry is
at a pivotal moment where generative AI development is accelerating rapidly, and
AI agents are becoming more specialized and capable of handling increasingly
complex tasks. This article is intended to provide the reader a deeper understanding
of how to leverage generative AI within the current technological limitations.
As developers continue to push the boundaries of what's
possible with large language models, we can expect the current limitations on
call sizes and processing speeds to expand. This will enable more sophisticated
AI agents that can handle larger document sets and more complex analyses.
However, the fundamental principle remains; success with generative AI requires
deep understanding of both construction processes and AI agent processes.
The future is not about replacing human expertise with AI,
but rather augmenting human capabilities with powerful analytical tools.
Author, Travis Olson, is a Director with Berkeley Research Group. He is effective at
developing and applying in-house artificial intelligence agents that extract
facts efficiently from construction dispute documents. He has more than
seventeen years of experience in heavy civil and commercial construction encompassing
major infrastructure projects including bridges, dams, water treatment
facilities, light rail systems, and commercial high-rise construction.
Editor, Thanh Do, is a structural forensic engineer and expert witness with Thornton Tomasetti, Inc. He specializes in investigations of construction/design defects and collapses, Design-Build delivery, and standard of care assessment. He also oversees the Forensic Visualization group at Thornton Tomasetti, which produces graphics and animations for trial exhibits/demonstratives.
