A single skyscraper can generate 3.6 million pages of contracts, pictures, blueprints, schedules, and RFIs, making construction data difficult for people to use.
2
Trunk Tools uses AI to organize construction documents, answer field questions with source files, and find discrepancies that can cause expensive rework.
3
Sarah Buchner argues that retrieval-augmented generation is becoming commoditized, while vertical agents that solve real construction problems can keep people in control and reduce bureaucracy.
Summary
Dr. Sarah Buchner introduces construction as a data-heavy industry where mistakes have physical and financial consequences. One New York City building gave Trunk Tools access to 3.6 million pages of contracts, pictures, blueprints, schedules, and RFIs. Buchner says construction loses 10% of its $15 trillion annual value to rework caused by data discrepancies. Trunk Tools organizes these files into what she calls the brain behind construction. Its Q&A system lets field workers ask questions by text message or web app and receive an answer with links to the original documents. Buchner shows how the system can expose conflicting information about a door, then proposes an agent that turns the discrepancy into an RFI and works toward resolving it. Her argument is that retrieval-augmented generation alone is becoming commoditized. The useful products will keep humans in the center while autonomous agents handle specific, real-world construction tasks.
Construction projects produce millions of pages that people cannot easily use
Buchner describes a New York City building that gave Trunk Tools access to 3.6 million pages of construction documentation. The files include contracts, pictures, blueprints, schedules, and RFIs. She compares the printed stack to one three times as high as the building itself. Construction creates this volume because many people make decisions every day and document them heavily. Buchner connects that behavior to the industry's exposure to litigation. The amount and format of the information make it difficult for field professionals to find the detail they need before work proceeds.
Data discrepancies turn document problems into physical rework
Buchner compares construction errors with bugs in software that appear after code is merged. In construction, the failure may appear in the field when concrete is poured or when a pipe does not fit into its opening. She says construction is a $15 trillion-a-year industry and that 10% of the work is rework caused by discrepancies in the data. She puts the annual cost at $1.5 trillion. Her point is that document inconsistencies are not an administrative nuisance. They can become expensive failures in the built environment.
Trunk Tools organizes construction files into a shared knowledge base
Trunk Tools puts construction knowledge in one place, which Buchner calls the brain behind construction. The system is built to digest and remember construction-specific file types. She presents this as a vertical application of retrieval-augmented generation for an industry that accounts for 10% of GDP. On top of this knowledge base, Trunk Tools is deploying AI-based agents. Buchner says the agents can have different levels of intelligence and can be aimed at particular tasks rather than treated as a general assistant.
A field worker can ask a document question through text or a web app
Buchner shows a traditional copilot workflow against 3.6 million pages of construction data. A user asks whether door number 2103 needs power-actuated hardware. The question can arrive through a text message or a web app. Within seconds, the system returns enterprise search results, a semantic answer, and the original document as a source. This gives field professionals a direct way to ask about project information without manually searching the entire document set. The example is a specific site question rather than a generic chatbot demonstration.
The answer can expose a discrepancy before it causes a field error
The door question produces an answer that contains conflicting information. Buchner says the discrepancy reveals that someone made a mistake somewhere in the 3.6 million pages, and that the mistake could become very expensive. She then opens one of the source files to show the kind of material involved: three-foot-wide tables filled with numbers that are effectively unusable for a person. AI is being applied to these documents because their structure and size make manual review impractical.
Agents can take action on a discrepancy instead of stopping at retrieval
Buchner proposes an agent that acts like an RFI when the system finds conflicting data. Rather than simply reporting that the documents disagree, the agent would work on resolving the discrepancy for the human. That lets the person return to their construction job instead of becoming stuck in paperwork. The distinction matters in her presentation: search and answers identify a problem, while an agent can begin the workflow required to address it.
The useful future is human-centered vertical agents
Buchner says retrieval-augmented generation is completely commoditized. She is speaking directly to people building retrieval systems and argues that the next layer should augment people with an army of agents. These agents should solve real-life problems in specific industries. In construction, that means working from project documents, identifying inconsistencies, and handling parts of processes such as RFIs while people remain responsible for the work. Her product direction is therefore tied to field tasks, not only to better document search.
"Instead of stopping where we have the data discrepancies, why don't you use an agent that is just actually an RFI and solving the discrepancy for the human so that the human can go back doing their job and not being a victim of bureaucracy."04:01
Who should watch
You build retrieval or copilot systems and want to see what a construction-specific product does after document search finds a conflict.
Your team works with large collections of contracts, drawings, schedules, or RFIs and needs field workers to get answers with source documents.
You are designing agents that should take action inside an industry workflow rather than only answer questions.
