Trunk Tools Launch: Disrupting the $15 Trillion Construction Industry with Autonomous Agents

How much documentation does it take to build a single skyscraper? For a New York City building using Trunk Tools, Sarah Buchner asks the audience to guess: more than a thousand pages, ten thousand, a hundred thousand, a million? Buchner reports that the project supplied 3.6 million pages of content, including pictures, blueprints, schedules and requests for information, or RFIs.

Buchner compares the printed documentation to a stack three times the building’s height. The scale makes construction an obvious candidate for retrieval-augmented generation, or RAG: retrieve relevant material from the project’s records, then use it to answer a specific question. The worker needs a usable answer from that collection while standing in the building those documents describe.

Construction produces this much documentation because many people make decisions every day, in an industry where litigation makes recording those decisions consequential. Buchner, who describes her PhD as combining AI and construction while crediting her team’s deeper expertise, compares the coordination problem to software development: individual contributions can look reasonable until they are merged. In construction, the integration failure appears in the field, when someone tries to pour concrete or fit a pipe into a hole and discovers that the pieces do not fit.

Conflicting records can become physical rework. Buchner estimates annual construction spending at $15 trillion and attributes 10%, or $1.5 trillion a year, to rework caused by data discrepancies. She gives no year, price basis or measurement methodology for those estimates. The operational problem is concrete: a discrepancy buried in project records can remain invisible until a crew acts on it.

The proposed foundation is a centralized construction knowledge layer. Rather than treat the corpus as interchangeable text, Trunk Tools specializes its RAG product for construction file types. The TrunkText diagram connects drawings, RFIs, schedules, submittals, contracts, change orders and bids to a shared system that reads, parses and structures project data. Buchner describes this as the brain behind construction; her claim of perfect digestion and memorization expresses the product ambition, rather than a demonstrated accuracy measurement.

Buchner describes construction as a vertical representing 10% of GDP, without specifying the geographic or temporal basis. Above the shared knowledge layer, she places an array of AI agents. Their capabilities vary, but they share a dependency: useful action requires access to the project’s construction-specific information. She then moves to a field example of the most familiar interaction, a copilot answering a worker’s question.

The copilot’s basic contract is an answer plus access to the original documents. Buchner says it answers questions over the 3.6-million-page collection within a few seconds; she supplies no latency benchmark. She also reports use on billions in construction volume, without specifying currency, period or deployment scope. The actual-site question concerns whether a particular door—identified verbally as door 2103, number two—needs power-actuated hardware.

A worker submits the question by text message or through the web app. The system returns enterprise search results alongside a semantic answer: relevant documents to inspect and a synthesized response to the question.

The completed answer exposes a coordination gap. The door schedule specifies power-actuated hardware, but the electrical drawings lack a power callout and panel designation. A simple affirmative answer would miss the installation problem: the hardware requirement appears in one record without the corresponding electrical details in another. The screen presents three source documents and a button labeled “Create RFI & send to Procore,” offering a route from discovery toward follow-up. Buchner identifies the discrepancy as a potentially expensive mistake.

Opening a source file makes the retrieval problem tangible. The underlying documentation includes sprawling tables packed with numbers, which Buchner describes as three feet wide. The useful output is therefore more than a located file: it brings a specific requirement and its missing counterpart into view, while retaining the route back to the records a person needs to inspect.

Once the system has identified conflicting documentation, stopping at the answer leaves the worker with the administrative task of pursuing clarification. Buchner proposes an agent that creates an RFI to move the discrepancy toward resolution. For the door example, the unresolved issue is how the specified powered hardware is supported by the electrical design. Creating the request would begin that coordination process; it would not itself establish that the design had been corrected or the door’s power provision resolved. The talk proposes this next step without showing a completed resolution. Its intended benefit is to return the worker to construction work instead of making them manage the paperwork around the conflict.

Buchner’s closing position is that RAG itself has become commoditized. In her view, the next opportunity is to keep the human at the center and surround them with agents that solve concrete problems. The door example gives that position substance: retrieval supplies the evidence, the answer exposes the inconsistency, and a specialized agent could carry the issue into the process used to resolve it. The product’s value grows when finding information leads to useful work on the person’s behalf.

Her proposed direction is vertical agents solving real-world problems, grounded in the records and workflows of a particular industry. She closes by inviting attendees to the Trunk Tools expo-hall booth, displaying a QR code and emphasizing that the company is hiring.