From Co-Pilot to System of Action: How Trunk Tools Rebuilt Itself in a Year
A year after its $40M Series B, Trunk Tools has rebuilt itself from a chat-style co-pilot into what CEO Dr. Sarah Buchner calls a “system of action”: AI agents that finish work end to end and hand tasks to each other, sitting neutrally on top of Procore, Autodesk, SharePoint, and email. New agents now ship every four to six weeks, the platform overrides human spec-section decisions about 15% of the time, and enterprise customers have started writing API rate limits into their contracts. Buchner also addressed the acquisition rumors, sort of.
When Trunk Tools announced its $40 million Series B led by Insight Partners in July 2025, Sarah Buchner took criticism for raising too much. Twelve months on, she returned to Bricks, Bucks & Bytes with a straightforward defense: building something foundational in construction takes cash, and the cash went into engineering and AI hiring that pushed the company to a release cadence of a new tested, evaluated agent every four to six weeks.
More interesting than the pace is what got rebuilt. A year and a half ago, Trunk Tools was, in Buchner’s own words, “a co-pilot for the construction industry”: a retrieval product with a chat interface. That product is gone. What replaced it says a lot about where construction software is heading.
Inside the rebuilt stack
Cortex, agent orchestration, and agent-to-agent handoff
The platform now has three tiers. At the base sit the integrations and preprocessing, including the drawings pipeline, which map construction-specific concepts into a knowledge graph. Buchner calls this package Cortex: a mix of Trunk Tools’ own models and publicly available ones that gives agents a brain-like structure to work from. “An agent without the cortex is pretty useless,” she said, “especially if they don’t touch drawings.” It’s a similar architectural bet to the one Primepoint’s ex-Meta founders raised $10M to pursue: structure the documents first, put language models on top.
Above Cortex sits agent orchestration and a harness for building new agents faster. The layer Buchner finds genuinely exciting is agent-to-agent handoff. Her example: a drawing review agent finds discrepancies in a drawing set, then passes them to an RFI agent that writes the RFI. No half-finished suggestions in a chat window. The company killed what she called its “half baked” agents entirely; everything visible now is file-triggered and completes work end to end.
The quality control behind that autonomy is unglamorous. One of the largest teams inside Trunk Tools’ AI group builds ground-truth and evaluation datasets, which generate a confidence score for every model decision. That’s what lets the platform override the spec section a human assigned to a submittal, which it now does about 15% of the time, only above a strict confidence threshold.
Not a Procore killer, on purpose
Systems of record store; systems of action work
Buchner was direct about the persistent assumption that Trunk Tools wants to replace Procore or Autodesk. “That was never the plan.” Rebuilding a system of record would be expensive and, in her words, boring. Those platforms won the collaborative-database market, and every jobsite also runs SharePoint, an ERP, and overflowing email folders. A system of action earns its keep by sitting neutrally on top of all of them: a submittal lands in Procore or SharePoint, the agent triggers, does the work, and pushes results back wherever they need to go.
The purpose, as she frames it, is deleting bureaucracy rather than software. Paying highly qualified superintendents to type data into systems of record at 5 PM instead of going home, or having three separate parties review the same submittal, is the waste she’s targeting. “If I get to a point where no human has to do data entry anymore into one of these systems of record, I feel like I have achieved my purpose on this earth.”
Understand the platforms before they reshape your stack
Weekly analysis of the architecture battles deciding construction software’s next decade.
Join 3000+ ReadersThe shakeout test, endorsed by a competitor’s shortlist
Four traits, and a new line item in enterprise contracts
Owen read out a recent LinkedIn post from Mark Schwartz, Trimble’s Group President for AECO Software, predicting a construction tech shakeout. Schwartz argued the survivors will share four traits: they’ll be platforms rather than point solutions, carry a credible AI story instead of buzzwords, stay interoperable so customers can swap components, and meet enterprise-grade security and data residency standards. Miss one and you might win the odd deal; you need all four.
Buchner didn’t find it provocative at all. Her reading: Schwartz is effectively saying you need a system of record and a system of action, you need Trimble and you need Trunk Tools. On interoperability, she added a detail worth filing away. This is the first year her enterprise customers, especially those above roughly $200 million in revenue, are negotiating API rate limits and MCP limits directly into contracts. Buyers want platforms, but they want plug-in, plug-out platforms. That matches what we keep hearing about buyers, not vendors, setting the real pace of ConTech AI.
I believe the future of most software stacks is going to be a database, a system of record or data lake, and then a bunch of different agents moving between inputs and data, doing the work in between.Dr. Sarah Buchner, CEO, Trunk Tools
Roadmaps, data terms, and those acquisition rumors
A three-year vision, a six-month plan, and a non-answer worth reading
Ask Buchner for a two-year roadmap and she’ll tell you she doesn’t have one. She keeps a three-year vision and a three-to-six-month roadmap, nothing in between, because staying flexible enough to absorb what generalized AI makes possible each quarter beats a long plan. Older companies, she argued, often lack the genetics to change course that often.
On data, Trunk Tools holds zero-retention, zero-training contracts with every large model provider it uses, a prerequisite for enterprise GC deals where diligence teams now ask about retention directly. Dustin DeVan claimed most new startups skip that negotiation and run on standard consumer terms, which, if true, quietly disqualifies them from enterprise contracts before the first demo.
And the acquisition rumors? Buchner noted they’ve circulated for two years, said it would be strange if a post-Series B company at her scale weren’t getting calls, and declined to comment further. Then she added a line that will keep the speculation alive: it’s a skill to know when to sell, and a skill to build a standalone lasting company, “and I’m trying to sharpen my skill on the second.”
| Dimension | System of record | System of action |
|---|---|---|
| Core job | Store and share the project’s source of truth | Do the work humans perform on top of that data |
| Examples on a jobsite | Procore, Autodesk, SharePoint, ERP, email | Drawing review, RFI drafting, submittal processing agents |
| Position in the stack | The database everyone writes into | Neutral layer moving between all data sources |
| How value shows up | Collaboration and auditability | Less data entry, fewer duplicate reviews, faster cycles |
| Competitive dynamic | Consolidated, hard to displace | Emerging, won on accuracy, trust, and delivery |
It’s Buchner’s term for a layer of AI agents that completes work rather than storing information: reviewing drawings, writing RFIs, processing submittals, then pushing results back into whichever system of record the project uses. It complements rather than replaces platforms like Procore and Autodesk, which remain the collaborative source of truth. (Source)
Cortex is Trunk Tools’ name for its preprocessing and model layer, released a few weeks before the episode. It ingests project data including drawings, structures it into a construction-specific knowledge graph, and combines the company’s own trained models with public ones. The agents on top all draw from this shared brain. (Source)
Through evaluation infrastructure. Trunk Tools maintains large ground-truth datasets that let every model decision carry a confidence score, and overrides only happen above a strict threshold. Buchner said the platform now changes the human-assigned spec section on roughly 15% of submittals because the data shows the model is right more often on that task. (Source)
In a LinkedIn post discussed on the show, Schwartz, Trimble’s Group President for AECO Software, predicted a shakeout in which surviving vendors share four traits: platform scope, a credible AI story, interoperability, and enterprise-grade security with data residency. Startups that won deals through endless custom builds, he warned, are waking up buried in technical debt. (Source)
Because interoperability has become a purchasing requirement, not a nice-to-have. Buchner said 2026 is the first year she’s seen enterprise customers, typically contractors above $200 million in revenue, write API and MCP access limits into agreements so they can plug agents and tools in and out as needs change. Buyers are protecting themselves against lock-in. (Source)
She wouldn’t say, beyond noting that acquisition rumors have followed the company for two years and that a post-Series B market leader not receiving calls would be strange. Her closing framing leaned toward independence: knowing when to sell is a skill, building a standalone lasting company is another, and she’s working on the second. (Source)
Whether the vendor holds zero-retention, zero-training agreements with the foundation model providers behind its product. Customer data belongs to the owner or the GC, and enterprise diligence teams increasingly ask for proof. DeVan’s claim on the show was that many young startups run on standard model-provider terms, which blocks them from serious enterprise deals. (Source)
Related Articles
Primepoint Raises $10M Seed to Build AI That Actually Reads Construction Drawings
“You Can Only Go As Fast As Your Customers”: The Quiet Brake on ConTech AI
