Zero-cost analysis at Throughline

In late 2024, while evaluating a private equity acquisition, I received a 96,000-row ERP export from the target company. Our analyst spent days building pivot tables. I dropped the same file into Claude, described the schema in a few sentences, and had structured analyses in minutes. Then I repeated the experiment on construction schedules at Mosaic, where I was VP of Product. Same result. Days of manual work collapsed into minutes of conversation.

The epiphany was that insight was no longer gated by technical skill. Experience with the business now mattered more than fluency with the tools.

Sitting here in 2026 that is not a novel observation. At the time it was revelatory.

That gap was too large to ignore. It became a company.

The thesis

LLMs dropped the marginal cost of serious analysis to near zero. Every question you can ask about a company's data can now be asked and answered in minutes, not days. That changes who holds the advantage. It used to belong to people who could write SQL or drive Excel. Now it belongs to people who know the right questions to ask.

It also changes what analysis looks like. Dashboards exist because analysis used to be expensive. You built the report once and stared at it for a year. When analysis costs nothing, you ask, get an answer, and throw it away. Disposable analysis replaces static dashboards. When the cost of insight is effectively zero, the limiting factor is knowing which questions to ask.

The company

Two partners and I founded Throughline Labs to build AI business intelligence for homebuilders. Mid-market builders run on fragmented ERPs, CRMs, accounting tools, and spreadsheets that barely talk to each other. From Mosaic, we knew those ERPs cannot be replaced. Switching costs kill every attempt. They can only be layered on top of. That was the product: a unified layer over the systems a builder already runs, queried in plain English.

I built the prototype. Natural-language queries over builder data across scheduling, purchasing, sales, warranty, and financials. A four-layer context architecture that customized the AI's behavior for each builder without fine-tuning a model. Self-debugging validation, so the system checked its own work before showing a number to an executive. Around it I built a universal builder data schema, the brand and website, and a prospect pipeline of 45+ builders, running demos and pilots personally. One of my co-founders led engineering on the v2 desktop app while I drove product.

Proof: twenty years of a builder, read in an afternoon

For an established homebuilder, I loaded two decades of history out of their construction ERP: every lot, schedule, budget, sales record, and warranty claim the company had ever logged. One analysis pass turned it into the picture the CEO had always run on instinct but never had the numbers to defend.

It ranked the vendors he suspected were trouble and settled it with data: which subcontractors ran chronically late, which drove the warranty backlog, and which quietly outperformed on both. It graded communities and floor plans on conversion and margin, separating the ones carrying the business from the ones dragging it. It found where the cost overruns actually lived, in soft costs and carrying costs tied to long cycle times rather than the trades everyone blamed, and it showed that the schedule baselines had drifted so far from reality they had stopped being useful.

None of this was new to him in spirit. He had an intuition about every one of these. What he had never had was the data to act: to retire a vendor, reprice a plan, reset a baseline. The analysis put a number behind each hunch in a single conversation, and going further cost nothing, because the follow-up questions were free.

Proof: the vendor scorecard

For a pilot builder, we scored a framing vendor against its peers using the builder's own construction data: schedules, work orders, budgets, and warranty records.

The headline finding was one a static dashboard cannot produce. On paper the vendor looked slow. Its lots ran far behind schedule. But the analysis separated the delay the vendor inherited from the delay it caused, and nearly all of the slip predated the vendor's crews ever arriving on site. Adjusted for what it inherited, the vendor finished right at the peer average. It also ran roughly half the rework rate of its peers, at lower cost variance, and its lots moved faster after framing, because a clean frame handoff compresses every downstream trade.

A dashboard shows you the slip. It never shows you the inheritance. The builder was about to judge one of its best vendors by the sins of the schedule it walked into.

The honest ending

Throughline did not work as a business. We ran pilots and had investor conversations; neither converted to committed revenue. Part of it was distribution. We focused on product over distribution, built well, and made almost no real attempt to sell beyond our own network.

The bigger reason was the ground moving under us. We had built a purpose-built agent harness for data analysis, and on one piece of it, pinned and refreshable reports that replace dashboards, we were ahead of the general tools. Then agentic coding broke open. By late 2025 it was clear that Anthropic's own coding agents would blow past a bespoke harness like ours, and a moat built on a capability the model makers are about to ship for free is not a moat. We wound the company down in early 2026.

I would rather say that plainly than dress it up. The thesis held. The company did not.

What survived

The methodology survived intact. Zero-cost analysis is how I run my consulting practice now. The same approach, dropping a messy dataset into an LLM, teaching it the schema, and asking the questions that matter, is what produced A project brain for a nine-figure build.

The rest of the career is at Work.