Compressing a two-week study into an afternoon

A cost segregation study used to take one to two weeks. It involves reading construction documents, invoices, and floor plans, then classifying every asset into the depreciation categories the IRS recognizes. It is slow, expensive, and almost entirely bounded by how fast a human can read.

We rebuilt it to run in a few hours. The interesting part is that the language model was the smallest piece of that.

The model is the easy part

Handing a document to an LLM and asking for a classification is a demo. Making it hold up against a regulated output is a different problem. Three things mattered more than model choice:

Ingestion determines everything downstream. Real-world construction documents are scans, mixed-orientation PDFs, spreadsheets exported to PDF, and photographs of invoices. If the extraction layer is noisy, no amount of prompt work recovers the signal. Most of our early accuracy gains came from work that had nothing to do with AI.

Classification needs to be auditable, not just correct. An IRS-aligned report has to show its reasoning. That pushed us toward a design where every classification carries its source span — the specific line item and document it came from — so a human reviewer can verify a decision in seconds instead of re-reading the source.

The human stays in the loop by design. The goal was never full automation. It was to move the expert from reading everything to reviewing exceptions. That reframing changed what we optimized for: not raw accuracy, but confident separation between the cases a human needs to see and the ones they don’t.

What I’d tell a team starting something similar

Spend your first month on the boring layer. Document ingestion, normalization, and a review interface that makes verification fast will outlast every model you swap in behind them. The model tier is the part you’ll change three times; the data path is the part you’ll live with.