Status report: the machinery that feeds tax workpapers
What we built for reading a client's paper pile into books, what is left, why it is not on sale yet, and an invitation to talk.
Contents
Hi, Mackenzie's team here. Boardy introduced us: you're looking for AI tax workpaper automation, and we've spent the last month building the machinery that feeds it. Here's where we stand.
Where your week starts
A client's pile of PDFs and photos, three bank exports in different formats, and a return where every line needs a schedule tying it back to a source a reviewer can sign off on. That tie-out is where we aimed.
What we built (October 19 to November 10)
- Ingestion. Clef decides whether a photo or PDF is something needed. Clairic reads what's in it. Clef then sorts each line into the books' categories. When Clef isn't sure, the line goes to an agent, and when the agent isn't sure, it goes to a person. A model's sort is stored as a guess, and the owner's word always outranks it.
- Bank exports in CSV, OFX and QFX.
- Books. From what it reads, it computes a journal, a trial balance and a profit-and-loss, cash or accrual, for any period.
- Tie-out. The part we think you'll care about: tap any figure and the documents behind it appear. Every stored fact records who said it (a document line, the owner, a reviewer or a model's guess), when it's true and when it was written down. That's the audit trail workpapers are made of.
- Cost. Clef and Clairic run on our own machine, so a run's model spend is zero.
Where it stands
On a test pile of a made-up owner's eight months of papers, it read 15 files in 7 seconds. On the first real run, Mackenzie's own October bank export, all three formats read the same 62 lines, with the same dates and the same total. Clef wasn't sure of any of those lines. The agents we'd set up for the test had questions for Mackenzie, but they never reached him that day, so the month didn't close.
The chain itself works. We've closed our own company's books with these same tools, several times: an agent running Claude Sonnet sorted by our business rules and asked Mackenzie about what it couldn't place. Once the papers are read, that part is fairly easy. What the site lacks is its own built-in agent for each business. It's outlined, not built.
What's left, and why
The site's sorting agent, and the tax layer. Ingestion was the hard part. Knowing the business rules of a tax situation in a given area is fairly straightforward; we just haven't decided which area to go to market in. That's also why we've held it off sale: we want a way to market it that doesn't rely on search, because Google holds pages about money to a stricter standard.
Let's talk
That's where you come in. If you work in tax, you already have what kept us off sale: clients who trust you with their money, without needing search to find them. Tell us your region and what else you'd want. You could be the one who decides which one we build first. Happy to chat.