Securitization Is Leaving the Spreadsheet. It Shouldn't Land in a Black Box
And we're making sure it doesn't.
August 27, 2026
•
5 min
.svg)
And we're making sure it doesn't.
August 27, 2026
•
5 min
There's a version of our product we could have shipped a year earlier. Wire up a large language model, point it at a securitization agreement, let it configure the deal, and demo the magic. It would have looked incredible in a pitch. It would have been irresponsible to put in front of anyone managing real money.
Here's the thing about securitization: this industry already knows exactly what it feels like to run trillion-dollar structures on a system nobody can fully inspect. It's called Excel. Ask anyone in a treasury or capital markets seat and you'll hear the same stories we heard in months of discovery calls: one broken formula silently mispricing an entire deal, two analysts rebuilding the same report every month, errors that don't surface until the first payment date. The spreadsheet was this industry's first black box. It just never marketed itself as one.
So when we set out to replace it, we held ourselves to a simple rule: whatever comes next has to be more inspectable than what it replaces, not less. An AI system that produces answers no one can trace doesn't improve on the spreadsheet. It just rebrands the problem.
That rule matched a conviction I've held since before this company existed: you earn the right to add AI by first building the data foundations and manual workflows until they're solid. In a market sprinting in the other direction, that order looks backwards. We followed it anyway, and built almost everything except the AI first.
Most of the conversation about "responsible AI" in finance focuses on the model layer: hallucination rates, guardrails, prompt design. Those matter. But in a regulated workflow, the decisions that actually determine whether AI is safe to deploy get made much earlier, in parts of the system nobody puts in a demo video.
Here's what I mean. In a spreadsheet, every number looks equally true. A verified figure, a stale link, and a hand-keyed guess all render in the same confident font, and nothing about the cell tells you which one you're looking at. Supervision starts with breaking that illusion. A system built for oversight knows whether a person has signed off on a value, remembers where every figure came from and who touched it, and keeps a closed period closed, so last month's report can't quietly change underneath you. None of that is AI. All of it is what makes a number trustworthy.
And if a platform can't earn that trust for a single number keyed in by a human, it has no hope of earning it for hundreds of fields extracted by an AI model from a 200-page agreement. You cannot supervise AI with infrastructure that was never designed to supervise anyone.
For the first stretch of building the OFS platform, that meant doing the unglamorous work up front. We built the manual flows before we automated any of them: the way an analyst configures a deal by hand, uploads a loan tape, runs a waterfall, records a covenant waiver. Not as a stopgap, but as the foundation.
That foundation gave us three things no AI model could have given us.
From the beginning, every meaningful value in OFS carries its status and its history. Nothing is just a cell in a grid, looking exactly as true as its neighbors. That sounds like bureaucratic plumbing until you introduce AI, and suddenly it's the entire trust mechanism.
Our Tile Builder lets your team construct deal logic (borrowing bases, eligibility tests, concentration limits) as a visual flow of nodes instead of spreadsheet formulas, because the logic that runs your deal shouldn't live in cell H47 of a workbook only one person understands. Every step is inspectable and audited.
Every action in the platform is written to an append-only log: every data change, approval, waiver, and period close. Closed periods lock, with every input frozen in a snapshot, so a reporting cycle never silently restates. The system doesn't trust us, its builders, to quietly fix things after the fact. That's the standard AI would have to meet.
All that transparency is what let us automate more. Once we realized we knew enough about each deal's structure to run the entire month-end distribution in-app, we let ourselves ship it under one condition: every computed figure has to open. Click any number in a completed waterfall and you land on the exact calculation snapshot that produced it: the balance, the rate, the day count, laid out on the same canvas your team already reads. Wrong number? Correct the input and the period restates, explicitly and on the record. In this system, peace of mind isn't a promise. It's the ability to look.
Once those rails existed, adding AI stopped being a leap of faith and became almost boring, in the best possible way. Every AI feature in OFS is a proposer, not a decider, feeding into the same review pipeline a human's work does.
When our AI reads a securitization agreement, it doesn't configure the deal. It proposes a configuration. I'll be honest: this wasn't a grand design principle at first. Early on, I spent weeks trying to make extraction from a 200-page agreement accurate enough to trust, and it never felt like enough, because no accuracy rate survives contact with a skeptical analyst. The breakthrough had nothing to do with the model. It was realizing the AI should have to cite its source.
Once every extracted value arrived with a citation to the exact clause and page it came from, plus a confidence score, everything fell into place. The analyst doesn't have to take the machine's word for anything, and doesn't have to scroll through 200 pages to verify it either. The evidence sits one click from the value. They review each field, approve or correct it, and only then does anything go live, with every approval logged. When the AI can't find something, it says so and hands the field to a person, rather than guessing.
The same pattern repeats everywhere. Hand the platform the Excel model your team runs a deal on today, and it comes back as Tile Builder nodes your team reads, reviews, and approves before anything runs on them. Ask the in-platform assistant about any deal, and its answers are built from the platform's own numbers, the same figures you can drill into by hand.
The result is a system where AI removes the re-keying, the re-building, and the weeks of setup, and removes none of the accountability. Judgment stays with people. The audit trail stays complete. The magic is real, but it's checkable. That, to us, is what AI-native actually means: not a model bolted onto the front of a product, but a whole system built so AI's work can be reviewed, approved, and audited like anyone else's.
I won't pretend this sequencing was free. Building deterministic infrastructure before flashy features is slower, and in a market where every competitor's homepage now says "AI-powered," slower is uncomfortable. There were weeks where the most important thing my team shipped was a period-locking mechanism nobody will ever screenshot.
But the discipline pays for itself the first time a practitioner asks the question they always ask:
For decades, the honest answer in this industry has been "because someone checked the spreadsheet, hopefully." With OFS, the answer is never "trust the model," either. It's: click the number. Read the clause it cites. Check who approved it and when. The system was built to be doubted, which, in my experience, is exactly what makes people comfortable relying on it.
The industry is going to spend the next few years replacing its spreadsheets. That's no longer in question. The question is what replaces them, and whether the people running these deals will be able to see inside it. My bet is that the winners in this category won't be the platforms with the most impressive AI. They'll be the ones whose software was built, from its first line of code, to show its work.
That's the standard we hold ourselves to. Securitization is leaving the spreadsheet. We're making sure it doesn't land in a black box.
Want To Talk To The People Behind It?
45 minutes with William - no sales team, no deck you haven't asked for. Whether you're evaluating an investment or exploring the platform for your team.