Home / DPA Insights / Private Credit
Why Extraction Was Never the Hard Part
We thought reading the numbers off a funding request would be the hard part. It wasn’t. Reading was easy. Knowing what the numbers meant was the real work.
Extraction is the easy ten percent of automating a private credit funding request. Judgment is the rest. A clean pipeline approved a request whose arithmetic was right and whose advance rate was not, because the agreement had been amended. A reviewer caught it, not the machine. Progress came from understanding credit documentation and how deals actually work, not from picking the cleverest model. AI cannot fix an unclear process.
That gap is the whole story. And it took a bad approval, caught by a person and not the machine, to make us see it.
01 / The judgment
The judgment lives in people, not in the process
In private credit, funding requests arrive in all shapes. Every party has its own style. Someone who knows the deal has to hold all of that in their head to get the calculation right.
The judgment lives in people, not in the process. That is exactly what makes it hard to automate.
Our assumption was simple. If we could reliably pull the key fields off the page, the rest would follow. Extraction felt like the mountain. Everything after it felt like plumbing.
02 / Where it breaks
Where plain automation breaks
Real requests are a different story. Formats change from deal to deal. Requests turn up with a stack of supporting documents. Credit agreements get amended, sometimes more than once, and the terms move with them. And an old change to a deal can quietly reshape today’s calculation.
A manual process survives all of this because skilled people adjust in the background. They lean on what they know about each deal. Rigid automation can’t do that. Unclear input just gives you unclear output.
So we tested. As generative AI matured, we ran the leading generative AI platforms across document understanding, orchestration, and reasoning. Each had real strengths. None solved the problem on its own.
03 / The case
The case that changed our minds
One request made the gap impossible to ignore.
It ran through a clean pipeline and came back approved. The arithmetic was right. The advance rate was not. The agreement had been amended twice since the deal closed, and the rate in force was no longer the one in the original document.
What saved us was a person. One of our reviewers looked at the approved request and something felt off. She had worked the deal and remembered there had been changes. She went back to the amendments, found the newer advance rate, and stopped a number that was wrong from going out the door. She had not run a check. She had applied judgment. That is exactly what the machine could not replicate. The model had read the request perfectly and still reached a conclusion she would never have signed.
That was the moment the problem reframed itself. A reviewer like her accounts for amendments and exceptions almost without thinking. Teaching software to do the same is far harder than teaching it to read text off a page. Extraction turned out to be the easy ten percent. The judgment behind it was the rest.
04 / What it taught us
What the failure taught us
The real problem was never reading. It was judgment. Once we accepted that, the questions changed.
We stopped asking how to extract fields faster. We started asking what a good reviewer actually does. Which agreement governs this request? Has it been amended? Do the numbers reconcile? Which exceptions need a person? Those aren’t extraction questions. They are reasoning questions, and they only make sense if you understand private credit operations from the inside.
Three things became clear.
- Business understanding mattered more than the technology. Progress came from knowing credit documentation and how these deals actually work, not from picking the cleverest model.
- The real complexity lived in the exceptions — the amendments, the odd formats, the historical changes that quietly reshape a calculation.
- AI, for all its capability, could read anything we gave it. It could not, on its own, know which reading mattered.
And often the gap wasn’t even technical. It was process clarity. AI can’t fix an unclear process. If a person would struggle to say which agreement governs a request, no model is going to guess its way to the right answer.
05 / Beyond private credit
The lesson travels well past private credit
Anywhere expert judgment sits between a document and a decision, whether in underwriting, trade finance, or claims, the temptation is the same. Automate the reading and call it done.
The reading is the easy part. The judgment is the work.
That reframing set up everything we built next. Once we knew the problem was judgment and not extraction, we could design for it. The system didn’t need to read better than a person. It needed to answer the questions a good reviewer asks, flag what deserved a second look, and leave the real decisions with people.
The reading is the easy part. The judgment is the work.
How we turned that into a working platform, built around specialized agents that work the way a review team does, is the subject of the next field note.
Next — Field Note #02Building for Judgment, Not Just Reading
One conversation. No tooling pitch.
Where does judgment sit in your credit process?
If your team is weighing automation for funding requests or other document-heavy credit workflows, we will walk through where judgment actually sits in your process before anyone talks about tooling.
Book a working sessionFrequently asked questions
Is document extraction the hard part of automating private credit funding requests?
No. Extraction turned out to be the easy ten percent. Modern models read funding requests accurately. The hard part is judgment: knowing which credit agreement governs the request, whether it has been amended, whether the numbers reconcile, and which exceptions need a person. Those are reasoning questions, not extraction questions.
Why do amended credit agreements break automated funding request processing?
Credit agreements get amended, sometimes more than once, and the terms move with them. A pipeline can read a request perfectly and calculate it correctly, and still produce the wrong answer, because the advance rate it used came from the original document and the rate in force has since changed. The arithmetic is right; the input is not. An old change to a deal can quietly reshape today’s calculation.
What does human-in-the-loop mean in private credit operations?
It means the system does not need to read better than a person. It needs to answer the questions a good reviewer asks, flag what deserves a second look, and leave the real decisions with people. Experienced reviewers account for amendments and exceptions almost without thinking, and that judgment is what the machine cannot replicate on its own.
Does this lesson apply outside private credit?
Yes. Anywhere expert judgment sits between a document and a decision — underwriting, trade finance, claims — the temptation is the same: automate the reading and call it done. The reading is the easy part. The judgment is the work.
