The 80% Problem: Why Your Strongest Credit Signal Never Reaches the Memo

By Steve Iskander, Founder and CEO, Intrepid

In this article

• Volume is not the same as value
• How context evaporates
• Making the 80% underwritable
• What to do about the 80%

Someone pushed back on my take on unstructured data in lending last week, and they were right to.

I had said that most of a lender’s decision data is unstructured. The reply: everything my risk team actually uses is already in a database.

That is a fair challenge and worth taking seriously.

Here is the distinction I would draw. Gartner and IDC have estimated for years that 80% or more of enterprise data is unstructured. That is a statement about volume, not importance, and plenty of that volume is genuinely noise. Nobody needs to underwrite off a video file.

But run the exercise on a single credit file and the picture changes.

The financials are structured. The spreads are structured. What is not structured: the explanation for why Q3 dipped, the customer concentration buried in an aging report, the covenant language sitting in a PDF, the email where the borrower flagged a supplier change, the site visit notes, the reason the last lender passed. This is the unstructured data in lending that moves a credit decision.

Now ask which of those changed your last credit decision.

In my experience it is rarely the ratio. It is the context around the ratio.

That context does exist. Someone read it. But it survives as a paragraph in a memo, then as a sentence in a committee summary, then as a number in a spreadsheet. Each step is reasonable. The compression is invisible. And when that borrower comes up for renewal in eighteen months, the person who read the file has moved on.

This is what I mean when I say the data is the bottleneck. Not that lenders lack data. That the highest-value part of it is stored in a form that only survives if a human remembers it.

The fix is not a smarter model reading the memo. It is making the underlying material machine-readable in the first place, so context travels with the credit instead of evaporating at each hand-off.

That is the layer we built Intrepid around. Financial data ingested in real time, structured, ready to underwrite against, so a credit team is reasoning about the business instead of reassembling it. Outcomes vary by portfolio and process, but the pattern is consistent: when inputs hold their shape, the analysis gets sharper.

The 80% number is worth arguing about. The compression problem is not.

For the credit teams reading this: what is the one piece of context about a borrower that you know matters and that lives nowhere but in someone’s head?

Volume is not the same as value

The often-cited figure that roughly 80% of data is unstructured comes from Gartner and IDC and refers to enterprise data by volume. Taken literally, a lot of that volume is genuinely noise. The sharper point for lenders is not about volume at all. It is that the specific pieces of information most likely to change a credit decision tend to arrive as prose and documents rather than as clean fields. A single number in an aging report, a sentence in a management letter, or a paragraph in an email can move a decision more than an entire page of ratios. That is why treating the 80% as a volume problem misses the point. It is a value problem.

How context evaporates

Context does not disappear all at once. It leaks. An analyst reads the full file and understands why a quarter looked soft. That understanding becomes a paragraph in a memo, then a sentence in a committee summary, then a figure in a spreadsheet. Each compression is reasonable in isolation, and each one strips a little nuance. By the time the borrower comes up for renewal, the person who did the original reading has often moved on, and the team reconstructs from the thinnest surviving artifact. The information existed. The system simply had no way to carry it forward.

Making unstructured data in lending underwritable

The fix is not a smarter reader of memos. It is making the underlying material machine-readable at the point it enters the file, so the context travels with the credit rather than being summarized away. When bank statement detail, aging reports, correspondence, and covenant language are structured on arrival, a renewal starts from the accumulated record instead of a blank page, and monitoring can watch the same signals continuously. The 80% stops being background and becomes something you can actually underwrite against.

What to do with unstructured data in lending

The practical response to the 80% problem is not to read more documents faster. It is to change where structuring happens. Instead of a human reading a file and compressing it into a memo, the material is turned into consistent, machine-readable fields at the point it enters the pipeline, with the original context preserved and linked. Bank statement detail, aging reports, correspondence, and covenant terms all become queryable rather than buried. Two things change immediately. Renewals stop starting from a blank page, because the structured record of the original underwrite is still there. And monitoring becomes continuous, because the same live fields can be watched over time. The 80% stops being an archive nobody reads and becomes a working part of the decision. None of this removes judgment. A slowing receivable can be a warning or a healthy new customer on longer terms. The point is to see the signal while you can still ask about it, instead of reconstructing it a year later from memory.

Related reading

→ Why AI is not the bottleneck in lending

→ How to monitor a loan portfolio after close

Frequently asked questions

Is 80% of lending data really unstructured?

The widely cited figure that 80% or more of data is unstructured comes from Gartner and IDC and describes enterprise data in general, by volume, not lending data specifically. It is a useful directional signal, not a precise measurement of any one credit file. The more important point for lenders is qualitative: the information that most often changes a credit decision tends to live in documents and prose rather than in structured fields.

What counts as unstructured data in a credit file?

Unstructured data includes anything that is not already a clean, queryable field: bank statement narratives, scanned tax returns, debt schedules in spreadsheets, covenant language in PDFs, and the email where a borrower explains a one-time dip or a supplier change. The financial ratios may be structured, but the context around them usually is not.

Why does context get lost between underwriting and renewal?

Context is lost because it survives as a paragraph in a memo, then a sentence in a committee summary, then a number in a spreadsheet. Each compression is reasonable on its own, but the nuance does not carry forward. By the time a borrower comes up for renewal, the person who read the original file has often moved on, and the reasoning has to be reconstructed.

How does Intrepid help preserve credit context?

Intrepid ingests financial data in real time and structures it to a lender’s credit policy, so context travels with the credit instead of evaporating at each hand-off. The decision stays with the credit team. Outcomes vary by portfolio and process.

Is unstructured data useful for credit decisions?

Often it is the most useful data you have. The context that explains a number, why a quarter dipped, who the concentrated customers are, what a borrower flagged, usually lives in documents and prose rather than structured fields. Making that context machine-readable is what lets it reach the decision.

Sources: Gartner · IDC

Intrepid makes the context that drives a credit decision machine-readable, so it reaches the decision. See how at intrepidfinance.io.

Published by Intrepid. Democratizing Access to Capital. intrepidfinance.io