Loading header...

Home / DPA Insights / Commercial Lending

Beyond LLMs: Building Reliable AI for Commercial Lending

AI can read the document. But can it tell you which actually matters?

The short version

Extracting data from a document is not the same as understanding its meaning and relevance to a lending workflow. Many “extraction” problems in rent rolls and operating statements are really business-rule problems: which number counts. DPA’s workflow runs AI Extraction → Business Rules → Validation → Human Review → Trusted Data. AI interprets, rules decide, people handle exceptions. On one ordinary row, the rule gives $4,300 Gross Rent. A model that added every figure on the row reported $10,300. The goal is human-efficient processing, not human-free processing.

From this buildOne row, two answers
$4,300 Gross Rent the rule gives: Monthly Base Rent + Monthly Other Income
$10,300 what a model reported after adding every figure on the row
3 months the wrong figure passed through extraction unchecked

A lender’s team had been working from a gross rent figure that was wrong. Not obviously wrong. The document had processed cleanly, the numbers looked reasonable, and nothing had flagged an error. The LLM had added every dollar figure on the row. It was simply adding the wrong things.

This is not an edge case. The Mortgage Bankers Association reports $875 billion of US commercial and multifamily mortgages are scheduled to mature in 2026. Every refinance needs fresh rent rolls and operating statements, and lenders need numbers they can defend, not just numbers that arrived quickly. Extracting data from a document is not the same as understanding its meaning and relevance to a lending workflow.

The problem: three questions lenders keep asking

Problem 1: Which number counts? Rent Rolls and Operating Statements look structured, but Gross Rent may be stated directly in one file, built from several components in another, and need certain charges excluded in a third. Many “extraction” problems are really business-rule problems.

Problem 2: Can we trust the number? A model that helpfully adds every dollar figure on a row can be wrong by a wide margin, and nothing on the page warns you. A confident answer is not always a correct one.

Problem 3: What happens when the document breaks the pattern? For example: vacant suites carrying an applicant’s name, duplicate “shadow” rows, evicted tenants, and page batches that return too few rows.

DPA’s solution: AI interprets, rules decide, people handle exceptions

DPA does not send a document straight from an LLM to a final output. Every step has a defined role, and each one checks the step before it:

AI Extraction → Business Rules → Validation → Human Review → Trusted Data

Which number counts? The LLM reads varied layouts, labels and multi-row units. A fixed rule then decides what counts: Gross Rent = Monthly Base Rent + Monthly Other Income, with Cost Recovery and deposits always excluded. In one residential format, Base Rent includes only lines coded RENT, SUBRENT, RENT-HP or HRV, and fees such as parking, pet, trash and utilities are left out.

Can we trust the number? Every Rent Roll and OSAR passes through DPA’s Statistical Quality Control (SQC) engine before delivery. Key totals are re-added from the source file where possible and reconciled against the prior period, alongside completeness, lease-date and business-rule checks. Rent and GLA outliers are flagged using variance and Z-score analysis, so analysts review only true exceptions.

What if the document breaks the pattern? Fixed rules rejoin split unit IDs, remove shadow rows, record vacant suites as VACANT and treat evicted tenants as vacant. Anything the rules do not cover goes to a reviewer instead of being silently accepted.

One Row, two possible answers

The case in the opening came from one portfolio review. The lender’s team had been working from a figure that was $6,000 per unit higher than the actual gross rent, and it had passed through extraction unchecked for three months. The row behind it was ordinary: $4,000 base rent, $300 other income, $1,000 cost recovery and a $5,000 security deposit. Labels change from file to file (“Rent,” “Base Rent,” “Contract Rent”), and reading past them is where the LLM helps. The rule then decides what counts: $4,000 + $300 = $4,300 Gross Rent. The model had added every figure on the row and reported $10,300. Nothing on the page warned anyone.

Validation closes the loop. Unit values are added up and compared with the total printed on the rent roll, and operating statements get the same treatment: totals and NOI are recalculated and compared with what the document reports. If they do not tie, a reviewer sees it.

Table 01 / One ordinary row, two possible answers
On the row Amount Does the rule count it?
Base rent $4,000 Yes. Monthly Base Rent
Other income $300 Yes. Monthly Other Income
Cost recovery $1,000 No. Always excluded
Security deposit $5,000 No. Always excluded
Gross Rent, by the rule $4,300 $4,000 + $300
What the model reported $10,300 It added every figure on the row

What DPA has achieved

  • Rules carry the weight where the answer is defined. Mapping every output field of a live rent roll pipeline showed that only some fields are pure AI reading tasks. Several come from the file name or master data, and for Gross Rent the rules carry more weight than the model.
  • The model’s answer is a check, not the final word. Totals are re-added from the source file, so a wrong figure is caught instead of passed on.
  • People focus on exceptions. Unusual layouts and conflicting information are routed to a reviewer. The goal is human-efficient processing, not human-free processing.

A system that processes every document automatically but sometimes returns a wrong value with no sign of the error can create more risk than value. DPA’s workflow automates most of the work, validates critical outputs and flags exceptions, so analysts get confidence in the data they are using, not just automation.

$875 billion in maturities. Every refinance needs numbers a lender can defend. That is not an extraction problem; it is a reliability problem. And reliability requires more than a confident model.

Can you defend the numbers in your rent rolls?

If your team processes rent rolls and operating statements for refinances or portfolio reviews, we will walk through which numbers count, how they are validated, and where exceptions should go to a reviewer.

Book a working session

Frequently asked questions

Is extracting data from a rent roll the same as understanding it?

No. Extracting data from a document is not the same as understanding its meaning and relevance to a lending workflow. Rent rolls and operating statements look structured, but Gross Rent may be stated directly in one file, built from several components in another, and need certain charges excluded in a third. Many extraction problems are really business-rule problems.

How should Gross Rent be calculated from a rent roll?

In DPA’s workflow, a fixed rule decides what counts: Gross Rent = Monthly Base Rent + Monthly Other Income, with Cost Recovery and deposits always excluded. On a row with $4,000 base rent, $300 other income, $1,000 cost recovery and a $5,000 security deposit, Gross Rent is $4,300. A model that adds every figure on the row reports $10,300.

How does DPA validate rent roll and operating statement data?

Every Rent Roll and OSAR passes through DPA’s Statistical Quality Control (SQC) engine before delivery. Key totals are re-added from the source file where possible and reconciled against the prior period, alongside completeness, lease-date and business-rule checks. Rent and GLA outliers are flagged using variance and Z-score analysis, so analysts review only true exceptions.

What happens when a rent roll breaks the pattern?

Fixed rules rejoin split unit IDs, remove shadow rows, record vacant suites as VACANT and treat evicted tenants as vacant. Anything the rules do not cover goes to a reviewer instead of being silently accepted.

Does reliable AI mean removing people from the process?

No. The goal is human-efficient processing, not human-free processing. Unusual layouts and conflicting information are routed to a reviewer, so analysts focus on exceptions.

Malav Shah
Malav Shah

Manager - Research and Data Operations,

Decimal Point Analytics Pvt Ltd

Malav Shah is Manager - Research and Data Operations at Decimal Point Analytics. This article is about why rent roll extraction is a reliability problem, not just an extraction problem.