A confident hallucination on a loan file does not look like an error. It looks exactly like a correct decision, right up until the loan goes non-performing. That is the quiet failure mode of general-purpose AI on a bank statement. And it is the reason document intelligence is suddenly the most interesting category in the enterprise AI stack.
Key takeaways
The reading layer, not the reasoning layer, is where enterprise AI is breaking. Frontier models score below 50 percent on real enterprise document reasoning tasks according to Databricks AI Research.
Confident hallucinations on numbers are the quiet failure mode. A bank statement total misread as a smaller number looks clean, passes review, and books a bad loan.
Document intelligence is becoming a distinct product category. Purpose-built intelligent document processing is replacing the default "use a frontier LLM to extract fields" approach.
Mid-market lenders need a version of this that enterprise platforms don't deliver. Accuracy, operations-owned workflows, fast time to value, and flat pricing. Four criteria rarely found together above the enterprise tier.
The bank statement that reads as right but isn't
Earlier this month, the CEO of Databricks made a public post that captured the document intelligence problem in a single line: if a large language model reads a receipt that says $3 + $4 = $8, it may parse it as $3 + $4 = $7, because its job is to predict the next token, not to verify the arithmetic.
In a consumer context, that is a rounding error.
In a credit application where the bank statement aggregates to seven million rather than the three million shown in the summary box, it is a different conversation. A downstream workflow that trusts the parse books a loan against inflated income. No one catches it in review because the document looks clean. The next flag comes from the collections team months later, when the borrower misses a payment against obligations they never had the cash to cover.
This is the quiet failure mode of general-purpose AI on lending documents. Not an obvious error, not a missing field, not a confidence score below the threshold. A clean-looking wrong number that reads as right.
The reading layer is the accuracy ceiling
Databricks AI Research published a benchmark earlier this year called OfficeQA, designed around real enterprise document tasks. Even frontier agents, the same models that can pass legal exams and debug complex code, scored below 50 percent on document reasoning. The bottleneck was not reasoning. It was reading.
That finding lines up with what lending operations teams have been observing for two years. The reasoning capability of frontier models has outpaced almost every prediction. The reading capability has not. Transcription of handwritten notes, parsing of cropped scans, reconciliation of totals across multi-page statements, extraction from tables with merged cells: these are still where the error rate climbs.
The implication is uncomfortable for the default LLM-first approach to document workflows. A reading problem cannot be solved with more reasoning. An agent that confidently misreads a number and then reasons brilliantly about the wrong number is not more useful than one that reads correctly and reasons adequately. This is why intelligent document processing is having a moment: it solves the reading problem before the reasoning problem even shows up.
Document intelligence meets intelligent document processing
The market is adjusting. Over the past year, the language has shifted from "use a frontier model to extract fields" to "use a purpose-built document processing layer, then let the agent reason over the structured output." The Databricks Document Intelligence announcement last week is the clearest expression of this yet: a dedicated product positioned as the foundation that every agentic workflow above it depends on.
The significance is not that Databricks built the product. It is that they judged the category large enough, and the reading problem structural enough, to put serious research and engineering behind it. Anthropic, Google, and Microsoft have made adjacent moves over the same period.
A new category is forming that sits between raw OCR at the bottom and general-purpose language models at the top. Its job is to convert messy documents into reliable structured data at accuracy levels that clear the bar for regulated workflows, and to do it consistently enough that the agents built on top can be trusted with credit decisions, claims, and compliance. This is what intelligent document processing has been quietly becoming, and what the market is now finally naming.
What mid-market lenders actually need from a bank statement workflow
The category is forming, but the available products are not evenly distributed across the market.
Enterprise platforms like the one Databricks just launched assume a lakehouse, a data engineering team, and the appetite for a platform integration project. That works for a bank with a thousand-person technology organization. It does not work for a non-bank lender in the Philippines processing two thousand loans a month, or a cooperative in Indonesia running origination through a small operations team with no SQL.
Mid-market lending needs four things that enterprise intelligent document processing rarely delivers together:
Accuracy on the documents that matter most. Bank statements, payroll certificates, business registration papers, tax returns, and identity documents. The bank statements and loan packages that dominate mid-market lending in Southeast Asia are the exact document types general-purpose models handle worst.
Ownership by operations, not IT. The person who understands how a loan package is reviewed is the same person who should be able to change a validation rule, add a document type, or reroute an exception. A no-code workflow builder is not a nice-to-have in this segment. It is a deployment precondition.
Fast time to value. Mid-market lenders do not have eighteen months. A document automation project that cannot show production throughput inside a quarter loses internal sponsorship before it gets there.
Economics that scale the right direction. Per-page pricing compounds painfully as volumes grow. Flat, volume-predictable pricing is structurally better fit for loan origination operations with consistent monthly throughput.
The signal from the top
When a company of Databricks' scale commits research resources and product surface to document intelligence, two things are true at once. The first is that they have identified a large enterprise opportunity and are moving to capture it. The second is that the underlying problem is real, the market is growing, and the segment below enterprise is materially underserved.
The lender processing two thousand loans a month in Manila has the same reading problem as the one processing two hundred thousand in London. The architectural answer is the same: a purpose-built reading layer that produces reliable structured data, and a workflow layer above it that matches how lending actually operates. What differs is the buying motion, the integration surface, the cost envelope, and the tolerance for implementation effort.
Floowed was built for this segment. Not because the technical problem is different, but because the surrounding context is. Lending teams across Southeast Asia need the accuracy that enterprise intelligent document processing platforms deliver, packaged and priced in a way that fits how their operations run.
The category is real. The validation is now public. The question for mid-market lenders is no longer whether to invest in document intelligence. It is which version of it actually fits the way you run your business. Book a demo if you want to see what that looks like in practice.
Frequently asked questions
What is document intelligence, and how is it different from OCR?
Document intelligence is an emerging category that sits between raw OCR at the bottom and general-purpose language models at the top. OCR converts pixels to characters but has no concept of meaning. Large language models understand meaning but hallucinate specific values like amounts and dates. Document intelligence combines layout-aware parsing with structured extraction to produce reliable, auditable output that downstream workflows can trust. The Databricks Document Intelligence launch in April 2026 was one of the clearest public signals that the category is consolidating into a distinct product layer.
Why do LLMs get numbers wrong on bank statements?
Large language models are trained to predict the next token based on patterns in their training data. When they encounter a number in a document, they do not verify it against the arithmetic of the page. If the visual context is ambiguous (cropped scans, handwritten entries, merged cells, faint print), the model fills in what is statistically likely rather than what is actually on the page. In a consumer context this is minor. In a credit decision where a misread of a bank statement total becomes the basis for an approved loan, it is a material risk.
Is Databricks Document Intelligence a fit for mid-market lenders?
Databricks Document Intelligence is built for enterprises already running the Databricks lakehouse with a data engineering team and an appetite for a platform integration project. For mid-market lenders processing hundreds to a few thousand loans a month without dedicated data infrastructure, the implementation model does not match the buying motion, the cost envelope, or the time-to-value requirement. Purpose-built intelligent document processing platforms sized for mid-market lending operations are generally a better fit for that segment.
What should mid-market lenders look for in document intelligence software?
Four criteria matter most: accuracy on the specific document types that drive credit decisions (bank statements, payroll certificates, tax returns, identity documents), workflow ownership by operations rather than IT, fast time to production throughput (within a quarter rather than a year), and flat predictable pricing rather than per-page economics that compound at scale. These criteria cut against enterprise-scale document intelligence products and in favor of purpose-built intelligent document processing platforms designed for lending operations.
%20(1).png)
