How multi-source data analysis cuts SME credit decisions from days to hours

Ask a lender where a five day decision goes and most will point at the credit committee. In our experience the committee is rarely the problem. The time goes to an analyst working through six months of bank statements line by line, rebuilding a cash flow picture with a calculator and a highlighter, usually three of the five days, before anyone with approval authority has even seen the file.

We built DocuMind to take over that step. This post walks through how it works in practice: where the parsing happens, what the confidence scores are for, and which of the extracted signals turn out to predict anything.

From PDF to signal set

Statements arrive in whatever format the bank produces: a clean CSV export if you are lucky, more often a scanned PDF, sometimes a photograph of a paper passbook. Extraction runs in three passes: layout detection first, then transaction parsing, then counterparty resolution. Every field comes out with a confidence score attached. Anything below threshold does not go into the decision; it goes into a queue for a human to check.

Whatever went in, the same schema comes out:

{
  "avg_balance_6m":      { "value": 1842000, "confidence": 0.98 },
  "inflow_monthly_avg":  { "value": 5210000, "confidence": 0.97 },
  "bounced_items_6m":    { "value": 2,       "confidence": 0.99 },
  "recurring_obligations": [
    { "counterparty": "lessor", "monthly": 240000, "confidence": 0.91 }
  ]
}

The signals that matter

We extract far more than we score on. Across our production volume, a fairly small set of signals ends up doing most of the predictive work:

Signal	What it captures	Weight
Balance floor, 6m	Liquidity buffer under stress	high
Bounce recency	Recent payment stress beats historical counts	high
Inflow concentration	Dependence on a single payer	medium
Obligation ratio	Recurring commitments against average inflow	medium

Weights are illustrative placeholders; production weights are scorecard-specific.

Where the hours go

Once extraction is automated, what remains on the decision clock is the work that genuinely needs time: a referral review, a site visit, a committee call on a marginal file. The mechanical part happens in minutes. Nothing about the committee got faster; the queue in front of it got shorter. On straightforward files, that turns a five day decision into a four hour one.

The analyst role does not disappear, but it moves. Instead of reading every statement, analysts review the flagged ones: low confidence extractions, contradictory signals, files the policy rules referred out. In practice, a single analyst can handle volume that used to need a team of ten.