Content authored by practitioners with experience at Amazon India, Intuit QuickBooks, and the Tata Group. Meet the team →
Vendor demos run on clean sample statements, so a lender cannot tell how an analyser will behave on its own files, its own bank mix and its own credit-policy definitions.
Run a fixed two-week pilot on 30 to 50 closed cases with an analyst answer key, compare each field, classify every disagreement as definition, tool error or analyst error, and apply pass rules written before the run.
A pilot set covering approved, rejected and known-bad files across bank categories and channels; a field list for income, obligations, balances and flags; a disagreement log; and a day 1, 3, 10 and 14 timeline.
A signed pilot read-out from the credit head and compliance with field-level agreement, time per file, rejected rows and a clear proceed, configure or stop decision.
A fair pilot of a bank statement analyser uses cases you have already underwritten, a fixed list of fields to compare and success rules you write down before the first file is uploaded. Two weeks is enough to test intake, field accuracy, credit signals, fraud flags and whether your analysts can actually use the report. Anything looser turns into a second demo.
For background on what the analysis should produce, see bank statement analysis in credit underwriting and how to read a bank statement for credit risk.
What goes into the pilot set?
Pick 30 to 50 closed cases from the last 12 months. The mix matters more than the count:
- Approved and rejected cases, roughly in the proportion your book sees, so the tool is tested on borderline files and not only clean ones.
- Three to five known-bad statements from past fraud or tampering cases, masked. The known-bad file test explains how to build these.
- Files from each bank category you lend against: private, public sector, small finance and co-operative banks, including scanned and password-protected PDFs.
- Each intake channel you use: web upload, and Account Aggregator data if you already take it.
- Your credit-policy definitions in writing: how you compute average bank balance, obligations, income and cash deposits.
Keep the data handling clean. The Reserve Bank of India (Digital Lending) Directions, 2025, issued on 8 May 2025, require need-based data collection with borrower consent, and the lender remains responsible for its service providers. Agree the data-sharing terms and the deletion date with the vendor before the files move.
What do you compare?
For every case, the analyst’s recorded result is the answer key. Compare the tool’s output field by field and log each difference with a reason.
| Test | What to give the vendor | What a pass looks like |
|---|---|---|
| Intake | All pilot files as received, including scans | Fewer than 5% of files rejected; every rejection has a stated reason |
| Field accuracy | Statements only, no answer key | Income, obligations and average balance agree with the analyst within your written tolerance, for example ₹500 or 2% per field, whichever is larger |
| Definitions | Your written policy rules | Each metric shows the rows used; differences trace to a definition, not a misread |
| Fraud flags | Known-bad files mixed with clean controls | Altered files flagged with page and row evidence; clean controls mostly quiet |
| Time per file | Nothing extra | End-to-end time clearly below your manual baseline |
| Report use | Two analysts read the reports | Analysts can write a credit memo from the report without reopening the PDF |
Record time per file too. If your analysts take 35 minutes a file today, measure the tool path including any review the analyst still does, not just processing time.
What does the two-week timeline look like?
| Day | Activity | Owner | Output |
|---|---|---|---|
| 1 | Intake check: upload all files | Credit operations | List of accepted and rejected files with reasons |
| 3 | First read-out on 10 cases | Vendor and credit analyst | Early disagreement list; definition gaps raised |
| 5 to 9 | Remaining cases processed; definitions configured where possible | Vendor | Full field-level comparison |
| 10 | Disagreement review | Credit head and analysts | Each difference classed as definition, tool error or analyst error |
| 14 | Decision meeting | Credit head and compliance | Signed read-out: proceed, configure or stop |
Worked example: reading a 30-case result
Illustrative figures. Of 30 closed cases:
- 24 agree on income, obligations and average balance within tolerance.
- 4 differ for definition reasons. In three, the tool counted all credits as income while your policy uses salary credits only; in one, it took the daily average balance where your policy samples the 5th, 15th and 25th of each month. These are configuration questions.
- 2 differ because of tool errors. One scanned co-operative bank statement had a misread amount on page 4; one missed an EMI of ₹7,250 with an unusual narration, which understated obligations on a ₹3,00,000 loan file.
So 24 of 30 (80%) agree as received, and 28 of 30 (about 93%) would agree once definitions are aligned. The decision turns on the 2 tool errors: are they isolated, and did the tool flag low-confidence reads so an analyst would have caught them? Write the pass rule in advance, for example “at least 90 percent agreement after definitions are aligned, and no silent misreads on known-bad files”, and the meeting on day 14 becomes a check, not a debate.
How should the pilot connect to a production decision?
A pilot that passes should leave you with three things: the configured definitions, a list of file types that still need manual review, and the compliance file for the outsourcing arrangement. If the tool passed only after heavy configuration, ask for those settings to be documented as part of the contract. The accuracy signals guide covers what to monitor once the tool is live.
How Terra Insight fits
TransactIQ is a managed bank statement analysis service on Terra Insight’s AWS Mumbai infrastructure, built for degraded PSU and co-operative bank statements and producing 40+ engineered credit signals along with fraud and tampering indicators for review. Evaluating lenders can request a representative-sample fit benchmark before any production integration. See the bank statement analyzer and NBFC lending pages.