10decodersDocument intelligence for fintech in India
Your credit team is not slow because the people are slow. It is slow because a human being has to open a 38-page scanned bank statement, find eleven numbers, and type them into a form. This is what changing that actually costs, what it returns, and what the RBI and the DPDP Rules expect you to be able to prove while you do it.
The document bottleneck in Indian fintech
Every Indian lending, payments and wealth business runs on documents it did not design and cannot control. A borrower uploads a bank statement as a photograph of a laptop screen. A merchant sends a GST return as a password-protected PDF. A branch scans a property deed at 150 dpi with the stamp paper half out of frame. Somewhere downstream, a person opens each one and reads it.
Document intelligence is the layer that reads those files first, extracts the fields you actually need, tells you how confident it is about each one, and routes only the uncertain cases to a human. It is not new technology. What has changed is that the economics finally work at Indian ticket sizes, and that two regulatory clocks now make the manual alternative expensive in a way it was not two years ago.
This piece covers six questions in order for CTOs, heads of credit operations, and COOs:
- Why bother at all, and what the return actually looks like
- Where it genuinely pays inside an Indian fintech stack
- Why you still need this if you already have a language model
- What the security and regulatory picture demands
- What it costs per page, including the line items nobody quotes
- How to get your own team competent rather than dependent
Why even try this? The return, honestly stated
The lazy version of this argument is a percentage. Someone tells you automation cuts processing time by seventy per cent and you are supposed to nod. We are not going to do that, because the number depends entirely on your document mix and you can compute it yourself in our calculator below.
The Account Aggregator did not solve this for you
The most common objection we hear from Indian lenders is that Account Aggregator makes document reading obsolete. Structured, signed, consented data arrives from the source bank; nobody has to parse a PDF.
That is true for the borrowers it reaches. Industry analysis published in 2026 places AA-enabled borrower coverage at roughly 38 per cent as of December 2025, with materially lower penetration across cooperative banks, regional rural banks and tier-2 and tier-3 geographies. Vendors reporting on AA-only lending programmes describe high application failure rates at the data-collection stage and significant applicant drop-off when a borrower is told their bank is not supported.
The realistic 2026 architecture is not AA or documents. It is both routes into the same decision engine, producing the same quality of output regardless of how the data arrived.
The three returns that are actually bankable
Where it pays inside an Indian fintech stack
Not every document workflow is worth automating. The ones that are share three traits: high volume, a stable set of fields you need every time, and a downstream decision that is currently waiting on a human to type.
| Workflow | What gets extracted | Why it is worth doing first |
|---|---|---|
| Bank statement analysis | Transaction lines, balances, salary credits, EMI debits, bounce records, counterparty patterns | Highest page volume, most manual minutes per file, directly gates disbursal |
| KYC and identity | PAN, masked Aadhaar, passport, driving licence, voter ID, address proof, photograph match | Every applicant passes through it; failures here block the whole funnel |
| GST returns and ITR | Turnover, filing period, tax paid, input credit, declared income, Form 26AS entries | The cross-check against bank credits is one of the strongest fraud signals in MSME lending |
| Cheque and NACH mandates | MICR line, IFSC, account number, signature region, mandate amount and frequency | Rejection rates on mandate registration are a silent, recurring operational cost |
| Collateral and property files | Sale deed parties, survey number, encumbrance entries, valuation figures, insurance cover | Longest documents, slowest reviews, highest concentration of legal risk |
| Trade and invoice finance | Invoice line items, LC terms, bill of lading, e-way bill reference, port and incoterm | Discrepancy checking is rule-based and repetitive, which is exactly what machines do well |
| AML and sanctions evidence | Entity names across scripts, beneficial ownership chains, adverse media extracts, filing references | Name matching across Indian transliteration variants is a hard problem worth solving once |
| Chargeback and dispute packs | Merchant evidence bundles, delivery proofs, terms acceptance records, transaction references | Network deadlines are fixed; missing them costs money regardless of the merits |
A sequencing note
Start with bank statement analysis if you are a lender, KYC if you are a payments or wealth platform. Both are high volume with a stable field set, which means you learn the operating model on a workflow that pays back inside the first quarter. Property and trade documents are more valuable per file but much harder, and they are a poor place to build your first pipeline.
Why use document intelligence when you already have Claude?
This question comes up in almost every scoping conversation now. You can hand a PDF to a frontier language model and get remarkably good extraction back. So why pay for a second thing? Because they are not the same layer, and the failure modes are different in ways that matter specifically in regulated lending.
| What you need | Document intelligence service | General language model |
|---|---|---|
| Provenance | Returns page number and bounding-box coordinates for every extracted value | Returns the value. Asking where it came from produces a plausible answer, not a verifiable one |
| Confidence you can act on | Per-field numeric confidence, so you can set a threshold and route below it to review | Expressed verbally and poorly calibrated; hard to build an auto-approve rule on |
| Cost predictability | Priced per page. A 40-page statement costs the same today and next quarter | Priced per token. Cost scales with document length and prompt design, and moves when you change models |
| Reproducibility for audit | Pinned model version returns the same output for the same input | Output varies between runs and between model versions, which is difficult to defend in an audit |
| Degraded scans and handwriting | Purpose-built OCR handles low-dpi scans, skew and regional-language handwriting at scale | Good on clean documents, less reliable on the photographed-screen files real customers upload |
| Reasoning across documents | Not its job. Returns fields, not judgement | This is where it wins outright: reconciling GST against bank credits, explaining an anomaly, drafting the reviewer note |
| Unseen document formats | Needs a trained model or a prebuilt that fits | Handles a format it has never seen, which makes it the right tool for your long tail |
Sending a 40-page scanned statement through a language model to find eleven numbers is not clever engineering. It is paying reasoning prices for optical character recognition.
Security, residency and what the regulator now expects
Two frameworks govern this in India and they pull in the same direction:
RBI Digital Lending Directions, 2025
Issued on 8 May 2025. Data collection must be need-based, backed by prior explicit consent, and supported by an audit trail. Borrower data must be stored in India, and where it is processed outside India it must be repatriated and deleted from foreign servers within 24 hours. Digital lending apps must be reported on the RBI's CIMS portal with designated officer certification.
DPDP Rules, 2025
Notified on 13 November 2025. Consent Manager registration opens around November 2026 and full substantive compliance is due on 13 May 2027. The Data Protection Board is operational and complaint mechanisms are live. Penalties reach ₹250 crore for failure to maintain reasonable security safeguards per violation.
The 7 controls to insist on before you sign anything
- Residency you can evidence, not assert: Processing in an Indian region, with a deletion guarantee that satisfies the 24-hour repatriation clause if any step runs outside India.
- Redaction before inference, not after: Mask Aadhaar digits and strip identifiers you do not need at the ingestion boundary, so the model layer never sees them.
- No training on your data, contractually: A zero-retention configuration on every hosted model endpoint in the path, written into the contract rather than inferred from a marketing page.
- Private network path: Private endpoints or VPC-scoped access to every service. No document bytes traversing the public internet.
- Field-level audit trail: Every extracted value stored with its source page, coordinates, model version and confidence score, retained for the life of the loan.
- Consent and purpose binding: Each document processed against a recorded consent artefact with a stated purpose.
- Retention and deletion that actually runs: A scheduled job that deletes source files and derived artefacts on a defined clock, with evidence it executed.
The cost of usage, without marketing arithmetic
Per-page pricing is public. Here are the published pay-as-you-go rates for major cloud document services as reported through 2026:
| Model | Per 1,000 pages | What it gives you |
|---|---|---|
| Read (OCR) | $1.50 | Text, lines, words, handwriting. No structure, no fields |
| Layout | $10.00 | Tables, checkboxes, reading order. The substrate for downstream reasoning |
| Prebuilt models | $10.00 | Invoice, receipt, ID and similar, where your document fits the trained shape |
| Document classifier | $3.00 | Routing: deciding what kind of document arrived before extracting it |
| Custom extraction | $30.00 | Your specific form shapes. Training is free; you pay at inference |
| Query fields | $10.00 | Ad-hoc field requests without training a model |
| Add-ons | $6.00 | High resolution, barcode, formula. Each one added to the base rate |
Line items that do not appear on pricing pages
- Storage and egress for source files retained with audit trails
- Compute calling the API, queuing retries and handling partial failures
- Human review queue for exception handling
- Exception tooling with side-by-side highlighting
- Model maintenance as formats drift over time
- Integration engineering into your loan origination system
What it would cost, and return, in your operation
Move the inputs to match your own intake. The model is deliberately conservative: exception files are costed at the full manual handling time you spend today, and the platform rate is the published pay-as-you-go price rather than a committed-volume discount.
Your intake today
Modelled outcome
- Reviewer capacity needed today29.1 FTE
- Reviewer capacity needed after10.2 FTE
- Platform cost per year₹25.34 L
- Cost per document today₹80.00
- Cost per document after₹38.56
- Reviewer hours released per year34,944
- Year one net position₹81.46 L
What this model does not include: the conversion value of faster decisions, fraud losses avoided, and the ongoing model maintenance cost as document formats drift. The first two make the case stronger and the third makes it weaker.
Assumes 9,240 productive reviewer minutes per month, US dollar to rupee conversion at 88, and that exception files consume the same handling time as they do today.
How to train your team so you are not permanently dependent
The organisations that get value from this are the ones where the credit operations team ends up owning the pipeline, not the ones where a vendor owns it and the team files tickets.
The Reviewer (Exception Handler)
The job stops being "read the document and type the fields" and becomes "judge whether the machine got it right, and if not, why". Reviewers learn confidence score meanings, systematic vs one-off failure patterns, and how to create valuable correction data.
The Operations Lead
Straight-through rate is a lever they control by moving a confidence threshold. They balance threshold trade-offs between reviewer labor costs and error risk tolerance.
The Engineer (Pipeline Owner)
Owns document ingestion, classification, routing, retry logic, provenance storage, and drift monitors. Backend engineering teams pick this up in weeks with reference implementations.
Compliance Lead
Needs to demonstrate, on request, where a specific number in a specific credit decision came from. They learn to walk the audit trail without engineering intervention.
A 90-day enablement shape that works
Days 1 to 30: Build a golden set. Two hundred real documents per type, labelled by hand, disagreements resolved by your best reviewer. This encodes what correct means in your business.
Days 31 to 60: Run in shadow mode. Machine and human process every file in parallel; nobody acts on machine output. Score against golden set weekly.
Days 61 to 90: Phased live rollout. Go live on high-confidence band only with hard rollback. Widen the band weekly based on empirical evidence.
Where this is not the right answer
We would rather say this now than discover it in month four:
Under 2,000 documents / month
The engineering and governance overhead does not amortise. Fix the intake form and the checklist first.
Bespoke / Unstructured legal instruments
When every document is genuinely unique with no repeating structure, use a language model directly with a human reviewer.
Upstream collection bottleneck
If files arrive incomplete and half your handling time is chasing customers for missing pages, fix collection workflows first.
Unowned exception queue
A pipeline with an unowned exception queue silently creates backlogs and auto-approval risks.
The working session: one day, your documents, your numbers
Most of what is above is generic until it meets your actual files. We run a one-day working session that turns it specific.
Intake teardown
A 60-minute call on one workflow, to establish whether there is a case here at all.
- Walkthrough of one document workflow
- Rough per-page cost model
- Straight recommendation on next step
Document intelligence day
One day with your credit operations, engineering and compliance leads on your real files.
- Accuracy benchmark on up to 200 files
- Live scoring against your ground truth
- Cost model & RBI/DPDP control map
- 90-day rollout & enablement plan
Blueprint sprint
Three weeks, ending with one workflow running in shadow mode against live volume.
- Everything in working session
- Working extraction pipeline
- Provenance store with coordinates
- Side-by-side review interface
Risk reversal guarantee
If the working session ends without at least one workflow carrying a quantified business case built on your own documents, there is no fee. If you proceed to a delivery engagement with us within 90 days, the session fee is credited in full against it.
Questions we get asked
Does Account Aggregator make this redundant in two years?+
It reduces the document lane over time; it does not close it. Coverage gaps in cooperative banks, regional rural banks and smaller geographies persist, and the eighteen-to-twenty-four-month history that seasonal underwriting needs still arrives as a file. Build for both routes and the question stops mattering.
Can we run this entirely inside our own environment?+
Yes. Container-deployed extraction and self-hosted or India-region model endpoints are both viable. It costs more per page than the managed service and it removes the residency and retention questions entirely, which for some regulated entities is the right trade.
What accuracy should we expect on Indian bank statements?+
Anyone quoting a single number without seeing your files is guessing. Accuracy on a clean digital statement from a large private bank is not comparable to a photographed cooperative bank passbook page. This is why the working session benchmarks on your documents before anyone puts a figure in a proposal.
How long before it pays for itself?+
Use the calculator above with your own numbers. In most Indian lending intake operations we have scoped, the payback on the one-time build lands inside the first year at volumes above roughly 10,000 documents a month, and does not land at all below about 2,000.
Who owns the trained models and the labelled data?+
You do. The golden set is your intellectual property and it is the most valuable thing the programme produces. Any arrangement where a vendor retains it should be declined.
Sources and further reading
- Reserve Bank of India (Digital Lending) Directions, 2025, issued 8 May 2025
- Digital Personal Data Protection Rules, 2025, notified by MeitY on 13 November 2025
- Published Azure AI Document Intelligence pay-as-you-go rates as reported through 2026
- Industry commentary on Account Aggregator borrower coverage and PDF-based statement analysis, 2026
