Reducto
Prospect brief · August 2026

Prepared for Klarna

3.4 million transactions a day. The hardest decisions still arrive as documents.3

Open banking covers the clean cases. The long tail behind affordability checks, KYC, merchant onboarding, and disputes arrives as bank statements, payslips, scanned IDs, and registry filings in hundreds of formats. Reducto turns that document long tail into structured data your models and agents can rely on, with every extracted value traceable to the exact place on the page it came from.

First: what Reducto is

Reducto turns complex documents into structured data your systems can rely on. Parse, extract, split, and classify run as APIs over anything the company receives: PDFs, scans, images, spreadsheets, more than thirty formats. Every extracted value carries a citation to its exact place on the source page and a confidence score, so a person or an agent can verify a number before acting on it.

Backed by a16z · In production at Citadel, Bridgewater Associates, Bretton, and Basis5

Current Account Statement

AFFORDABILITY PACKET · SCANNED PDF · P.2 OF 6

detected · table
DateDescriptionAmountBalance
01 JulSALARY · ACME AB+34,20051,384
03 JulRENT · BOSTAD STHLM-11,50039,884
05 JulLOAN PMT · NORDBANK-2,15037,734

This statement is provided for information purposes only. The bank accepts no liability for decisions made on the basis of this document. Figures may differ from cleared balances. For questions regarding this statement please contact your branch...

classified · boilerplate · filtered out

structured output

doc_type····· bank_statementclassified ✓
holder······· J. Anderssonconf 0.997
net_income_mo 34,200 SEKconf 0.998
rent_mo······ 11,500 SEKconf 0.996
debt_svc_mo·· 2,150 SEKconf 0.995
source······· p.2 · rows 1-3bounding box
Illustrative; names and figures are fictional. Every extraction ships with field-level confidence and a citation back to the source page.
99.6%Precision and recall on the only independent long-document extraction benchmark1
225/225Only system to complete every document: avg. 358 pages, ~88,700 fields each1
5B+Pages processed in production, 1B+ a month2
1.07MMerchants on Klarna, up 49% in a year: a swelling onboarding document load3

Capabilities

Built for the documents a lender actually receives

Bank statements

Hundreds of formats, one schema

Transaction tables extracted across bank and country layouts, degraded scans and rotated pages included, into the normalized income and expense record your risk models expect.

Identity documents

Scanned IDs, routed and read

Passports, national IDs, and proof-of-address documents classified by type and country, then parsed, so the long tail your IDV vendor rejects stops falling to manual review.

Merchant KYB packets

One upload, many documents

Registration certificates, ownership declarations, and company financials arrive as a single mixed file. Split and classify separate them; extraction structures each one.

Dispute evidence

Mixed media into case files

Order confirmations, tracking screenshots, photographed receipts, and correspondence threads parsed into one structured record an adjudicator or a triage agent can act on.

Provenance

Every value carries its source

Bounding-box citations point each extracted field back to its page and cell, with field-level confidence. For decisions a regulator can ask about, the input is auditable.

Deployment

Deployed to a bank's bar

SaaS, hybrid VPC, full VPC, or air-gapped on-prem. SOC 2 Type II, encryption in transit and at rest, zero data retention with 24-hour deletion, pinned model versions.2

MENU

Where this lands across Klarna

Nothing here has been tested with Klarna yet: every item is an offer, not a claim. Each one names the workflow it serves and the teams closest to it, because a document going wrong is never abstract. Someone's credit decision, onboarding queue, or dispute case feels it.

I

Where we fit first

The workflows where document accuracy touches the credit-loss line most directly.

Affordability and income documentsStrong fit

Bank statements, payslips, and income proof behind manual-review and Fair Financing decisions, extracted into a normalized affordability record. Klarna provisioned $794M for credit losses in FY20253; the document long tail feeding those decisions is where silent misreads compound.

for credit risk · underwriting · in production at Basis, on bank statements in hundreds of formats5

owns Pedro Camargo, Global Head of Credit Risk Modeling · cares the CRO organization

KYC identity documentsStrong fit

Passports, national IDs, and proof of address at bank-grade onboarding scale: 15.8M banking consumers and growing3. Classification routes each document by type and country; parsing handles the scans that would otherwise queue for a human.

for onboarding · financial crime prevention · in production at Bretton, on KYC and AML documents for regulated banks and fintechs5

owns the compliance organization · cares Sebnem Erener, Head of AI Legal, Regulatory & Ethics

Merchant KYB packetsStrong fit

1.07M merchants, up 49% in a year3, each onboarded on registration documents, ownership declarations, financials, and screening evidence across 26 countries' formats. One heterogeneous upload, split, classified, and extracted per component.

for merchant onboarding · AML · in production at Bretton5

owns merchant onboarding and financial crime teams

II

Next in line

Same platform, same accuracy bar, ready when a team is.

Dispute and chargeback evidenceAvailable

Klarna's own evidence guidelines ask merchants for order confirmations, delivery proof, receipts, and correspondence4. Those packets arrive as PDFs, screenshots, and photos; parsed and structured, they feed adjudication instead of a manual reading queue.

for disputes · recoveries · merchant experience

owns disputes and fraud operations

AML, sanctions, and SAR filesAvailable

Screening hits, enhanced-due-diligence files, and narrative documents parsed with citation grounding, so every statement in a filing traces to its source. Provenance that survives a regulator's review.

for financial crime prevention · compliance · in production at Bretton5

owns the compliance organization

Collections and hardship correspondenceAvailable

Free-text letters, hardship documentation, and attachments across a $10.5B receivables book3, classified and extracted so recovery workflows start from structured facts instead of an unread inbox.

for collections · customer operations

owns collections and recoveries teams

III

The platform layer

For the AI platform and engineering teams: the layer everything above runs on.

The accuracy layer under every agent Klarna buildsPlatform

Klarna has shown it builds its own AI3. Reducto is not another agent: it is parse, extract, split, and classify as APIs your agents call, plus a native MCP server for agent frameworks. Your models and your underwriting logic stay yours; Reducto is the clean, cited input underneath them, never the model itself.

for AI platform · agent engineering · ML infrastructure · 99.6% precision/recall on the independent LongExtractBench (micro1, Jul 2026)1

owns Kate Canney, Accountable Lead, Automations & Agent Channels · cares Kevin Lindemann, Director of Product, AI

Deployed to your barPlatform

SaaS, hybrid VPC, full VPC, or air-gapped on-prem. SOC 2 Type II, encryption in transit and at rest, zero data retention with 24-hour deletion, and pinned model versions so results reproduce exactly.2

for infrastructure · security · compliance

owns the CTO organization

How to order

Put your documents through the test.

Bring the ten documents your team trusts least: the statements that misparse, the IDs that fall to manual review, the merchant packets that take days. Reducto returns them structured, cited, and scored, and you judge the output against your own numbers.

Reducto

Malia Lake

Business Development

malia@reducto.ai