ContactBook a call

Industries/ Finance/ Document reading and triage

Bank statements and payslips, read on arrival, checked against the file

We build the layer that reads what applicants and clients send, pulls out the figures your team needs and checks them against what was declared. Clear documents move on with the source attached, and anything ambiguous or above a threshold you set goes to a person.

An operations analyst, a young man with curly hair in a navy jumper, in a bright back office comparing a scanned statement on one monitor with extracted figures on a second, a few envelopes on the desk.
Used by
Case handlers, underwriters, paraplanners and operations
Takes the place of
Opening each PDF and retyping figures into the system
Connects to
Your core system, CRM, document store and shared inboxes
Built with
Our document intelligence service

Where it hurts today

A normal week, before it exists

A single mortgage or business loan case can bring in a dozen documents: three months of bank statements, payslips, an SA302 tax calculation in the UK or W-2s in the US, proof of address, ID, a valuation and a signed declaration. They arrive as email attachments, phone photographs and scans, often bundled into one PDF. Someone opens each one, works out what it is, renames it and saves it to the case before any checking starts.

Then the checking. A case handler reads the payslip and types the gross and net pay into the system, then reads the statements to see whether the salary credits match, whether any direct debits came back unpaid, and whether a regular payment suggests a commitment nobody declared. It is careful, skilled work, and much of it is reading numbers off one page and typing them into another.

At an advice firm the pile looks different but behaves the same way: letters of authority, provider valuations, transfer forms and annual statements, each one opened, checked, renamed and filed by a paraplanner. When volumes rise, the queue grows, applicants wait, and the few documents with something genuinely wrong get the same attention as the dozens that are perfectly fine.

What it does

Document reading and triage, feature by feature

Reading the page is the easy part. What makes it safe in a lender or an advice firm is what happens after.

  1. 01

    Sorted and split on arrival

    A combined PDF from a broker is split into its documents, each classified as a bank statement, payslip, ID, valuation or form, and filed to the right case with a clear name. Pages that are upside down, cropped or unreadable are flagged rather than guessed at.

  2. 02

    Bank statement and payslip extraction

    Employer, pay date, gross and net pay, account holder, statement period, opening and closing balances: each field is extracted with its own confidence score. A certain employer name and an uncertain net pay figure are treated differently, so a person looks only at the field in doubt.

  3. 03

    Checked against what was declared

    Extracted figures are compared with the application or client record. Does the name match, is the statement within the period your policy requires, do the salary credits agree with the payslip, has this document appeared on another case? Each mismatch is listed with the page it came from.

  4. 04

    Thresholds you set

    Your credit or operations lead decides what goes straight through and what goes to a person: values above a limit, income that differs from the declared figure, statements older than your policy allows. The rules are written down, changed only by named people, and every change is recorded.

  5. 05

    A review screen built for checking

    Documents in doubt open beside their extracted fields, each highlighted where it was found on the page. The reviewer confirms or corrects in seconds and never retypes the document. Corrections are captured, so the fields your team keeps fixing are the ones that improve.

  6. 06

    Statement notes, not decisions

    For bank statements, the system can list what an underwriter usually looks for, such as returned payments, regular transfers to another lender or gambling transactions, each linked to the line on the statement. It lists them. The underwriter decides what they mean.

A day with it

How the day runs instead

A day in a lender's case team with documents read as they arrive.

  1. 7.30am
    Broker

    A broker uploads a case pack for a self-employed applicant: two years of tax calculations, six months of business statements and ID, combined in one file. By the time the case team logs on, it has been split, classified and filed to the case.

  2. 9.00am
    Case handler

    The case handler opens the review queue. Most overnight documents passed every check. Four need a person: a payslip whose net pay is below the declared figure, a statement outside the date range, and two photographs too blurred to read.

  3. 9.20am
    Applicant

    The applicant whose photograph was blurred gets a message asking for a clearer copy, with a note on what went wrong. He retakes it on his phone from the same link, and it is read and checked again straight away.

  4. 11.00am
    Underwriter

    The underwriter opens a case that is already populated: income extracted and matched, statement notes linked to each line, and the valuation flagged because it is above the review threshold. She checks the source pages and makes the decision in the core system.

  5. Month end
    Operations manager

    The operations manager reviews the accuracy report: which fields were corrected most often, and which document types went to review. Payslips from one payroll provider keep being corrected, so that layout is added to the test set.

What it connects to

Works with what you already run

Documents are collected where they already arrive, and the results go into the systems your team works in.

  • Shared inboxes and upload links

    Documents are collected from the case mailbox, a broker portal or your client portal's upload step, so applicants and brokers keep sending them the way they do today.

  • Core lending or back-office system

    Extracted figures, check results and the filed documents are written to the case or client record through its interface, with each value linked to the page it came from.

  • Document store

    Originals are archived in your document management system or a cloud store such as SharePoint, named consistently and linked to the record, with retention periods applied as your policy sets them.

  • AI document service

    Reading is done by a document model or a specialist extraction service, for example from Microsoft Azure or AWS, chosen per document type and run under terms and in a region agreed with you.

The first version

Start small, then grow it

A first version should be useful on day one and small enough to get right. This is where we would usually draw the line, and what tends to follow once people are using it.

In the first version

  • One document type, usually payslips or bank statements, read from your real cases
  • Splitting, classification and filing for combined PDFs from brokers and clients
  • Checks against the declared figures in the application or client record
  • A review queue showing each extracted field beside the source page
  • Everything routed to a person at launch, with automation raised field by field

Once it is in use

  • Further types such as tax calculations, valuations and letters of authority
  • Statement notes for underwriters, linked to individual transactions
  • Instant requests for a new copy when a photograph cannot be read
  • Monthly accuracy reporting per field and per document type

Buyer's guide

Before you commission one

Four decisions worth making before you automate document handling.

  1. 01

    A specialist product, or a build

    For high-volume, well-understood documents such as bank statements, specialist income verification and bank statement analysis products exist, and one may suit you better than a build. A custom pipeline makes sense when your document mix is broad, when the checks depend on your own case data, or when results must post into a core system those products do not reach.

  2. 02

    What to gather first

    A hundred real, anonymised documents of the type you would start with, including the bad scans and the awkward payroll layouts. Add the checks a case handler makes on each one, the thresholds your credit policy sets, and how long documents must be kept. That set becomes the test the system is scored against.

  3. 03

    Who should own it

    Operations owns the queue and the day-to-day thresholds. Credit or underwriting owns what counts as a mismatch and what must go to a person. Your data protection lead signs off what is extracted and kept. Name all three people before building, because a pipeline nobody owns drifts quietly as documents change.

  4. 04

    How you will know it is working

    Watch the time from a document arriving to a case ready for underwriting, how many documents go to review and why, which fields reviewers correct most often, and how many cases bounce back for a missing or unreadable document. Accuracy should be re-scored on your test set after every change, never assumed.

Terms worth knowing

The words you will meet

Intelligent document processing
Software that reads documents the way a person would: it classifies each one, finds the fields that matter and extracts them as data. Modern versions cope with layouts they have not seen before, unlike older template-based OCR that broke whenever a payroll provider changed its payslip design.
Confidence score
A value the system attaches to each extracted field, showing how sure it is. It lets routine values move on while uncertain ones go to a person. The thresholds are a business decision: an employer's name and a net pay figure deserve very different levels of caution.
Human in the loop
A design where a person checks or confirms the system's output before it counts. In lending it also has a legal side: in the UK, rules in force since 5 February 2026 give people the right to human intervention in significant decisions made solely by automated means, and the EU GDPR has similar protections.

Handled with care

The details that matter here

Documents in this sector carry identity, income and account details. These are the things we handle with your credit, operations and data protection leads.

  • A person decides

    The system reads, checks and lists. It does not approve or decline applications or judge affordability. That matters beyond good practice: in the EU, the AI Act treats AI that evaluates people's creditworthiness as high-risk, with those obligations now due from 2 December 2027. We design the routing so decisions land with a named person.

  • Scored on your own documents

    Before anything runs live, we test it on a set of your real, anonymised documents with known correct answers and report the results field by field. You decide which fields may pass without review, and you can send any field back to manual checking at any time.

  • Sensitive data handled narrowly

    Bank statements show far more than income: health payments, family matters, religious giving. We agree with your data protection lead what is extracted and kept, limit the full documents to the roles that need them, and apply retention periods rather than storing everything indefinitely.

  • A trail for every figure

    Every value in your system keeps a link to the page it came from, the version of the pipeline that read it, its confidence and who confirmed or corrected it. When a decision is reviewed later, the trail shows how each figure was arrived at.

Guides

Worth reading first

Questions

What buyers ask first

Can AI document processing make lending decisions?

Not in what we build. It reads documents, checks them against what was declared and lists what an underwriter would want to see. The credit decision stays with your people and your existing decision system, with the track record your auditors and funders rely on. If you already run an automated decision engine, the extracted data can feed it, but the rules and the responsibility stay where they are today.

How accurate is AI document processing on bank statements and payslips?

It depends on your documents more than on the model. A clean PDF statement downloaded from a bank reads very well. A phone photograph of a creased payslip under a kitchen light does not. That is why we score it on your real documents before proposing anything, and why each field carries its own confidence score, so uncertain values reach a person rather than your system.

Do brokers and applicants have to send documents differently?

No. Coping with what already arrives is most of the point. That said, if open banking can replace uploaded statements for some customers, or a broker can send structured data instead of a PDF, that is more reliable than reading a document, and we will point it out rather than build extraction you do not need.

Can it detect fraudulent documents?

It can flag things worth a closer look, such as figures that do not add up, a statement that disagrees with the payslip, or a document already seen on another case. It is not a fraud detection service, and we would not present it as one. Where your firm uses a specialist fraud provider, documents can be passed to it at the step your fraud lead chooses.

Where are our customers' documents processed and stored?

That is decided with you before anything is built. Options range from a cloud document service under terms your data protection lead has reviewed, through processing pinned to a particular region, to models running inside your own cloud account. You get a written data flow, and retention follows the periods your anti-money laundering and lending record rules require.

Send us the documents your team retypes every day

Share a handful of anonymised examples, including the awkward ones. We reply within 48 hours and will tell you which are worth automating first.

More for finance

What else we build for finance and banking