Use caseDocument AIWorkflow automation
Mortgage applications where the documents chase themselves and the underwriter gets a checked file
A live document list per application, reminders on WhatsApp or email, and every payslip and bank statement read and cross-checked before an underwriter sees it.
A blueprint, not a client story. The business described is illustrative; the architecture, integrations and trade-offs are real, and this is how I would build it. By Ergini, .
The short version
A document collection workflow for mortgage brokers and small lenders. It keeps a checklist per application, sends reminders on WhatsApp or email with one upload link that needs no login, reads what arrives (payslips, bank statements, tax calculations, ID, property documents), checks it against the application and each other, and flags inconsistencies or signs of editing. The adviser gets an indexed, lender-ready file. It does not assess creditworthiness: the adviser recommends, and the lender's underwriter decides.
- Best for
- Mortgage brokers and SME lenders whose case managers spend their days chasing, renaming and checking borrower documents.
- Connects to
- Broker CRM or loan origination system (Acre, Mortgage Brain, Europace, Encompass), WhatsApp Business Platform, Gmail, Upload link, Open banking provider, Document store with retention rules
- The AI does
- Reads payslips, bank statements, tax documents and IDs in any format, and writes reminders and questions to borrowers in plain language.
- People do
- Advise the borrower, judge anything unusual in the income or accounts, choose the lender, and underwrite.
- Built as
- AI Workflow Automation, usually $15K - $30K
The case that is always one payslip short
Say you run a mortgage brokerage in the UK with ten advisers and four case managers. Each adviser carries thirty or forty live cases, and every one needs the same paperwork before a lender will look at it: proof of identity and address, recent payslips and a P60 for employed borrowers, SA302 tax calculations and tax year overviews for the self-employed, bank statements, proof of deposit, and later the property details.
The case managers spend most of their day chasing, and borrowers never send everything at once, even when there is a portal. An email with a list goes out; the borrower sends two of the seven items as phone photos on WhatsApp; a payslip arrives as a screenshot with the dates cropped off; the bank statement is page one of three. Each arrival is opened, renamed, checked, filed in the CRM and ticked off before the next reminder goes out. By the time the file is complete, the first payslip is too old for the lender the adviser chose, and the chase starts again.
Then the lender's underwriter finds what nobody had time to check: the employer on the payslip is not the one on the application, or the salary credits on the statement are lower than the payslips say. A case that looked ready comes back with conditions, and the borrower hears about it from a frustrated adviser. Ocrolus, which sells document automation to US lenders, reports that indexing documents inside Encompass saves about two hours per loan; that is its own figure, but it shows where lenders see the time going.
A German Baufinanzierung broker working through Europace has the same week with different paper: salary statements, the annual income tax certificate, tax assessments, and a long list of property documents such as the land register extract, floor plans and the living area calculation.
Between the fact-find and the lender
The model reads documents and writes messages. Checklists, windows, arithmetic and anything close to the lending decision are code or people.
01 Trigger · Acre, Europace or LOS API
The fact-find is complete
The adviser marks the case ready in the CRM. The workflow builds the checklist from the borrower's circumstances (employed, self-employed, gifted deposit, remortgage) and the target lender's requirements, which live in a table the case managers maintain.
02 System · WhatsApp Cloud API, Gmail API
Ask once, on the borrower's channel
One message with the full list and an upload link that needs no account or password, on WhatsApp or by email as the borrower chose. The link always shows what is still outstanding. Photos and PDFs sent back in the chat or by email are accepted and filed the same way, because refusing them only slows the case.
03 AI model · Vision model, structured output
Read what arrives
Each file is classified and extracted: employer, pay date, gross, net and year-to-date figures from payslips; holder, period, balances and transactions from statements; figures from tax calculations; name, number and expiry from IDs. Photos are straightened first, and a statement split across uploads is put back together.
04 Plain code
Check each document and the whole file
Statements must cover the lender's period with no gaps and balances that run on; year-to-date totals on consecutive payslips must add up; net pay must appear as a credit from the same employer; names and addresses must match the application; every document must be inside the lender's freshness window.
05 Plain code
Look for signs of editing
PDF metadata and structure, font and layout inconsistencies, and totals that do not reconcile are scored as signals, each with its evidence. Where the borrower connected open banking, statement figures are compared with the bank's own data. The deeper version of these checks is the document fraud detection build.
06 Decision
What happens with this document?
Rules decide; the model's reading is only an input.
- Complete, consistent and inside the window then ticked off, renamed to the firm's convention and filed in the CRM
- Missing pages, expired, or unreadable then a specific follow-up goes to the borrower
- A mismatch with the application, or an editing signal then stopped for the case manager with the evidence, and nothing is said to the borrower until a person decides
07 AI model
Write the follow-up
Short and specific: 'your March statement is missing page 2', not 'please resend your documents'. Reminders speed up as the case deadline approaches, switch channel if one goes unanswered, and stop the moment the checklist is complete. For the first weeks, each follow-up waits for a case manager's one-click approval.
08 Person
The adviser reviews the file
A one-page summary: what is in, what was checked, open questions, and for self-employed borrowers an income working labeled as a working, not a decision. The adviser decides what to raise with the borrower and which lender to submit to.
09 Result · CRM, lender portal
An indexed, lender-ready pack
Documents renamed and ordered the way the chosen lender wants them, uploaded to the CRM or the lender's portal with the check results. Every request, upload and check is logged, and a failed upload is retried under the same document ID, so nothing is filed twice.
Two payslips and a bank statement that do not quite agree
Most documents pass without anyone opening them. This is what a case manager sees when one does not.
Case M-1127, remortgage, day 6 of collection
Borrower · WhatsApp
Uploaded the payslips and the bank statement through the link. Is that everything now?
- extract(file: "payslip_june.pdf")employer matches application / paid 26 Jun 2026 / gross 4,250.00 / net 3,212.40 / year-to-date gross 12,750.00
- extract(file: "payslip_july.pdf")employer matches / paid 28 Jul 2026 / gross 4,250.00 / net 3,212.40 / year-to-date gross 17,500.00
- extract(file: "statement_jun_jul.pdf")holder matches / 1 Jun to 31 Jul 2026 / 4 of 4 pages, balances run on / salary credits: 26 Jun 3,212.40, 28 Jul 2,987.40
- check_file(case: "M-1127")year-to-date: July should show 17,000.00, payslip shows 17,500.00 / July net pay 3,212.40 but bank credit 2,987.40, difference 225.00 / July payslip produced by different PDF software than June
Collection workflow · WhatsApp
Thanks, your payslips and bank statement have arrived and we're checking them now. The only thing still outstanding is your latest mortgage statement, which you can add through the same link.
- The borrower is not told about the mismatch. The case manager gets the three findings with the documents side by side.
Case manager · CRM
Asked the borrower to download the July payslip directly from the employer's payroll portal. Holding the case until it arrives.
- It may be a payroll correction or an edited document. The workflow does not decide which; it makes sure a person looks before a lender does.
What each borrower document is checked for
Lists vary by borrower and lender. These are the common items and what the workflow verifies for each; case managers add whatever a lender asks for.
| Document | What is extracted | What is checked |
|---|---|---|
| Payslips | Employer, pay date, gross, deductions, net, year-to-date, tax code | Dates inside the lender's window; year-to-date arithmetic across consecutive slips; net pay matches a bank credit |
| Bank statements | Holder, account, period, opening and closing balances, transactions | Full period with no gaps; balances run on; name and address match the application; large unexplained credits listed for the adviser |
| SA302 and tax year overview (UK), tax assessment (Germany) | Tax year, total income, trading profit, tax due | The overview matches the calculation for the same year; figures match the application |
| Photo ID and proof of address | Name, date of birth, document number, expiry, address | Not expired at expected completion; names consistent across every document |
| Proof of deposit | Source account, balance, gift letter details | The funds appear on a statement in the borrower's name, or a gift letter is present with the donor's statement |
| Property documents (Germany: land register extract, floor plans, living area calculation) | Address, owner, encumbrances, areas | The property matches the application; the extract's date is inside the lender's window |
The credit decision stays out of the software
The EU AI Act lists evaluating the creditworthiness of individuals as a high-risk use, with those obligations now due from 2 December 2027. This workflow stays outside that by design: it collects and checks documents and never scores a borrower.
The AI model
Read every upload, including photos and screenshots
Borrowers send documents in every format there is.
Write reminders and follow-ups
Specific requests get answered; vague ones get ignored.
Plain code
Build the checklist from the case and the lender's list
Lender requirements are data in a table, not model knowledge.
Check dates, arithmetic, completeness and consistency
The same checks, the same way, on every file.
Score editing signals and compare with open banking data
Evidence for a person, never a verdict.
Prepare a self-employed income working
Arithmetic with every figure linked to its source document.
A person
Decide what a mismatch or signal means
It may be a payroll error or an edit; only a conversation tells.
Recommend the product and the lender
Regulated advice, owned by the adviser.
Assess affordability and creditworthiness
The lender's underwriter decides; the workflow produces no score.
Where document chasing breaks down
Self-employed income
Two years of tax calculations, overviews and accounts, profits that move, and lenders that average them differently. The workflow extracts the figures and prepares a working with every number linked to its source, clearly labeled as a working. It never presents a qualifying income; that is the lender's calculation.
Documents in another language or currency
A borrower paid in euros, a statement from a Polish bank, a German tax assessment on a UK application. The model reads them well, but lenders may want certified translations and their own exchange rate. The checklist adds a translation requirement for lenders that ask for one, and figures are converted in code at a rate the adviser can see.
Documents that expire while the case waits
Chains stall and payslips age. Each document carries its expiry for the chosen lender, and the workflow re-requests before the file goes stale rather than after an underwriter notices. When the adviser switches lender, the list and the windows are recomputed instead of assumed.
Borrowers who stop answering
A chase that repeats the same generic list teaches people to ignore it. Reminders name the one or two items left, switch channel after an unanswered attempt, and hand over to the adviser after a set number, because a silent borrower may have a reason worth a phone call. Plain, specific messages are also what the FCA's Consumer Duty expects.
A document that talks to the model
An applicant can hide text in a PDF telling an AI reviewer that the figures are verified. Extraction runs with no tools that act, output goes through a strict schema, and every check that matters is code, so hidden instructions have nothing to change. The prompt injection defenses write-up covers the patterns.
Financial documents under GDPR
Payslips and statements are personal data with obvious misuse potential. They stay in EU or UK hosting, access follows the case, borrowers are told how their documents are checked, and files are deleted on the schedule the firm's record-keeping duties allow. Attachments sent over WhatsApp pass through Meta's Cloud API, which belongs in the DPIA; the GDPR-compliant AI development page covers that paperwork.
Does your CRM's client portal already do enough?
Check your own systems first. In the US, Ocrolus, Blend and Floify cover document collection and analysis around Encompass and other origination systems, and they are mature products. In the UK, broker CRMs such as Acre and the tools around Mortgage Brain, and in Germany the Europace ecosystem, handle cases and lender submissions; see what their client-facing document features already cover. If borrowers already upload through a portal and your case managers mainly need a checklist, switch on what you have.
A custom build earns its place when the chasing itself is the problem: borrowers who live on WhatsApp and will not log into another portal, documents in five formats, checks your portal does not run (year-to-date arithmetic, bank credit matching, editing signals), and lender windows and lists that keep changing. It also fits SME lenders whose origination system stores documents but reads none of them.
The usual shape is a layer beside the CRM: the CRM stays the case record, the workflow owns the checklist, the chase and the checks, and the indexed file lands back in the CRM for the adviser to submit. Nothing about the lending decision moves into it, which keeps the EU AI Act analysis short.
How you would know it is working
A blueprint has no results to report, so here is what I would measure from the first week instead, on your own data.
- Days from fact-find to complete file
- Per case and per borrower type, compared with the months before, split into time waiting on the borrower and time waiting on the firm.
- Case manager touches per case
- Documents a person had to open, rename or check by hand, which should fall as the checks earn trust.
- Lender conditions after submission
- Requests from underwriters for documents or explanations, tagged by whether the workflow had already flagged the issue.
- Signals and their outcomes
- Every flagged mismatch or editing signal with what the case manager found, so false positives are tuned down and true ones are counted.
- Reminder response time by channel
- How fast borrowers respond on WhatsApp versus email, used to set the default channel and cadence.
What a build like this costs
This is built as AI Workflow Automation, which runs $3.5K - $60K overall. A build like this one usually lands in the multi-step workflow tier: $15K - $30K, 3-5 weeks. The first working version runs on your real data well before the end of that window.
What it costs to run
Model costs run from a few cents to tens of cents per case, depending on how many pages borrowers send. WhatsApp templates outside the 24-hour window are billed per message by Meta, open banking providers charge per connection or per month, and document storage is a small line.
What moves the price
- Which CRM or origination system, and how open its API is for reading cases and writing documents
- How many channels: email only, or WhatsApp with templates and opt-in handling
- How many lenders' document lists and freshness windows need to be maintained
- Whether open banking is part of the flow, and through which provider
- The depth of the editing checks, from arithmetic and metadata to a full forensic review
Who this is for
- UK mortgage brokerages whose case managers chase documents on WhatsApp and email all day
- German Baufinanzierung brokers on Europace collecting income and property documents from borrowers
- SME and specialist lenders whose origination system stores documents but reads none of them
- Firms that have had a lender find an edited payslip before they did
Questions people ask about this
Can AI check payslips and bank statements for a mortgage application?
Yes, for the mechanical checks: dates, completeness, year-to-date arithmetic, names and addresses, and whether net pay shows up as a credit on the statement. The model reads the documents and code runs the checks. What it should not do is judge affordability or creditworthiness, which is the lender's decision and, under the EU AI Act, a high-risk use.
How do I automate mortgage document collection?
Build the checklist from the borrower's circumstances and the lender's requirements, send one request with an upload link on the channel the borrower prefers, and read and check each document as it arrives. Follow-ups name exactly what is missing and stop when the list is complete. The adviser gets an indexed file with the check results and decides what to raise with the borrower.
Is this a high-risk AI system under the EU AI Act?
Not as designed. The Act lists creditworthiness evaluation and credit scoring of individuals as high-risk, and the Digital Omnibus moved those obligations to 2 December 2027. This workflow collects and checks documents and produces no score or recommendation about the borrower. If you later want it to recommend lending decisions, the analysis changes, and so does the design.
Can it detect edited or fake documents?
It raises signals, not verdicts: arithmetic that does not reconcile, metadata showing a different producing application, layout and font inconsistencies, and figures that disagree with open banking data. Each signal arrives with its evidence and goes to a case manager. For lenders who need a deeper forensic layer, the document fraud detection blueprint covers it.
Do borrowers have to log into another portal?
No. Each application gets one upload link that needs no account or password, and photos or PDFs sent back on WhatsApp or by email are filed the same way. Borrowers opt in to WhatsApp first, and reminders sent more than 24 hours after their last message use templates Meta has approved. If borrowers can chat with an automated assistant, it says so up front, as Article 50 of the EU AI Act requires for EU borrowers.
How much work is it to set up?
A first version with one CRM, email and WhatsApp reminders, the core checklist and the checks above usually fits the multi-step tier of AI workflow automation shown on this page. After launch, keeping lender lists and windows current is ongoing work for the case managers, which the build turns into a table edit rather than a code change.
Sources