Fourteen Documents for One Mortgage, and Three Still Missing
By Ergini, Software & AI Developer
A composite story. The company and the people in it are invented. The problem, the rules and the system are real, and the full blueprint is in the use case library.
TL;DR
A composite story: at an invented mortgage brokerage in Brighton, a first-time buyers' file sits at eleven of fourteen documents on day 23 while the estate agent chases the adviser. Two of the three missing items were sent and could not be used. The workflow I would build keeps a checklist per case from the lender's own requirements, asks once on the borrower's channel with one upload link, reads every photo and PDF, checks dates, arithmetic and names in code, and names exactly what is still missing. Case managers approve the follow-ups, and the lender decides the loan.
A payslip under a mug
The photo shows a payslip on a kitchen table, half under a mug of tea. Kerry zooms in. The employer's name is there, and the gross and the net pay. The pay date is under the mug.
Kerry is one of three case managers at a mortgage brokerage in Brighton, and the photo is the latest arrival on case 2231: Chloe and Ravi, first-time buyers, a two-bedroom flat and a deposit gifted by Ravi's parents. Chloe sent it to the work phone on WhatsApp with one line: "Is this the one you needed?"
Nearly. The case needs fourteen documents before Gemma, their adviser, can submit it to the lender she has chosen, and Kerry's tracker says eleven are in. It is day 23 since the fact-find, and the three missing items have been missing for most of that time: Chloe's June payslip, which this photo almost is; Ravi's tax year overview for 2024 to 2025; and a statement from his parents' account showing the gift leaving it.
Gemma, meanwhile, has an email from the estate agent. The vendor wants to exchange by the end of the month and would like to know whether the buyers' mortgage offer is "on its way". It is not on its way. It is waiting for a mug to move.
Eleven of fourteen, day 23
The brokerage is invented, and so are Kerry, Gemma, Chloe and Ravi. The week is the one most brokers will recognize: eight advisers with thirty or forty live cases each, three case managers, and every case waiting on the same kinds of paper before a lender will look at it. Photo ID and proof of address. Recent payslips and a P60 for the employed. SA302 tax calculations and tax year overviews for the self-employed. Bank statements. Proof of deposit.
Borrowers never send it all at once, even when there is a portal. An email with the list goes out, and two items come back as photos on WhatsApp; a bank statement arrives as page one of three; a payslip is a screenshot with the dates cropped off. Each arrival is opened, renamed, checked, filed in Acre, the brokerage's CRM, and ticked off in Kerry's tracker, a spreadsheet beside it, before the next reminder goes out. The reminder is the same list again, because writing a specific one for every case on her list is not something a person can do every morning.
And the clock runs the wrong way. By the time the last document arrives, the first payslip can be older than the lender's window allows, and the chase starts again for a document that was already in. When a file does reach the underwriter, what nobody had time to check comes back as a condition: the salary credit on the statement is not the net pay on the payslip, or the deposit is not where the application says it is.
Never sent, unusable, or expired
Nadia, who owns the brokerage, gets in touch after a month in which three purchases were delayed for the same reason. On our first video call I ask Kerry to share her tracker, and I ask one question about every document marked outstanding: was it never sent, sent but unusable, or sent and then expired while the rest of the file caught up?
Case 2231 answers it well. Chloe has sent her June payslip twice, once with the mug and once as a screenshot with the dates cropped off. Ravi has sent his SA302 for 2024 to 2025 twice, because he thinks it is the tax year overview; nobody has told him they are two different documents. Only the parents' statement has truly never arrived, and that is not a reminder problem. His parents do not want to send their bank statements to people they have never met.
That settles what the build is for. The chase is not failing because borrowers are careless. It fails because it says "please send your outstanding documents" to people who believe they already have, and because nobody checks an upload until someone has time to open it.
There are things it will not do, and I name them on that first call. It will not assess affordability or creditworthiness, or score anyone: the adviser recommends and the lender's underwriter decides. It will not tell a borrower about a mismatch in their documents. And it will not keep reminding someone who has gone quiet. After a set number of reminders the case goes to the adviser, because a silent borrower may have a reason worth a phone call.
One list, one link, the borrower's channel
The build starts with a table the case managers will own. Every lender the advisers submit to gets its document requirements as rows: what it wants, how recent each item must be, in what order it wants the pack. They edit it when a lender changes its mind. When the adviser marks the fact-find complete in Acre, the case's checklist is built from the borrowers' circumstances (employed, self-employed, gifted deposit) and the chosen lender's rows.
The ask then goes out once, with the whole list, on the channel each borrower chose, with one upload link that needs no account or password and always shows what is still outstanding. Photos sent back in the chat and PDFs sent by email are filed the same way, because refusing them only slows the case. WhatsApp needs the borrower's opt-in, and anything sent more than 24 hours after their last message goes as a template Meta has approved.
A vision model reads whatever arrives, straightening photos first and putting a statement split across uploads back together. Everything after the reading is code. 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, and 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 window, and IDs must still be valid at the expected completion.
Case 2231, as the checklist sees it
Here are the fourteen documents on the day the photo with the mug arrives, as the checklist would hold them:
| Documents | What the checks found | Next |
|---|---|---|
| Photo ID for both, proof of address | Names and address match the application; both IDs valid past the expected completion | Filed |
| Chloe's April and May payslips, P60 | Year-to-date adds up; net pay matches the salary credits on the statement | Filed |
| Chloe's June payslip | Two uploads, neither shows a pay date | Follow-up |
| Ravi's SA302s for both years, tax year overview for 2023 to 2024 | The overview matches the calculation for the same year; figures match the application | Filed |
| Ravi's tax year overview for 2024 to 2025 | Not received: that year's SA302 arrived twice instead | Follow-up |
| Joint bank statements, three months | Two uploads put back together; balances run on | Filed |
| Gift letter | Present, signed by both donors | Filed |
| Donors' statement showing the gift | Not received after two reminders | To Gemma |
The model writes the follow-up, and in the first weeks Kerry approves each one with a click before it goes. This one reaches Chloe on WhatsApp the same morning:
Thanks, Chloe. The pay date on this payslip is hidden in the photo, so we can't use it yet. Could you send your June payslip again with the whole page showing? And one for Ravi: we have his tax calculation for 2024 to 2025 twice, but we also need his tax year overview for that year. It's a separate page he can print from his HMRC online account, under Self Assessment. Same link as before, and that's everything we need from you two.
Plain, specific messages are also what the FCA's Consumer Duty expects. The donors' statement gets no third reminder: it goes to Gemma with the history, and she calls Ravi's mother to explain why the lender needs to see the money leave her account, and who will see the statement. The whole route, from fact-find to a lender-ready pack, is in the blueprint for loan document collection.
Last month's cases, run again
Before a single message reaches a borrower, the checks run on files the brokerage has already closed. Last month's completed cases go through the reading and the rules, and every flag is compared with what actually happened: what Kerry caught, and what underwriters came back with as conditions. Flags that were noise get tuned down, and every condition the checks should have caught becomes a test case.
The replay also catches something I got wrong. In one case Gemma switched lender halfway through, after the first lender would not lend on the property. My first version kept the checklist it had built for the first lender, so a payslip that was recent enough for that lender stayed ticked, although the second lender's window was shorter. The underwriter had asked for a newer one. Now, when the adviser changes the lender in Acre, the list and every window are recomputed from the new lender's rows, and anything that no longer qualifies goes back on the chase.
Kerry approves, Gemma decides
- Follow-ups wait for Kerry at first. Once she trusts them, only the unusual ones do. Reminders speed up as the case deadline nears, switch channel when one goes unanswered, and stop the moment the checklist is complete.
- Mismatches stop for a person. A salary credit that is not the net pay, a name that differs, a sign that a PDF was edited: each goes to the case manager with the evidence side by side, and the borrower hears nothing until someone decides what it means. The deeper checks are the document fraud detection build.
- The adviser gets a summary, not a pile. One page: what is in, what was checked, open questions, and for Ravi an income working labeled as a working, with every figure linked to its source. The lender and what to raise with the borrowers are Gemma's calls; affordability is the underwriter's.
- Documents cannot give orders. A PDF can hide text telling an AI reviewer the figures are verified, so extraction runs with no tools that act and a strict schema, and every check that matters is code. The prompt injection write-up has the patterns.
Payslips and statements stay in UK or EU hosting, access follows the case, and files are deleted on the schedule the firm's record-keeping duties allow. Attachments sent on WhatsApp pass through Meta's Cloud API, which belongs in the DPIA.
Fourteen of fourteen
Kerry's tracker is retired; each case's checklist is the tracker. Her mornings start with the exceptions: uploads that failed a check, follow-ups waiting for approval, cases handed to an adviser. A photo with a mug in it still arrives now and then, and the reply asking for the whole page goes out that morning instead of whenever she reaches it. When case 2231 reaches fourteen of fourteen, the pack goes to the lender renamed and ordered the way that lender wants it, with the checks attached, and Gemma can tell the estate agent the case is with the lender, complete.
Check your own systems before building any of this. Acre and the tools around Mortgage Brain in the UK, and the Europace ecosystem in Germany, handle cases and lender submissions, and in the US Ocrolus, Blend and Floify are mature products for document collection. If borrowers already upload through a portal and your case managers mainly need a checklist, switch on what you have.
A 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, and lender lists that keep changing. It sits beside the CRM rather than replacing it, and usually fits the multi-step tier of AI workflow automation. The full blueprint has the flow, the checks per document and the ways document chasing breaks down.
Frequently asked questions
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 why, 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.
Can AI check payslips and bank statements for a mortgage application?
Yes, for the mechanical checks: dates inside the lender's window, 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. It should not judge affordability or creditworthiness, which is the lender's decision and, under the EU AI Act, a high-risk use.
Do borrowers need to log into a portal to send mortgage documents?
No. Each application gets one upload link that needs no account or password and always shows what is still outstanding, 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.
Should a mortgage broker use the CRM's client portal or build document collection?
Check what you already pay for first: broker CRMs such as Acre, the Europace ecosystem in Germany, and US products like Ocrolus, Blend and Floify all handle documents in some form. A build earns its place when the chasing itself is the problem: borrowers who live on WhatsApp, documents in every format, checks your portal does not run, and lender windows and lists that keep changing.