Use caseRetail banking
Claire applies from her phone on Sunday. On Monday an agent sees her bill is four months old, and the account waits days for a new one.
Each proof is typed and read on uploadDates and names are checked against the bank's rulesThe agent clears every gap before the account opens
Driver's license, valid until June 2031. Last name, first name and date of birth match what Claire declared in the app.
source · Driver's license, front and backOne point needs a person · the check says which, with its source
How long does it take a bank to notice that a utility bill is four months old? For a Sunday application, about a day, plus however long the customer takes to answer. What fills the queue is proofs that are slightly off: a maiden name on the bill and a married name on the license, a street number that differs by one. Each one sends the file back and forth for a week.
The bank moves that check to the minute of upload. The AI recognizes each proof, extracts its dates and names, and runs the rules compliance wrote: which proofs count, how old they may be, what must match. Karim starts from a verdict with its reasons, clears what he can, and decides before the account opens.
None of these screens is new to an onboarding team. The gain sits in their order: the proofs are checked when they arrive, the gaps reach the agent with their reason, and the account opens on a file a person has cleared. We follow Claire's application below, one proof at a time.
Claire uploads her proofs on Sunday evening. The license is read and checked within the minute.
Claire photographs the front and back of her driver's license, then uploads a PDF electric bill. The bank's app calls the repository API: an onboarding file is created, and each document lands in its slot.
The AI recognizes a state driver's license and extracts the fields the checks need: last name, first and middle names, date of birth, license number, issue date and expiration date. Each field comes with its confidence and its place on the image.
The checks are the bank's own, written by compliance in plain language: is this an accepted identity document, is it valid today, do the name and date of birth match the declaration.
An address proof has its own rules: who issued it, how old it is, and whether it matches the declaration. This one fails two of them.
The AI recognizes an electric bill and extracts the account holder, the service address and the issue date. The bill is dated July 8, 2026, four months ago. The bank accepts proofs up to three months old.
Lowell on the license, Garner on the bill. The check looks across the documents on the file and finds the middle name on the license: Garner. It writes that the bill is likely in Claire's maiden name, points to the field, and leaves the check open.
On Monday at nine, Karim opens his onboarding queue. Files marked Complete have already moved to the account opening step. He filters on Attention required, sorted by the age of the application.
One check to confirm, one proof to replace, then the account. Compliance closes the act with a rule update.
Karim opens the license in the viewer and zooms in on the name block: Garner is printed as her middle name. He clears the last name check with a note. Only onboarding and compliance can open these images, and each opening is recorded.
The bill date cannot be cleared. Karim asks the assistant for a request to Claire. The draft says what is fine, what is missing and which proofs she can send instead, taken from the bank's list. He sends it to the app.
Claire uploads her October internet bill an hour later. It lands in the same slot, the address checks run again on it, and the status moves to Complete. The file moves to account opening.
Sylvia, the compliance officer, owns the rules. This month she adds the latest W-2 or 1099 to the accepted proofs of address. The change is a sentence in the rule, live for the next upload, with no release.
Because these are identity documents, the bank runs the extraction model on its own premises. No image leaves its network, and every call is logged with its model and its date.
Each application lands in one of three states. Only the middle one needs Karim, and it tells him which proof to look at.
Accepted proofs, valid dates, names and address that match. Each check leaves its reason on the file, and the application goes on to account opening.
What follows: the file moves to account opening, where an agent opens the account.A bill too old, a name that differs. The check says which, with the document and the line.
What follows: the agent clears it or requests a new proof; the check runs again when it arrives.An expired passport, a document not on the list. The AI lists what passes first, then names the rule that fails.
What follows: the agent decides what to ask for, and the customer is told why.It opens no account. It refuses no applicant. It clears no gap on its own. Across the ten steps on this page, every decision on the file is taken by an agent, with their own rights and their own trace.
The AI recognizes, extracts and checks. The rules are written by compliance, and the identity images stay on the bank's own infrastructure when the bank requires it.
Taken from Claire's application. Your own numbers will follow your list of accepted proofs and your weekend volume.
Where these numbers come from. They are counted on the scenario described on this page, with example data. They are not a customer result.
Onboarding, compliance and IT all sign off before a check like this goes live. Each gets something different out of it.
Every photo gets looked at, just to find the one detail that is off.
Approving a check you cannot read is out of the question.
The app, the back office and the archive each keep a copy of the same license.
With the proof of address. It fails most often, its rules are simple, and its date is easy to check. A few weeks of real uploads will show how many files go through clean.
Identity documents join the same check afterwards, with the list compliance already keeps.
Ready to try
Bring your list of accepted proofs and a few anonymized samples. We will run the same checks on them with your onboarding and compliance teams.