Why a claim takes nine days when the work takes forty minutes

Ask an underwriter or a loan officer how long it takes to assess a straightforward file and the answer is usually somewhere between thirty minutes and an hour. Ask the customer how long it took and the answer is a week and a half.

That gap is the whole business problem, and it is worth being precise about where it goes.

The nine days, broken down

We have sat with teams and timed this. The shape is remarkably consistent:

Day 1 — the file arrives incomplete. Almost every file does. Photographs taken at an angle, one page missing, an ID that is unreadable at the corner where the number is.

Days 1 to 5 — the chase. Somebody calls. The customer promises to send it. They send the same photograph again. Somebody calls again. This is the largest single block of the nine days and none of it is work.

Day 5 — re-entry. Details are typed out of the photographs and into the system. Twenty minutes, plus the errors that get discovered later.

Days 6 to 8 — the queue. The file now waits behind other files, because the person who assesses it also spent their week chasing documents.

Day 8 or 9 — forty minutes of actual assessment. The part everyone is trained for, and the only part that requires judgement.

So: roughly ninety percent waiting and re-entry, ten percent expertise. Nobody designed this. It accumulated.

What can be moved, and what cannot

The assessment cannot be automated, and we would not propose it. Underwriting, credit decisions and claim liability are judgement, they carry regulatory weight, and a machine making them alone is how an institution ends up in front of a regulator explaining something it does not understand.

But the eight and a half days around it are not judgement. They are logistics.

Collection can happen where the customer already is. The file arrives on WhatsApp, because that is where the photograph was taken. An agent asks for the documents one at a time, in order, in the customer's language.

The check happens in the conversation, not after it. Is this the right document, are all the pages here, is the number legible? The customer is told immediately, while the document is still in their hand — not five days later by a phone call. This single change removes most of the chase, because the chase exists entirely because nobody looked at the photograph until later.

Extraction happens from photographs, not clean scans. This matters more here than in the markets these products were designed for. Nobody has a scanner. Everybody has a camera and poor lighting.

The complete file joins the queue. The assessor opens a file that is complete, legible and already extracted, and spends their forty minutes on the forty minutes.

What actually changes

Not the assessment time. The elapsed time — the number the customer experiences and the number that decides whether they renew, refer or complain.

And the second-order effect matters more than the first: the assessor stops spending most of their week on logistics. That capacity is what has been missing all along.

The part institutions here should not skip

If you are a regulated institution, the interesting question is not whether the technology works. It is what you can show.

Before anything faces a customer, three things should exist in writing:

1. What the system may never decide alone. Liability, credit approval, rejection, anything with a legal consequence. This list is written and agreed, not implied.
2. What is logged. Every document received, every extraction, every escalation, every decision and its basis. If your audit trail is a screenshot folder, that is a finding waiting to happen.
3. How it was tested. Not "the vendor demonstrated it". A test set built from your own real historical files, with numbers: extraction accuracy, false positives, how often it escalated when it should not have and, more importantly, how often it did not escalate when it should have.

That third one is the one nobody does, and it is the one that will be asked about first. It is also the reason we run evaluation and red teaming as a separate service, including on systems somebody else built — because being told your AI works by the people selling it to you is not an assessment.

What we would ask you first

• What proportion of your files arrive complete on the first attempt?
• How many days of your average file are chase, and how many are assessment?
• What does an assessor's week actually look like, hour by hour?
• If a regulator asked how a decision was reached on a specific file from March, could you show them?

If the last one made you pause, start there rather than with the agent.