⚡ At a glance

• The problem: A lot of people drop out of eKYC because they don't have a document and have no idea how to get it.

• The idea: An AI guide inside the KYC flow that knows which documents each person needs, and where and how to get the ones they're missing.

• Where it applies: Mobile wallets, bank accounts, loans, insurance, pretty much any financial product in Bangladesh that runs KYC.

• Type: A concept, built on my KYC work at ShopUp.

The stuck moment

When I designed the document loop for ShopUp's loan KYC, we fixed a lot: people only saw the documents they needed, knew why we needed each one, and got feedback after every upload.

But one problem never went away. Someone opens "Trade license" and they just don't have one. At that point the app has nothing more to say, so the person closes it, calls support or waits for a field agent to visit, and some simply give up.

Our loop told people what to upload, but it never told them how to get it.

Why a checklist isn't enough

Each financial product asks for a different set of documents, and the list changes depending on who's applying.

A checklist can show this, but it can't help with the hard part, which happens outside the app:

• Documents come from different places. e-TIN comes from the NBR website, a trade license from the city corporation, municipality or union parishad, and a bank statement from the bank.

• Documents depend on each other. You need an NID to get an e-TIN, and a rent agreement and holding tax receipt to get a trade license.

• The rules aren't obvious. An e-TIN is free and takes a few minutes online, but most applicants I've seen didn't know that.

All this information is out there, it's just spread across offices, websites and people. An AI guide can bring it into one place, right where the person gets stuck.

The guide

The guide lives inside the KYC flow and shows up when someone is stuck.

What it does:

• Asks a few questions to build a document list for that person

• Explains each document in one line

• When someone taps "I don't have this," it shows how to get it: where to go, whether it can be done online, what to bring, and roughly how long it takes

• Saves progress, so they can leave, get the document and come back

• Reminds them later

What it never does:

• Decide if someone is eligible

• Approve or reject documents on its own

• Fill in government forms for them

• Ask for anything the product doesn't need

The guide helps people get ready, while approval stays with the lender.

Where its knowledge comes from:

Offices, fees and rules change all the time, so if the AI answers from memory, sooner or later it will send someone to the wrong office. That's why the guide only answers from a checked source list covering every document, where to get it and what you need for it. The operations team owns that list and keeps it updated.

Caption: Most documents need other documents first, and the guide shows the order.

Design decisions

1. Ask a few questions first.
Three questions are enough to build the list: what you're applying for, what kind of business you run, and whether it's registered. A tea stall owner and a registered wholesaler shouldn't see the same list.

2. Give the next step, not the reasoning.
Research on AI explanations found they make people accept the AI's answer whether it's right or wrong. So instead of explaining how it decided, the guide just tells people what to do: "Go to the NBR website. You need your NID and phone. It's free."

3. Voice in, pictures out.
In a Microsoft Research study with low-literate users in India, no one finished a banking task using text menus, while voice and pictures worked much better. So people can ask in spoken Bangla, and answers come with an example image of the document.

4. Read back before doing anything.
AI still works worse in Bangla than in English, so the guide repeats what it understood before it changes anything: "You don't have a trade license and your shop is in Mirpur. Right?"

5. Don't block everything on one document.
One missing document shouldn't hold up the whole application. People can park it, finish the rest and come back later, and that category gets a new status: Getting it.

6. Keep a person one tap away.
"Talk to a person" is always there, and the field agent sees the same plan, so they pick up where the guide left off instead of starting from zero.

The flow

1. Your documents: Trade license shows as To do.

2. "I don't have this": The guide asks one question: "Is your shop registered with the city corporation?"

3. The path: A short card with where to go, what to bring and whether it can be done online, plus a link and a "Remind me" button.

4. Back to the list: Trade license now shows "Getting it," and the rest of the application keeps moving.

Caption: From stuck to moving again in four screens.

Caption: Everything you need to get a trade license, on one card.

Takeaways

• The hardest part of KYC happens outside the app. We spent years improving uploads, but people were getting stuck before they ever reached that screen.

• AI is most useful when it connects the dots. Checking photos helps, but knowing how offices, documents and rules link together is where it really makes a difference.

• Keep the AI small and honest. Let it guide people to the next step, leave decisions to the lender, and keep a real person close by.

Most KYC flows stop at telling people what to upload. I think the next step is helping them actually get it.