The 10-minute loan

A retail bank wanted to put AI into its personal-loan process.

Underwriting a loan has five steps. Market the product. Take the application. Review and approve. Final diligence. Execute.

The obvious target was the middle. A loan officer spent an hour reviewing each file. AI could do that review in seconds.

So they automated it. An hour of work, gone.

And almost nothing changed.

The file still sat five days in a queue, waiting for a loan officer to free up for the next step. The customer still waited. The process still took the same nine days, minus one hour buried in the middle.

That is the trap with AI. You automate a step inside a process that stays frozen, and you get a rounding error.

Some banks did something different.

They asked: if the approval is instant, why does the whole thing still take nine days? And they built a new product. Apply now, approved in ten minutes.

That is not a faster step. That is a different business.

But it forces everything to move. Marketing sells a ten-minute promise. The application routes for an instant decision. Diligence and execution have to scale to match.

Automating a step needs an engineer.

Redesigning the workflow needs someone with the scope to touch all five.

Most companies hire the first and skip the second. Then they wonder where the transformation went.

Where are you automating a step when you should be redrawing the workflow?

Source: Andrew Ng, The Future of AI Agents, Interrupt 26 (LangChain, June 2026)