How AI Reduced Manual Driver Verification by 75% — Operations Case Study. Part 2

In my previous article, I described how we were able to reduce the workforce by 75% by changing the user verification flow.

How It Was Before

• Driver registers an account in the ride-hailing app


• Uploads photos of needed documents by himself/herself


• Uploads photos of the car


• Waits in line until somebody from the support team can pick up the profile and verify it manually

The Bottleneck

Mostly it works fine, but during high loads, marketing campaigns, or some events, there were spikes of registrations. And sometimes up to 10 people in support working in parallel were not able to handle this load, because they also had chat communication support.

So the most critical part is not to save cost on support actually. It is a benefit, but still not the goal. The goal was to handle spikes and make it possible for drivers to go online as fast as possible.

Solution

• After uploading all photos, instead of going to the human support queue, drivers go to the AI queue


• The AI model verifies everything using detailed prompts. It checks car number plates with documents, checks car photos from all angles, checks avatar, verifies the quality of photos, name in profile and documents, etc.


• Once it is validated, the driver is marked as verified by AI and is ready to go online


• For the new drivers we always ask passengers "is the car correct in reality?" If NO, then it will be immediately deactivated and passed to human support


• In case of failure, humans step in and check it manually

You may think that AI can make mistakes and pass some "bad" drivers, but humans may too.

The Tech Stack

Self-hosted Qwen3-VL-30B-A3B-Instruct-FP8.

Why? Because it fits on one GPU, is cheap enough to run, and good enough to handle photo verifications.

Backend uses GoLang and NATS queue to handle load and queue all verifications. Nothing special here to describe. It's just a FIFO queue.

So the goal was to speed up the verification process, but the benefit is also cost reduction on human support.

We may like it or not, but if this automates the process and helps save some money, why not use it.

The next step is obviously to automate chat support. Half of the questions are very simple and general, it obviously must be automated.