Good AI Is Frictionless—Until It Shouldn’t Be

Shriram Rajagopal is Chief Product Officer at VitalEdge Technologies.

​If you write a lot, you probably use autocorrect and grammar check. Autocorrect works in the background, fixing obvious issues so you can type as fast as you can think. Grammar check stops you—not to fix it for you, but to ask if you really meant it.

A good AI platform also knows when to let you go full speed, and when to introduce some much-needed friction. It lets the user get work done as quickly as possible, while taking key opportunities to say, “Are you sure?”

Most platforms are optimized to be frictionless, which is a problem because it feels the same whether the output is right or wrong. The model still performs, but doesn’t tell the user when to take a closer look.

The key is in the balance, and knowing how to get the most out of both speed and friction.​

Whether you build AI tools or buy them (or both), your job is making sure users can actually get to the tool and use it without getting frustrated.​ Here’s how I determine whether an AI product has the right amount of friction, based on my experience building AI-driven software for equipment dealerships.

Remove obstacles.

A field technician I spoke with went on a service call for a bulldozer. Standing at the machine, he opened his AI application, pointed his device’s camera at the bulldozer’s serial-number plate and scanned it to automatically bring up the most common repair issues.

When you consider that the alternative for the technician is taking off his gloves, squinting in the sun and typing in a 17-character string of numbers and letters, you understand why the scan matters. An AI feature that cannot be easily reached or operated is an AI feature that, for all practical purposes, does not exist.

Lower the cognitive burden.

In human-factors-guided application design, we need to manage the cognitive burden.