Why Your AI Needs a Warm-Up Lap: Lessons from Tesla
Tesla doesn’t let you use their autopilot straight away. You first need to calibrate the cameras, which takes 25-30km of driving. It’s frustrating but it’s the only way to onboard someone on a trustworthy autopilot.
To reduce the frustration, they’ve shipped a calibration gauge that shows you where you stand. That’s brilliant!
We recently did the same at Cycle App as we realized that our AI needed a history of context-specific and human-verified data before it started delivering great results. Paradoxically instead of letting users jump straight to the autopilot, we force them to go through a calibration phase. And, like Tesla, we shipped our own AI calibration gauge:
Before your calibration is at 100%, you can’t enable the autopilot features. We believe it’s the only way to create trust as you onboard someone on an autopilot. It’s higher friction to get started but also higher friction to churn from the product.
It’s particularly true for AI products. If time to value is too short without any calibration needed, that’s a sign of a bullshit AI product that gives you results you could easily get with chatGPT – I wrote a blog post that introduces the Bullshit AI test as a way to spot whether some AI is bullshit.
Getting results you can trust
Let’s zoom into Cycle’s use case for a minute. We’ve built an autopilot that automatically processes your incoming customer feedback across multiple sources. If what you want is drop a CSV full of messy feedback to get instant trends without putting any effort, we’re not the right tool for you – ChatGPT is.
A trustworthy autopilot must be calibrated before being enabled. It shouldn’t take you more than 5-7 days to fully calibrate Cycle’s AI. By the end of it we can guarantee exceptional AI accuracy so it’s worth the wait ⏳
AI calibration process
What happens during calibration? You basically:
1/ customize your setup (eg define your taxonomy and choose the types of insights you want AI to look for in your feedback);
2/ capture and process real customer feedback (assisted with AI) in order to provide a history of context-specific & human-verified data that AI then uses to adapt to your context.
That’s it! There’s more friction than dropping an Excel file into ChatGPT and hoping for the best… but you’re guaranteed to uncover non-bullshit customer insights at scale 🤷♀️ Your choice!
Easy onboarding low friction = easy off-boarding, low friction 😊
PS: when you’re done with the generic chatGPT approach and start looking for an autopilot that knows the ins & outs of robust feedback systems – and that keeps getting better the more you use it – then we’ll be around and happy to help 👋
(Btw, we’ve created a bunch of prompt templates and a tutorial on how to start from a messy feedback CSV and get bullshit insights in no time with ChatGPT, let me know if you’re interested)