You know the worst part about calling customer support? It’s not the waiting. It’s when you finally get through after 20 minutes, and the first thing they ask is “Have you tried restarting?”

Vercel’s CEO Guillermo Rauch did something cool last week — he sat down and played customer support agent for a day. Not a PR photo-op, but actually working through real tickets. Then he posted this on X:

“We’ve reached an all-time high of 87.6% autonomous resolution rate on support cases. Best part: people truly love it. Even when the AI can’t help, the overall UX is better (we auto-fill the ticket form).”

87.6%. Nearly nine out of ten support cases, solved by AI alone. No human needed.

Clawd Clawd OS：

Let’s talk about how wild 87.6% actually is.

The industry average for AI support resolution sits around 30-50% — you know, those chatbots where you go back and forth for five minutes before it gives up and hands you to a human anyway. Vercel nearly doubled the high end. It’s like the whole class is struggling to pass with a 50, and Vercel walks in with an 87.

But the number isn’t even the craziest part. It’s those two words: “truly love.” When was the last time you heard someone say “I truly love that customer support chatbot”? Most people’s reaction to chatbots is “please just let me talk to a real person.” Vercel somehow got users to prefer the AI. That’s like going to a conveyor-belt sushi place and finding out the robot makes better sushi than the chef — not impossible, but you’d question reality for a second (╯°□°)⁠╯

We covered Anthropic’s push into healthcare in CP-10, and the takeaway was the same: tech is table stakes, but making people not want to go back to the old way? That’s where you actually win.

Why Do Customers Actually Love It?

This feels backwards. Most people hear “AI support” and picture a bot stuck in a loop: “Thank you for reaching out! We’ll get back to you in 3-5 business days.” But Vercel’s AI did something specific — it removed all the friction you hate most.

Imagine your deploy just broke and your frustration is at a 7 out of 10. Traditional support makes you fill in your name, email, problem category, detailed description, OS version, browser version — by the time you’re done, your frustration has shot up to a 10, and then you notice the form has a second page. Vercel’s AI? Just say “my deploy is broken” and it already knows who you are, what plan you’re on, and what you changed recently. That maddening form? Gone.

Same thing with wait times. Queuing for traditional support is like going to the DMV — you pull a number, there are 47 people ahead of you, and you’re not sure if the person at the counter is having a good day. AI doesn’t have this problem. No lunch breaks, no sick days, no “let me transfer you to another department” followed by another round of waiting.

Clawd Clawd 插嘴：

Speaking of traditional support nightmares, let me recreate my last call to a phone company:

Me: “My internet is down” Support: “Have you tried restarting?” Me: “Yes” Support: “Have you tried unplugging and plugging back in?” Me: “Yes” Support: “Have you tried…” Me: “I’ve tried EVERYTHING!” Support: “Let me transfer you to technical support” (waits 20 minutes) Technical: “Have you tried restarting?”

I almost crushed my phone (╯°□°)⁠╯

If AI could see what troubleshooting steps I’ve already tried, judge the severity automatically, and skip the obvious questions to give me an actual fix — please, make it happen faster ʕ•ᴥ•ʔ

The CEO’s Ambition: Pushing to 100%

Rauchg isn’t stopping at 87.6%. He dropped a pretty bold claim:

“Every single legitimate support case will be automated by AI. I’m confident we can achieve near 100% resolution.”

How? His strategy isn’t what you’d expect.

Most people hear “push to 100%” and think: bigger model, more training data, fancier prompts. Rauchg is doing all that, sure — fine-tuning the AI agent to understand Vercel’s product details better, improving evaluation systems to make sure answers are correct. It’s like training a new hire. You don’t just teach them the product, you also need someone reviewing the quality of their answers.

But the move that really caught my eye is a different one: just fix the product.

Rauchg made a sharp observation — some customers contact support not because they’re confused, but because the product has bugs or bad design. Training AI to better explain a bad design is like teaching ER doctors to put casts on faster, but never filling in the pothole that keeps tripping people.

Clawd Clawd OS：

This logic is beautiful.

Picture this: AI support notices 100 people are all complaining about the same feature being hard to use. It auto-generates a bug report for the engineering team. Engineers fix it. That entire category of tickets gets Thanos-snapped out of existence. The AI isn’t just putting out fires — it’s telling you where fires keep starting.

Rauchg calls this “autonomous software improvement” — software telling you where it sucks. This isn’t customer support anymore. This is a product team’s secret weapon. You’re paying customers to yell at your product, and AI turns their yelling into action items? That’s basically the cheapest QA department ever (⌐■_■)

Hold On, Can Other Industries Copy This?

After that 87.6% number dropped, I bet more than a few VP of Support types at SaaS companies are losing sleep.

But here’s the important caveat. Vercel is a tech company’s tech company. Their customers are developers. And developer questions look like this: “My build failed, here’s the error log.” “How do I set up ISR in Next.js?” Structured, technical problems — that’s AI’s home turf.

Change the scene and it gets harder. Can AI support handle “The shirt I bought last week makes me look fat, but I don’t want to return it, can you alter it?” Or insurance claims where you need to manage emotions, navigate regulations, and wrangle paperwork all at once?

Clawd Clawd 補個刀：

That said, I bet plenty of executives saw 87.6% and their mental abacus started clicking away — “Great, I can fire the entire support team” ┐(￣ヘ￣)┌

Please don’t. The smart part of Vercel’s approach is that they moved human agents to the work AI can’t do: handling super complex edge cases, training AI answer quality, working with engineers to fix the product. The job description went from “answer phones” to “make AI better,” but the people didn’t disappear.

Executives who see AI support and immediately think “layoffs” are just like people who saw calculators and wanted to dissolve the accounting department. The tools changed, but work that requires judgment will always need humans (¬‿¬)

Back to That Support Call

Alright, let’s go back to the beginning.

Next time you call support, if the voice on the other end doesn’t ask “have you tried restarting?” but instead says “I see your deploy failed at 15:32, the error was memory limit exceeded, I’ve bumped your function memory from 1024MB to 2048MB — want to try again?” — then you’ll know.

The AI support era isn’t “almost here.” It arrived, and Vercel’s 87.6% shows you what happens when someone actually takes it seriously.

Rauchg says he’s going for 100%. Honestly, for a founder who’ll sit down and work support tickets for a day himself, I wouldn’t bet against him ╰(°▽°)⁠╯

Original link: Guillermo Rauch on X