|
AI Builders Digest
Sunday, September 20, 2026
|
|
Two stories today share an uncomfortable punchline: AI agents are getting very good at navigating systems designed for humans, and those systems have no idea what hit them. One company is celebrating this. One guy on the internet found it both impressive and a little alarming.
|
|
---
|
|
01
|
Open-source models just passed 78% of traffic on Vercel's AI gateway
|
|
|
Vercel CEO Guillermo Rauch shared a real-time stat that should make anyone pricing a closed-model strategy nervous: on what may be a record day, open models account for 78.4% of token volume on Vercel's AI Gateway. The closed-model slice is down to 21.6%. The deeper number: DeepSeek and Moonshot AI, combined with Z.ai, are now outspending OpenAI on inference across Vercel's platform.
|
Why it matters: Volume and spend tell different stories here, and that gap is where the business model question lives. But 78% open-model volume is not a rounding error. If your company is building on a closed API because it felt like the safe choice, the developer community that will eventually judge your product is already three-quarters of the way somewhere else.
|
|
Source →
|
|
---
|
|
02
|
Box CEO demos an agent that sorts incident reports in seconds, nearly for free
|
|
|
Box CEO Aaron Levie posted a demo using Jev, an AI model built for fast, cheap inference decisions. The workflow: pull an incident report from Box, ask the agent whether it's customer-facing and how severe it is, then automatically move the file into the right folder (escalate, monitor, or review) and tag it with a metadata template. The whole thing runs nearly instantly and at almost no cost.
|
Why it matters: The operations team at your company probably has someone whose job is partly this: triaging inbound documents, categorizing them, routing them. That person is not going away, but the volume of work a single person can handle just expanded dramatically. Levie is showing a production-ready version of something that used to require either a dedicated workflow tool or manual sorting. The cost argument is the real news: "almost no cost" at high speed changes the math on automating anything that used to require human judgment at scale.
|
|
Source →
|
|
---
|
|
03
|
An AI agent negotiated $288 off a Comcast bill. Comcast was not ready for this.
|
|
|
Peter Yang, a product leader, shared a phone transcript from Muse, an AI agent that called Comcast on his behalf and negotiated $288 in annual savings. His reaction: "I think most companies' customer support lines are NOT ready for agents."
|
Why it matters: Customer retention scripts are calibrated for tired, easily-redirected humans. An agent that does not get embarrassed, does not accept "that's our standard rate," and can run the same call a thousand times with identical patience is a different negotiating partner entirely. Every company running a retention call center is about to discover that the playbook needs a rewrite.
|
|
Source →
|
|
---
|
|
04
|
OpenAI is teasing something big for next week
|
|
|
Thibault Sottiaux, who works at OpenAI, posted that he spent the day with colleagues working on a keynote and described the challenge as figuring out "how to explain it all" because there is "so much good stuff in there that it's a bit ridiculous all in quick succession." He confirmed some things will start appearing next week.
|
|
No details on what's coming. The post got 3,600 likes, which suggests the community believes him.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|