THE NUMBER
$300 a day. That is what tech investor Jason Calacanis said on the All-In Podcast his AI agents cost him after he turned them on without a cap - more than $100,000 a year per agent, completing a fraction of the job, and asking out loud: "When do tokens outpace the salary of the employee?" That math is what this newsletter is about.
3 THINGS HAPPENING RIGHT NOW
He tripled his capacity - then saw the monthly bills.
Nova, a freelancer who spent four months building AI agent workflows for 12 client projects, handled 73% of customer inquiries without touching them himself. His personal project capacity tripled; response time dropped from four hours to under two minutes. Then he ran the numbers: at least $19 a month for Flowise plus $30-80 in OpenAI API costs. And platform updates broke two live client bots without warning.
Three-quarters of companies that deployed AI customer-service bots shut them down
A survey of more than 2,500 AI decision makers, covered by tech publication The Register, found nearly three-quarters rolled back or shut down AI customer-service agents after deploying them. The survey was by Sinch, a vendor with a stake in the finding; The Register covered it independently.
Video production just became something you can delegate to an agent
OpenMontage, an open-source system that turns AI coding assistants into a full video production studio, picked up more than 3,700 GitHub stars in a single day last week and now has more than 18,000 total stars. It runs 12 pipelines and 400+ agent skills: script writing, visuals, voice, music, final cut. Hiring video production out costs hundreds to thousands of dollars; OpenMontage handles the full pipeline for the cost of running the model.
THE DEEP DIVE
The 30% He Refused to Automate
Sidharth Barman runs a freelance consulting business. When he built an agent to handle client onboarding, he could have automated all of it - the intake form, project scoping, the contract, the kickoff call. He automated 70% and stopped there on purpose.
The 30% he kept: pricing negotiations, calls where the client sounded hesitant, any judgment call about whether a project was actually a good fit. Not because the agent couldn't handle them - because those were the moments where his business ran on relationship and judgment, not process.
One outcome he did not expect: clients who were not a real fit dropped off early in the automated flow without consuming hours of calls with prospects who weren't going to convert. Every client who made it through arrived aligned and prepared.
What he'd do differently if starting over: identify what needs human judgment before building anything. Protect those parts.
This applies to any service business where part of the value is the relationship - consulting, coaching, legal, creative work. The agent handles the volume. You handle the judgment. That split is the strategy.
ONE THING TO TRY THIS WEEK
Before you build an agent, watch it do the job once.
Calacanis found what his agents cost by running them. Barman figured out what not to automate by thinking it through. There is a middle step: test the task live before building the system.
Open Claude Code in any folder on your computer.
Pick one task you do repeatedly: answering inquiries, replying to reviews, following up with leads.
Find a real example from your business: an actual email, review, or message.
Type this:
Here is a task I do regularly:
[Describe what you normally do with this - 2-3 sentences]
Here is a real example:
[Paste the actual email, review, or message]
Do what I would normally do with this. When you are done,
show me the result - and tell me what you were not sure about.
Read what Claude produces. Would you send that? Use it? If yes, you have a working proof of concept before building any infrastructure. If no, you know what to fix.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
Next issue: the tiny AI-run team - a handful of people who staffed an entire company's worth of functions with AI agents, and the numbers they hit doing it.
Manu