Lana Lopez starts every Sunday with brunch now. A year ago, that would have been unthinkable. As co-founder of Kulshi, a prediction market valued at more than a billion dollars, her Sundays used to disappear into a full-day audit of every metric, every promise, and every missed deliverable across her fast-growing company. Then her engineering team built an AI system that connects to everything: emails, docs, project updates. Now, as she put it, the week practically reads itself. That shift from buried to free is what separates the 15% of people who report getting real value from AI and the majority who open a tool, poke around, and close it confused. Six leaders with decades of experience between them mapped out exactly how that gap closes.
Why most AI use stays stuck at the search-bar stage
A recent study of 22,500 American employees found that 52% now use AI at work, but only 15% use it daily. The productivity numbers tell the real story: people using AI for one or two tasks report a 45% productivity gain. Push it to seven or more tasks and that number climbs to 90%. Same tools. More kinds of work handed over. The difference is the hours spent understanding how to actually build on them.
Eric, who runs the digital economy lab at Stanford and studies what AI does to work and wages, offered the clearest on-ramp for beginners. Instead of hunting for the right use case, he sits down with Claude or ChatGPT and says: tell me how you can be useful to me, and ask me questions. ‘You literally say keep asking me questions, interview me,’ he explained, ‘and it’ll say okay, what’s your job, do you have kids, what are some of the problems you worried about last week. And it’ll come up with like 10 recommended things that you can be using it for.’ When asked what he would pay to give up AI for a month, Eric did not hesitate: ‘It’s tens of thousands. It would be like tearing off my left arm.’
His one standing rule, and it was echoed by every leader in the conversation: AI is the analyst, never the strategist. ‘I almost never 100% delegate something to AI,’ he said. ‘For me it’s always co-working and collaboration.’
Turning one repeated task into a system that runs itself
Peter Yang left product roles at Meta, Reddit, and Roblox to go solo, and has since built 16 apps without writing a line of code. His entire operation runs on what he calls a skill library: documented prompts for every repeating process, from podcast post-production to newsletter editing to social distribution. The self-improvement loop he built is simple. After a back-and-forth conversation with AI, he adds one instruction: ‘Based on our conversation, can you please update the skill to try to get there in one shot faster?’ The skill rewrites itself. He also keeps a plain text file he calls a personal advisor, loaded with context about his business, his goals, and his decision rules, and each week it sends him a brief covering revenue, content performance across his last 30 days, and how comparable channels are performing. For Substack, which has no API, it uses browser automation to pull the data that would otherwise take him an hour to gather manually. His ceiling for AI, stated plainly: ‘AI gets me to 90%. The last 10% is mine.’
Sal Khan at Khan Academy is watching his engineers run five to ten AI agents simultaneously to write and review code. The organization’s Anthropic bill runs about $1.2 million a year across roughly 200 engineers, and it is growing. One engineer spent $3,000 in compute costs in a single day. Khan’s reaction when he found out what the engineer had actually built: ‘That was a great use of $3,000. Go spend more.’ A feature that would once have taken until the next school year shipped in about a month after an engineer built a working prototype at a weekend hackathon. Khan’s own daily habit is a prompt he runs every morning to a system connected to his Slack and Gmail: ‘What’s falling through the cracks?’ He has also learned to push back on AI sycophancy directly. ‘I’ve learned to say be critical of me, really push back,’ he noted, ‘and even then sometimes it won’t.’
Aaron Levy, who runs Box, a $4 billion company, keeps his prompts long and stores them in documents he returns to constantly. His top recommended tools for anyone starting out: Codex, Claude, and Perplexity. The workflow he will never do manually again is market research: ‘I’ll just never go to Google and type each company in. I’m going to have an agent go and fan out, do all of that, and then maybe I’ll click the underlying sources and verify something.’
The Sunday afternoon Lana Lopez got back
Lana Lopez’s AI system, built by her own engineering team, tracks every update from across the company, flags overpromises, surfaces missed deliverables, and compares week-over-week metrics. ‘Sundays are very heavy days for me,’ she said, ‘because it’s when I stop and look at the entire week. And now I’m actually able to have brunch on Sunday because I have a lot more time to think about things.’
The one rule every leader in the conversation kept returning to: start with one task you do every week, document it fully, and build a skill from it. The brunch is the proof it works.
