|
Last week I was in a session with the operations lead at an investment fund. Her quarterly investor capacity calculation was due that day. She had saved a skill for it a while back. She had her doubts, and said so out loud. She fully expected it to fail. She ran it anyway. Midway through, the skill paused and asked her a question. Was a new investor exempt from the limit? It also picked up transfers that hadn't been booked into the capital activity tab yet. Something she would likely have missed on a busy day. When the headroom number came back, she checked the change against the prior quarter. "So this is actually perfect." Time saved? "Maybe 15 minutes, 20 minutes." That sounds small. Fifteen minutes a quarter is an hour a year. What happened next is key. She and I added two lines to the skill: ask about exemptions up front, and always add transfers to the activity tab. Every question a skill asks you mid-run is a line missing from its instructions. Think about how you'd train a new analyst. The first time they run the quarterly numbers, they come to your desk four times. You answer, and if you're smart, you write the answers down. By the third quarter, they come once. The skill works the same way, except it never forgets what you told it. So the bigger win here is that next quarter, the skill knows what she knows. It'll stop asking questions. If you want to try this loop yourself: 1. Pick one recurring calculation or report. Something you do monthly or quarterly. Save the steps as a skill or a saved prompt. 2. Run it on this period's data. Keep a running note of every clarifying question it asks and every correction you make. 3. Turn each one into a rule. After the run, add them to the instructions. "Always ask whether a new investor is exempt before calculating." "Always include unbooked transfers." 4. Run it again next period and count the questions. Fewer questions means the skill is learning your job. I like this as a measure because it's honest about where AI is today. It won't know your exceptions on day one, just like that new analyst didn't either on their first day on the job. I am taking one new client in 2026. Apply to work with me here. When the spot fills, I'll be looking at my client roster for 2027 so ping me if that timing works better for you. Alex |
As an AI Coach, Advisor, and Agent Builder, I help organizations and business leaders harness the power of artificial intelligence to boost productivity and streamline operations. I enable organizations to navigate the transformative landscape of AI, educating teams, identifying operational and strategic opportunities with AI and creating a framework for safe and transparent use of data in the organization.
This week I went to a lunch at Stanford Law with Ethan Mollick and Kevin Roose. If you haven't read them: Ethan Mollick is the Wharton professor behind Co-Intelligence, the book that showed readers how to use AI productively. His latest, Co-Existence: The Next Phase of AI, is about working well in the uneven zone where AI's abilities and ours overlap, the jagged frontier. Kevin Roose wrote Futureproof, and his new book, The AGI Chronicles, draws on more than 150 interviews to tell the story...
This week I spent a session with a consulting firm. Their clients care a great deal about confidentiality, so the firm doesn't record or transcribe client meetings. Someone types notes live in Notion instead, in one long running page for each client. Then the typing starts again. After the meeting, a person re-enters the takeaways into Asana as tasks. Then a person writes them out a third time in a follow-up email. Same information, three rounds of typing. I'm tired just thinking about all...
Last week I was on a governance call with the general counsel of a hedge fund. The call was part of a firm kick off for the AI program. One of our first topics was: do we give people access to Claude Cowork or Claude Code? Here's how I reframed it. Cowork is an "ask for permission" environment. You pick the folders it can touch. Code is "ask for forgiveness." It was built for developers, and it starts with broad access to your machine. It related to a conversation a few days later around...