July 20, 2026

---

Hi Everyone,

Two or three people on your team are getting a lot done with AI, and you don't fully know how.

Most of what they've figured out — the prompts that work, the skills they've built — lives in their personal chat histories, and when they move teams or leave the company, that capability goes with them.

Ethan Mollick at Wharton calls these people "secret cyborgs" — employees who use AI heavily and don't tell anyone, either because they're worried about backlash or because they haven't thought to share.

You don't need a months-long rollout to bring what they know into the open. A short exercise, run team by team, is enough to start.

Start with the tasks people already repeat

When you decide to systematize AI, the instinct is to catalog everything: every prompt, skill, and use case anyone brings up. That usually creates more work than value. Start with the tasks that come up every week. Write those down properly and ignore the rest.

Two examples of what this looks like in practice:

1. At West Monroe, a 2,000-person consulting firm, an internal AI platform called Nigel holds 278 prompts organized by topic and job role, with about 12,000 all-time uses. Nothing exotic — just the tasks consultants keep doing, like client emails, code, and dataset analysis, captured once and shared with the whole organization.
   ​
2. At BBVA Bank, a legal team built a single GPT for one recurring workflow: signatory-authority questions from branch managers. That skill now handles more than 9,000 queries a year, has freed up three people to work on other things, and has hit a quarter of the legal team's annual savings goal on its own.

Where to start

Pick one team — marketing, sales, ops, finance, whichever you're closest to. You're looking for the three to five highest-frequency, highest-value repeated tasks people are already doing with AI.

1. Ask, don't audit: Send a short message to each team lead: "Curious what you and your team use AI for on a normal week. Not looking to police anything, just want to see what people have figured out."
   ​
   Frame it as recognition rather than compliance, because if people think there's a downside to answering honestly, they won't answer honestly.
   ​
2. Pick what's frequent and valuable: For each task that comes back, score two things: how often it comes up in a normal week, and how much time or quality is at stake when it does. Multiply the two, and circle the top three to five.
   ​
3. Write each one down properly: Not just the prompt — for each task, capture when to use it, what good output looks like, where the reference material sits, and one named person who owns it.
   ​
4. Add a check-in date: Six months is a reasonable starting point. By then, either people have stopped using the task or the model has moved on, and the guidance is out of date.

Don't overthink where you keep it. A shared Notion page or Basecamp project works, and so do Claude Projects or custom GPTs if you're already using them.

Go deeper

👉 IT Brew: Best practices for building a prompt library — see how companies organize prompt libraries that people actually keep using.

👉 Ethan Mollick: Detecting the Secret Cyborgs — why your best AI users may be keeping their work to themselves — and how to change that.

👉 Zapier: How Zapier rolled out AI org-wide and drove 97% adoption — inside the rollout that got 97% of Zapier’s team using AI at work.

👉 BBVA: Using ChatGPT for legal queries and marketing — how one legal GPT answers thousands of questions a year and frees lawyers for higher-value work.

Coming up tomorrow

In tomorrow's issue, you'll learn what happens when you tell your team "I might be wrong." The research surprised us.

That's it for today!

P.S. What's the AI task you'd miss most if it disappeared tomorrow?