I delivered a 15-minute keynote at Jio Institute’s Convergence 2026 at NTU on Thursday.

The topic was “Data Storytelling” - a bit jarring in the middle of an AI event. Shailesh picked it and I just rolled with it. A spent several days worrying, “How the heck do I say about data storytelling, when most of my recent workshops and talks are about the death of my data storytelling approaches?”

After a discussion with ChatGPT and Claude, I settled my usual strategy these days:

  1. Agents know more than me anyway, so I don’t see what value I add.
  2. So instead, I’ll survey the audience and share insights. That survey data is new.

I had ChatGPT analyze the audience and create the survey, even implement it using gws (didn’t know it could do that!) and had it analyze the results.

Two survey results were striking.

(ASIDE: it took a few painstaking hours and last minute panic to pick the two. There were several insights. Maybe some othere were more important. I just happened to choose these.)

  1. What mainly stops you from handing that task over to AI today?
    Top answer: “I’d have to check it anyway”. In other words, verification.
  2. Imagine AI became 10× more reliable and could securely access all your work systems. What part of your work would you still want to keep for yourself?
    Top answer: The final judgement call. In other words, delegation.

Which got me to do think: How do I actually verify and delegate? The lightbulb 💡 moment was realizing that I could actually find this out from my own data.

So, on the bus ride, I asked ChatGPT:

  1. How do I verify AI’s work?
    • Go through our chats in August 2026.
    • Find all techniques I used to verify / make it easy for me to verify the output.
    • Factor in their effectiveness.
    • Share as a prioritized list with examples
  2. How do I delegate to AI?
    • (… same thing)

Here are my top verification techniques:

  1. Run it. Don’t just read it. In other words, tell agents to run test cases. Works well with code, but you can programmatically verify mathematical proofs and even contracts.
  2. Triangulate sources. Ask it to find evidence from multiple sources.
  3. Set the test first, i.e. define your acceptance criteria - or at least, tell it to define acceptance criteria first, so you can verify if that’s OK.
  4. … and so on for 18 of these.

I didn’t verify the verification techniques! But it’s something I plan to go back to and see if (a) there’s something I should do more of and (b) there’s something I’m not doing enough of.

Here are my top delegation techniques:

  1. Draft only. NEVER send. That retains control with very low effort.
  2. Work independently. I tell it to ask me only if it really needs to.
  3. Implement, test, fix. I ask it to go ahead with the implementation without review when tests are available.
  4. … and so on for 18 of these.

I didn’t verify these either, and I’m not as happy with these - perhaps because I didn’t understand them well, or I didn’t ask the question well, or I haven’t spent enough time exploring autonomy / delegation. But I do plan to explore more.

What I learnt / re-learnt:

  1. Even if you have nothing to teach, discovering from the audience can teach.
  2. Your workflow logs are an insight source if people want to learn from you.
  3. Speaking under-prepared is scary but educational.