Back to the podcast
EP04034:27 · Nov 5, 2026

Flip It Inside Out. Let the Agent Do the Work and Tag In the Human.

Most teams start with AI by listing what their people do and asking which tasks a bot can take over. Jay thinks that's backwards. Start with the work, let the agent own it, and have it pull in a human only where judgment matters.

Jeff and Jay dig into what that means for planning next year. They cover why software companies need to sell the outcome instead of features, and why a SaaS tool you can clone in a weekend is worth zero. They also share Jeff's take on boiling the customer journey down to the five to seven moments that matter.

In this episode
  • 00:00 Cold open
  • 00:18 Welcome
  • 00:55 Remembering Adam Robinson
  • 02:00 Planning next year when AI changes how the work gets done. Does token spend raise your cost to serve?
  • 04:58 Start with the problem you're actually solving, not the tools (Good Strategy Bad Strategy)
  • 09:44 Call recordings hold most of your customer context
  • 10:29 Most companies don't use enough tokens yet for cost to serve to matter
  • 12:13 Flip it inside out: the agent does the work and tags in the human. The Spawn example.
  • 16:31 A Figma product manager's agent that runs its own A/B tests and asks him to approve the merge
  • 18:59 Arca isn't a software company, it's a wealth manager. Gainsight sells renewals as a service.
  • 22:15 SaaS cut services to protect gross margin. Now the value flips back to the service.
  • 23:11 "I could clone our PSA tool in a weekend." SaaS is just a CRUD database.
  • 26:24 Data moats and action moats
  • 27:11 Can I vibe code it?
  • 28:32 Agents should ship with a data model, and open source starter kits have a big role to play
  • 29:46 The five to seven customer moments that matter, designed agent first
  • 31:22 Ops becomes the team that runs the agents
  • 32:00 Want your team to get good at AI? Teach them one tool.
  • 33:18 An AI build day in teams of two
Key takeaways
  1. Define the constraint first. Name the problem holding back growth, retention, or expansion before you pick a tool. Then plan by quarter or month, not just by year.
  2. Flip it inside out. Design the work so the agent runs it end to end and pulls in a human for judgment calls and direction changes.
  3. Sell the work, not the software. Customers want the renewal, the approved timesheet, the profitability report. If you don't have a data moat or an action moat, features alone won't hold up.
  4. Cut the journey to the moments that matter. Most teams overbuild customer touchpoints. Find the five to seven that count, and design each one so the agent does the work first.
  5. Teach one tool deeply. Skills learned in one agent platform carry over to the others.
Mentioned in this episode
  • Good Strategy Bad Strategy by Richard Rumelt
  • "How the most AI-pilled company in the world works," a podcast episode about Spawn [add link]
  • How I AI with Claire Vo, the Figma product manager episode [add link]
  • Latent Space with Aman Singh of Arca, on how agents run a wealth management firm [add link]
  • Can I vibe code it? (canivibecode.com)
  • Satya Nadella's "SaaS is just a CRUD database" remark
  • Balboa is working toward Anthropic certified partner status

Want the thinking behind the episodes?

The newsletter finds the idea. The podcast pressure-tests it out loud. Get it every Sunday.

Read the newsletter