Vol.034 — Two Ways to Build AI Workflows. One Scales Fast and Hides the Cost. One Scales Slow and Owns It.
Date: 2026-05-26 | Newsletter
Key Summary
Two signals landed in the same 72 hours: Anthropic shipped Claude for Small Business with 15 native integrations and Dario Amodei publicly tightened his timeline for professional job displacement; on the other side, a Singapore Reddit post described a company that replaced three humans with one AI, saves hundreds of thousands per region per year, and absorbs a 5% error rate because the math still closes positive. Zenta’s argument is that the 5% is not a tolerance — it is a design choice. Path A processes everything, ships the errors, and learns nothing from the complaints. Path B flags those same 5% of cases before they ship, runs a human review with full context, and turns every override into research data about where the model still fails. Same headline error rate; completely different operations. For ASEAN SMEs, the stack difference is real too: Claude for Small Business shipped into QuickBooks, HubSpot, and PayPal, while most regional operators run Xero, LINE, Zalo, GrabPay, and DuitNow. The integrators matter less than the design question — is the human built into the loop as a quality and learning mechanism, or bolted on after the errors surface? Do it simple. Design the human in. Don’t bolt them on.