The editor layer

The call AI would have gotten wrong

Six weeks out from the biggest trip on my desk, 25 travelers across six families headed to Kenya and Tanzania, timed to the great migration, Qatar Airways cancelled the group's flights.

Every rebooking tool, and every AI agent I could have pointed at the problem, would have solved for the same objective: restore the original plan as closely as possible. We did something else. We rerouted the group through London and made the disruption the trip's strongest opening, absorbing 12 time zones in a way the kids could handle. In the reserve itself, the group stayed at Olive Bush Camp over the high-end Mara names, because on safari, location beats brand.

None of that was a data lookup. It was judgment about timing and sequencing, and about when the obvious answer is wrong.

What the machines actually do

En Route Luxe, the advisory behind that trip, is not an AI demo. It books $500K+ a year, grown entirely through repeat bookings and referrals. And it runs on an AI operating stack I built and ship daily: agents for destination research, itinerary assembly and QA, and the content pipeline. Automated guardrails check every itinerary before a client sees it: closure dates, transfer buffers, ticketing rules, contingency paths.

The stack matters because of what it does not do. Agents draft. A human reviews. The judgment calls, supplier selection, disruption response, taste, stay human on purpose. That boundary is not a limitation of the tools. It is the product decision.

Where the pattern came from

I learned this at enterprise scale before I ran it in my own business. At Zillow I spent six years building document intelligence: a platform processing 1M+ real estate documents a year, with multimodal LLM document understanding in production. The model was never the product. What earned professionals' trust was the system around it. Golden datasets and per-field accuracy gates decided what rolled out. Confidence-based routing decided what a human reviewed. Automated accuracy went from ~85% to 98%+, and the compliance-critical fields a human reviewed held at ~100%.

Brokers did not trust it because the model was impressive. They trusted it because the failure modes had owners. The same pattern holds at both scales: if a rule can be a script, it is a script; guardrails run before output ships; human judgment sits where the stakes concentrate.

The editor layer

I call this the editor layer, because the work looks like editing. The agents produce more than I ever could alone. The value is deciding what ships.

Adding AI features is easy now; every roadmap has them. The harder job is subtracting: turning off the automation that degrades the product, and refusing to automate the decision where a customer's trust actually lives. Shreyas Doshi has argued that "tools have never been a significant source of alpha" in product success, and AI tools do not change that. AI raises the floor for everyone at the same time. What compounds is the editorial judgment sitting above the tools.

That is the PM job now. Not writing better prompts. Deciding, in your product and your domain, what must never run on autopilot.

What are you choosing not to automate?