95% of organizations see no ROI from GenAI (MIT NANDA, 2025). And the interesting part is why: the models are more than good enough. The bottleneck has moved to two deeply human things. Communication with whoever knows the domain, and the ability to redesign a process.
Yesterday's constraint (model quality) is today's non-issue. The failures I see in production are design failures, not capability failures. That changes where the competitive advantage lives: in the workflow, the expertise and the infrastructure you build around the model. Never in the model choice itself.
The right stack is upside down
People → Workflows → Agents → Infrastructure. People (domain competence + judgment) sit on top. The model is one component inside the Agents layer, not the foundation. Design agent-first, not model-first.
The principles I actually use
- Model ≠ Problem. If the agent fails, the flaw is in the design.
- The treasure is what experts don't articulate: the sub-optimal practices they've normalized. One diagnostic question: what would change your life that we've written off as impossible?
- Every piece of data you bring in and model well is a brick that stays yours.
- Redesign the process from zero under one constraint: one person + these AI tools. Automating the human org chart is the recipe for failing with style.
- Predictable quality via evals and pass^k, not after-the-fact review. The expert's standard goes inside the system.
- The harness is the durable asset. Tools, contracts, evals, observability: yours. The model: swappable.
- Start from a high-value deliverable. Days instead of weeks, and measure the quality gain, not just the speed.
- Multiplier, not a margin discount. AI is for unlocking previously impossible work.
The long version of this method is the flagship article on yempik.com. This is the version I keep in my pocket when I talk to a CEO. It works in both rooms.
// the more technical you are, the less the technology is your job.