AI Agents Signal End of Middle Management, Redefining Work
By Ray Dombrowski ·
A new playbook suggests companies must build autonomous agents per customer to maximize lifetime value and abandon traditional workflow structures.
On a recent episode of "a16z," the discussion around AI quickly moved beyond talk of chatbots into radical corporate redesign, presenting a vision that challenges decades of established labor structure. The core argument—that middle management concepts are largely obsolete in an agent-run company—is startling enough to make any seasoned observer pause and re-evaluate their own career path.
The speakers detailed Kavak’s transformation from a used car marketplace into what they describe as an "agent-run company." On the podcast, the speaker argued that investment must shift from knowledge workers to tokens because superhuman agents can be built. This is not merely about integrating ChatGPT; it requires rebuilding core APIs and systems so autonomous agents can perform tasks. The goal, according to Kavak’s playbook, is to maximize Customer Lifetime Value (CLV) by building "an agent per customer," giving each a virtual machine and memory of the customer's entire history. This shift fundamentally changes the measure of success: it moves from being transactional (counting cars bought/sold) to relational.
The End of the Workflow Graph
The architectural bet described is profound. Kavak moved away from building multi-agent workflows or graphs toward a robust system centered on long-running agents with hard goals. These agents are not just glorified call center scripts; they act as "mega experts," handling complex processes like buying a car, which involves financing, insurance, and trade-ins simultaneously. The results cited—tripling the NPS and increasing conversion rates by over 2.1x compared to previous human teams—are compelling data points that cannot be ignored when analyzing productivity gains.
However, this model requires total organizational commitment. On the podcast, it was emphasized that success depends on generating data and feedback loops by putting agents in front of customers (getting "evals") to continuously fine-tune them. This is where the rubber meets the road for any industrial policy analyst: continuous improvement doesn't come from a mandate; it comes from measured failure within a live market.
The New Labor Contract
For those of us who have spent decades judging economic health by payroll, the implications are dizzying. Kavak’s approach mandates that all employees—from CEOs to mechanics—must undergo retraining through initiatives like the "Jedi Academy." This acknowledges that every job will change, requiring people to collaborate with new technology rather than becoming specialized AI engineers themselves.
The structure itself is described as highly flat and senior-team focused, rendering middle management concepts obsolete. While this promises unprecedented efficiency—evidenced by reducing car loan approval time from months to under three minutes—it also suggests a massive structural labor shock. The human role shifts from executing processes (the job) to supervising the system that executes processes (the oversight).
Measurement and Mandate
What holds up in this vision is the relentless focus on measurable business outcomes over superficial KPIs. Measuring success by "Did it convert? Is the customer happy?" rather than counting calls or minutes is a necessary corrective to much of modern service labor management. Furthermore, the advice that AI adoption must be top-down—requiring leaders to set a clear, vertical strategy for 3–5 years—is sound industrial policy. Bottom-up hackathons are often merely novelties; structural change requires executive mandate and capital allocation toward specific "tokens."
However, I remain skeptical of the speed and completeness of this transition. While the promise of agents handling 96% of interactions is impressive, the historical record shows that organizational inertia—the very concept of middle management—is sticky. The true challenge isn't building the agent; it’s dismantling the human structures built around inefficiency and process control over decades.
The next wave of industrial policy won't be about optimizing existing labor roles; it will be about managing the painful, rapid obsolescence of entire organizational layers. The most valuable asset in this new economy will not be the AI itself, but the ability to rapidly retrain a workforce capable of supervising and debugging autonomous systems.