Why the Age of AI Needs Technical Program Management More, Not Less
Every few months someone tells me AI will automate program management. They're half right: AI is coming for our artifacts. It's the judgment underneath them that just became the scarcest resource in the industry.
I've now run launch programs on both sides of the divide — consumer devices and platforms at Google, and frontier AI models at the Chan Zuckerberg Biohub, where our team shipped a world model of protein biology to the global scientific community. The mechanics of the job changed more in that transition than in the fifteen years before it. The necessity of the job went the other direction: up, sharply.
What AI is actually doing is separating two roles that were always distinct but were often performed by the same person carrying the same title. The coordination manager — whose function was information relay, task tracking, status aggregation, and progress synthesis — is being replaced, because that specific work was always a symptom of a poorly designed system. AI executes it better, faster, and without the organizational drag that comes from routing everything through a single person.
The systems-architect manager is experiencing something different. Their function is to decide how human judgment, AI output, onshore capacity, and offshore execution combine into something that produces measurable results. They design the decision-rights architecture that determines what requires a person and what can be delegated to a process. They build the feedback loops that improve team performance without their constant presence. When they are out, the team does not collapse. When AI becomes available, they extend it into the architecture rather than being replaced by it. That is the role I've built my career around: the systems-architect manager, designing the operating system of the organization.
What's genuinely different
"Done" is now a probability distribution. A traditional launch has a feature checklist; you either shipped the API or you didn't. A model launch has eval scores. The product is probabilistic, which means the definition of done is a threshold on a benchmark, a bar on safety behavior, a judgment call about whether the win-rate is real or an artifact of the eval set. TPMs who grew up on binary exit criteria have to learn to program-manage toward evidence, not completion. Milestones become experiments: you can schedule the training run, but not what it teaches you.
Compute is the new factory slot. In hardware, I learned that a booked production line doesn't negotiate. GPU capacity has exactly that character — reserved months out, costing millions, with an opportunity cost measured in every experiment you didn't run. A large training run is the closest thing software has ever had to a tape-out: enormously expensive, effectively irreversible, and demanding that dataset decisions, architecture bets, and eval readiness all converge before the clock starts. That convergence is a program management problem, and the price of getting it wrong is no longer a slipped sprint.
The dependency graph grew new dimensions. Model programs braid together research, engineering, infrastructure, data procurement and licensing, safety review, and go-to-market — disciplines with different definitions of rigor, different clocks, and different failure modes. At the Biohub we tracked model milestones, GPU cluster utilization, and pipeline health across distributed teams in real time, because no single person could hold that graph in their head. The seams between functions — where programs have always lived and died — have multiplied.
The field moves weekly. Annual planning assumed the ground held still for a year. In AI, a paper published on Tuesday can invalidate a roadmap assumption by Friday. The answer isn't to abandon planning; it's to plan the way researchers experiment — direction held firmly, details held loosely, with explicit checkpoints where new evidence is allowed to change the plan.
Why that makes TPMs more necessary
Here's the irony in the automation argument: the parts of program management AI genuinely automates — status collection, tracker hygiene, meeting summaries, dependency charts — are the parts that were never the real job. I've written before that TPM maturity runs from tracking to unblocking to shaping. AI is rapidly commoditizing level one. What it cannot do is decide which irreversible bet to take when the eval results are ambiguous and the cluster reservation starts Monday.
AI didn't shrink the job. It burned away the clerical shell and left the judgment exposed — and judgment is the part that was always in short supply.
Meanwhile, the demand side exploded. Every AI organization is now running the hardest cross-functional program in its history, usually for the first time, usually under competitive pressure, often with researchers who have never shipped and engineers who have never trained. Someone has to hold the whole system: the science, the compute economics, the safety bar, the launch commitments. Organizations that treat that as overhead discover the cost the first time a nine-figure training run starts without its data pipeline validated or its evals ready.
What to carry forward, what to leave behind
From the old world, keep the discipline of irreversibility: spend disproportionate energy on the decisions you can't take back — training runs, data licenses, public capability claims — and give teams autonomy on everything else. Keep backwards planning; a model release rehearsed from launch day in reverse surfaces eval gaps and infra dependencies months earlier. Keep making bad news cheap to report, because in AI the bad news arrives as a quiet metric, not a loud crash.
Leave behind the identity that was built on artifacts. If your value proposition is the tracker, AI is your replacement. If your value proposition is judgment at the seams — knowing which clock each discipline runs on, which risks are load-bearing, and when the plan should bend to the evidence — you're not being automated. You're being promoted, whether your title changes or not.