A point of view on AI, backed by the work.

Generative and agentic AI have made the breadth of software engineering — the plausible and conventional code — nearly free. The first casualty is the junior engineer, whose old work (the boilerplate, the well-specified tickets, the first-pass tests and debugging) was both the team's production and its apprenticeship.

Cheap breadth raises the value of two scarce human contributions: depth, the judgment of which answer is right in this system, and divergence, the perspective from outside the consensus. Left to inertia, an organization slides toward a substitution equilibrium — review theater where the signature stays but verification drains out, a starved apprenticeship, and output that converges on the model's mean.

The alternative, complementarity, is not a slogan. It is a set of concrete management moves, grounded in a worked enterprise data-pipeline case with a labor-share projection to 2032 and a clear-eyed section on the risks and conditions of reversal.

Fig. 1 — the cost of breadth falls toward zero while the value of judgment rises. Where they cross is where management has to change.

Most enterprise AI programs are approved on conviction — a striking demo, a competitor's press release, the fear of being left behind — and then quietly written off. The paper starts from the uncomfortable number that roughly four in five never earn back their cost, and argues the failure is almost never the model. It is the absence of a discipline for deciding what to fund, when to stop, and how to price the risk that a project simply will not land.

Working from eight enterprise case studies, it builds that discipline into four analytical layers. An initiative moves from a go / no-go readiness check, through a decomposition of where return actually comes from and a stage-gate that releases money in tranches, to a single A-to-F grade — with a risk-adjustment table that tunes the discount rate by category, so a speculative bet and a routine automation are never valued on the same terms.

The payoff is a shared language for the funding decision. Boards get one comparable grade across competing bets, and teams get a defensible reason to stop a weak project at the next gate instead of letting it drain a roadmap for a year. It is the same framework behind the executive curriculum I teach to the C-suite.

Fig. 2 — the four layers, exploded. Eight case studies rise to the layer each one tested.

Commanding a swarm the usual way does not scale: an operator drowns the moment every drone needs its own heading, target, and rule for what to do when the plan breaks. Drone-Megaddon borrows the grammar of the real-time strategy game instead — the operator sets high-level intent, formations, objectives, and rules of engagement, and the system resolves it into the per-agent control underneath.

The idea is not new to the AI moment. The original interface was built in 2014 as a Carnegie Mellon capstone: a command layer for autonomous swarms years before the tooling to run them was widely available.

The 2026 extension, “Co-Intelligent Swarm Control with Language and Vision Models,” closes the loop with today's models. The operator can command in natural language, and the swarm can see, interpret, and adapt to a scene rather than execute a fixed script.