Two data points from the past 12 months frame the practical state of enterprise artificial intelligence (AI). On July 14, Demis Hassabis—Nobel laureate and CEO of Google DeepMind—published a framework arguing that AI’s impact could reach 10 times that of the Industrial Revolution, at 10 times the speed. Meanwhile, MIT researchers studying real deployments found that roughly 95% of enterprise generative-AI pilots are producing no measurable return. Both are true at once. The technology is compounding, but most corporate implementations of it are not. For business leaders, the relevant question is no longer whether to adopt “generative AI” but why so many adoptions fail—and the answer is rarely the model. It is the absence of a defined human and AI operating pathway: an explicit account of which decisions machines will make, which decisions humans will keep and how each will improve the other. Organizations automating against near-term incentives—cost, headcount, the optics of an AI strategy—are skipping that step. Three decisions, made before deployment, separate the 5% from the rest.
- Decision 1: Redesign the Process Before Automating It.
The economics here are well established. Brynjolfsson, Rock and Syverson showed that general-purpose technologies pay off only after firms make complementary investments—reorganizing workflows, retraining people, redesigning processes—which is why productivity gains lag adoption, sometimes by years. Automating a broken process simply produces a faster broken process. The practical exercise is process archaeology: for each workflow slated for AI, ask which steps exist because they create value and which exist because they always have. In a meaningful share of cases, the highest-return intervention is deleting a step, not automating it. Budget for the redesign, not just the license. - Decision 2: Assign the Division of Labor Explicitly.
The best long-run dataset on human-machine collaboration is three decades of computer chess and its lessons transfer directly. Machines proved superior at tactics, search and tireless calculation; humans stayed comparatively stronger at strategy and judgment—and notably, today’s best human players are the strongest in history because they learned from the machines. Current AI systems show the same asymmetry: Geoffrey Hinton describes their capability as jagged—superhuman on some tasks, unreliable on adjacent ones. The operational translation is a decision-rights map: machines get scale, search and first drafts; humans keep judgment calls, client relationships and accountability for outcomes. Firms that write this down build a capability that compounds. Firms that treat AI as a substitution exercise acquire a cost saving and a fragility at the same time. - Decision 3: Deploy it as Research, Not as Finished Infrastructure.
The builders themselves say the technology is unsettled—Ilya Sutskever describes the field as moving from the age of scaling to the age of research, meaning even the current recipe is provisional. That argues for ordinary engineering discipline applied rigorously: staged rollouts, measurement against predefined baselines and reversibility. If a deployment cannot be measured or unwound, it is not a strategy; it is exposure. This is also the discipline that converts uncertainty from a threat into an option—teams that instrument their deployments learn faster than competitors that simply switch systems on.
To be clear, none of this is an argument for delaying adoption. The developments cited at the start of this article point in the opposite direction. The upside is equally concrete: the AlphaFold line of research is already accelerating drug discovery and similar systems are being pointed at materials and energy. The frontier debate, including Hassabis’s own warnings about technology advancing faster than our understanding of it, will continue, and it matters. But for commercial organizations the decisive action is closer to home. Define the human and AI pathway before the deployment schedule. Redesign before you automate. Write down who decides what. Measure, stage and stay honest about what nobody yet knows. The road ahead is genuinely unknown—but the organizations that take it deliberately, with humans and machines improving each other, will be the ones that end up in the 5%, and amazed at what they achieve there.
