Adapting agile and product-discovery practices to AI delivery: where outputs are probabilistic, evals are part of the definition of done, and shipping cadence i
Agile for AI teams adapts iterative delivery to probabilistic work: experiment-shaped backlog items with kill criteria, eval thresholds as definition-of-done, model-dependency planning, and ceremonies that handle 'it works 87% of the time' honestly.
Quiet enabler: teams that adapt delivery practice ship AI consistently while cargo-cult agile stalls, delivery leads fluent in both idioms are valued across enterprise AI programs.
Adapted, yes: iteration suits AI's experimental nature, but stories need eval-based acceptance, kill criteria, and tolerance for negative results as legitimate outcomes. Unadapted scrum theater fits badly.
Time-boxed experiments with decision exits: the commitment is the learning and the decision, not a feature. Two-week spikes with explicit kill/continue gates keep uncertainty inside the cadence.