AI in advertising used to mean creative generation. In 2026, the more interesting work is happening underneath the campaign — in the measurement layer that decides what was reached, what worked, and where to move next.
Audience resolution that doesn't break under fragmentation
Modern probabilistic and graph-based models resolve the same person across linear, OTT, mobile and smart-TV exposure with far higher accuracy than rule-based stitching. The result is a deduplicated reach number you can actually defend.
Outcome modelling that stops at causality, not correlation
ML-driven outcome modelling — when paired with proper experimental design — separates what exposure caused from what would have happened anyway. That's the difference between attribution that comforts and attribution that decides.
Optimisation that respects constraints
Reinforcement-learning-style optimisers now reallocate budget across screens within real-world constraints (minimums, partner commitments, frequency caps) instead of producing theoretically perfect plans nobody can run.
The interesting AI in adtech isn't the model — it's whether the model survives a Monday-morning planning meeting.
What still needs human judgment
Strategy, brand context, business priorities, market signals — these don't come out of a model. The role of AI is to make the evidence layer cheaper, faster and more honest, so the human decisions sit on better ground.