Just came across a very interesting simulation study: What happens if you make AI agents simulate the behavior of typical shoppers – and then test marketing campaigns on them before launching anything in the real world? Researchers built a tiny virtual town and populated it with AI agents. Just 11 agents and 10 locations. Each agent had: • a budget • an energy level • food and shopping needs • memory • social connections • its own persona Then DeepSeek-V3 generated their plans, conversations, and purchase decisions. So the agents went to work, ate, talked to each other, remembered things, invited friends somewhere – and independently decided where to spend money. The actual experiment was very simple: One fried chicken shop offered a 20% discount in the middle of the week. The researchers then watched what happened. • The shop’s revenue increased by 51% from Day 2 to Day 3 – even after (!) the discount. • Its market share grew from 30% on Day 1 to 41% on Day 3. • But the TOTAL market didn’t really grow. So the discount didn’t create much new demand – it mostly pulled customers away from a competitor. • The local diner lost market share and saw revenue fall by 7%. • The coffee shop stayed relatively stable. Even more interestingly, the strongest effect didn’t happen immediately. The authors interpret that as information spreading through the agents’ social network: they talked, made plans together, invited friends – and interest in the place increased through their interactions. The simulation also produced something that looked like loyalty: some agents returned to the fried chicken shop during the promotion, while the diner kept broader appeal even without discounts. So in theory, systems like this could let marketers test: • discounts • promotions • price changes • new product launches • competitive scenarios on synthetic consumers before going into the real market. But there’s a very important BUT. This is still a proof of concept, not evidence that AI can reliably predict real human behavior. The simulation only had 11 agents. Their decisions depend heavily on prompt design, the chosen LLM, the rules of the environment, and even how the virtual town is structured. Imo, this is definitely something to watch. Because if these simulations eventually get calibrated properly on real-world data, part of CustDev, pricing experiments, and campaign testing may happen on AI agents first – and only then on humans.