Too Many AI Workers Can Actually Make Tax Mistakes Worse

When companies sell goods across borders, they must determine the correct VAT (a sales tax) for each transaction. This is a perfect job for AI, but there's a design question: should you use one versatile AI "agent" (a bot that can think and use tools) or split the job among many specialized agents, like a team of workers? This new, open-source study tested exactly that.

Researchers ran over 4,400 simulated tax cases, comparing one AI worker against teams of two, three, four, and five specialized AI agents. They found that the middle ground — around two to four agents — got the best accuracy (83% correct), beating both the single worker (72%) and the five-agent team (77%). But the results didn't meet their pre-defined success threshold, so we can't yet say the "sweet spot" is real. The single agent wasn't clearly worse, either.

The most striking finding: when the researchers deliberately injected a single "hallucinated" record — a plausible-sounding but made-up piece of data — every configuration's accuracy dropped dramatically. Teams with many narrow agents were hurt worst. This suggests that splitting work into too many small pieces makes the system more fragile, not more reliable.

For anyone who will rely on AI for taxes, accounting, or other precise tasks, this is a cautionary tale. Breaking a job into many AI helpers isn't automatically better, and even one small AI error can cascade. The good news: the researcher shared all code, data, and raw traces, so others can continue testing. The practical takeaway? For critical tasks like taxes, use AI carefully, verify outputs, and don't assume "more agents" means "better results."