You should. It's brilliant.

A general chatbot answers from memory, and memory ages. The numbers that decide your answer — tax brackets, contribution limits, Medicare premiums — reset every January, while a model's training data always runs months behind. So we keep a Facts Layer: more than 80 annually-indexed figures across 14 domains, each traced to a primary source, injected at the top of every advisor's instructions so the current number beats the remembered one.

Answers here don't arrive as anonymous paragraphs. They name their sources — the actual authority, like an IRS revenue procedure or a CMS notice — and carry a stated confidence level. When the advisor is exercising judgment rather than citing a rule, the answer says so instead of dressing opinion up as fact.

One generalist model holds one conversation. WithDave is 5 advisors covering 30 topics — business and money, legal, life planning, government benefits, insurance and risk — each topic with its own hand-built instructions, its own verified figures, and its own sense of what a professional in that field would flag.

A professional's first instinct isn't to answer — it's to ask. Our advisors are instructed to do the same: before committing to numbers that depend on your filing status, your state, or your timeline, they ask the follow-up questions a professional would ask. A generic chat will happily give you a confident answer to an under-specified question. Ours narrows the question first.

A chat answer is a dead end — you read it, you close the tab. Here the answer is the start: it becomes action items you can track, documents you can generate (167 types, from checklists to letters), morning briefings that watch the topics you care about, and a dashboard that remembers — so when the rules change in January, you find out from us, not from a penalty notice.

Before every milestone, model change, or prompt edit, we run an adversarial exam: 135 golden questions across 21 topic batteries, built specifically to bait the stale figures and confident half-answers general models produce — and graded through the real product path, not a lab mock. When the exam catches us, we fix the layer that failed, not just the answer.