Every bookkeeper is being told to "use AI" — but almost no one is being told which jobs are actually safe to hand over. In this session, Tanya Hilts, CPB, FCBP will share three real tasks she gave AI in her own firm: documenting internal processes, drafting client communication, and reviewing financial data. Two worked well. One produced a confident, professional-looking result that was completely wrong — twice, in two different ways. She’ll walk through what happened, why the failure didn't look like a failure, and the one practice that would have caught it every time: validate the input, validate the output. Attendees will leave with a practical framework for deciding which AI tasks belong in their workflow today, and which ones still need a human checking the math. Key Takeaways: A simple framework for sorting AI tasks into "safe to delegate" vs. "still needs review" — regardless of which AI tool you use Why a well-designed process reduces AI risk but doesn't eliminate it, and what actually closes the gap Real examples from an active Canadian bookkeeping practice, not hypotheticals