Every Licensed Professional Becomes an Insurer — My AI Just Passed the Singapore CPA Exam 10 for 10

The creator built an AI agent and ran it against past papers for the Singapore CPA exam (SCAQ). It passed 10 times out of 10. He had originally planned to talk about the New Zealand CPA exam, which his AI also passed, but the source materials there sit in a grey zone — copyright and exam-board terms get murky the moment you describe where the questions came from. Singapore's public exam body publishes past papers and answer keys as PDFs, so this whole video can be discussed cleanly.

SCAQ is no toy exam. Four modules — financial reporting, audit, tax, and governance — taken online under webcam proctoring. Not the kind of test you cheese with question-guessing. That's the workload the AI was put against.

The AI's edge comes from three layered tricks. First, hybrid retrieval — vector search and keyword search run in parallel, results merged. Clause numbers and legal terms favour keyword search, which pure embeddings can miss. Second, HyDE: instead of searching with the question, ask the AI to draft a hypothetical answer first, then search with that. In the source material, the answer text sits closer to the real answer than the question text does.

Third, reranking. Pull a wide 30 candidates, then use a separate specialist model to filter down to the actual 5 needed. Before these three layers landed, average score was 52 — just over the 50 pass line. After: low 70s. The system moved from barely-passing to comfortably-passing in one upgrade. The AI finishes in under an hour where humans take three. Its failure modes are oddly human — 5–10 points off on image and table questions (multimodal isn't there yet), occasional confusion on Singapore's 8→9% GST change, audit weaker than financial reporting. Same skew real candidates show.

The headline isn't 'AI became a CPA.' The AI can't sign a tax return or take liability for advice given to a real client. The author's actual claim is sharper, and structural.

What clients pay licensed professionals for shifts. Today they pay for research, analysis, and a deliverable. Tomorrow they're paying for liability absorption — the professional's malpractice insurance is what's really being purchased. Today professional indemnity insurance is a side item, with the main bill being the professional service fee. That gets inverted. Premium becomes the main line item. The professional sits between an AI doing the research and a human doing the sign-off, taking a premium to absorb the risk. Once that flips, new insurance products and new liability designs follow. The author's prediction: licensed professionals who embrace AI scale 10× on case volume and run on premium-based pricing. The holdouts keep billing hourly and get squeezed. That's the 10–20 year view he says became unmistakable from this one experiment.