The honest version, from someone who trades manually and uses software to support it rather than replace it.
Where it genuinely helps
- Screening and ranking. Reducing thousands of names to a short list.
- Summarising. Concalls, filings and long documents.
- Bookkeeping. Journalling, tagging, computing what your rules would have done.
Where it quietly hurts
- Backtest-shaped confidence. It is trivially easy to find a rule that worked on the past. Out-of-sample discipline is the whole game.
- Non-stationarity. Markets change regime. A model trained on one regime is confidently wrong in the next.
- Signal laundering. A model output feels objective. It is still a hypothesis with a number attached.
Your own take goes here — what you actually use, and one thing that didn’t work.