Using the abbreviated JS/DS screening flow as a practical example of changing the workflow around AI rather than simply inserting AI into an existing process. Problem: the original process was thorough but too expensive for testing new search directions. Change: use a faster screening path for batches, then spend deep DS effort only on the …
How to place Human-in-the-Loop checkpoints based on risk, consequence, ambiguity, and reversibility. Problem: automation is attractive, but not every decision should be automated. Low-risk work can be automated; high-risk or consequential decisions get HITL checkpoints. Broader lesson: design human review into the workflow intentionally.
Why finding evidence is not enough, and why interpretation, provenance, and context still matter. Problem: finding evidence is not the same as interpreting what it proves. Covers governance + knowledge quality + “what does the data actually say?” Broader lesson: better retrieval does not fix poor interpretation.
The problem with fragmented AI tools, what the system needed to do, and how it helps one person work with more scale, consistency, and control. Problem: AI tools were useful, but fragmented. Need: shared knowledge, consistent rules, repeatable decisions, evidence tracking, feedback loops. Value: the system lets one person do more, more consistently, with less …