What it reads
- Aggregate behavior: retention, funnels, where users drop off, which early actions go with users sticking around
- Every active flow’s results, compared against the holdout
- Your app context (below)
- Your own database, read-only, if you’ve connected one
What it proposes
Each proposal is one change: create a flow, adjust one, pause one, or retire one. Removing rules that stopped working matters as much as adding new ones. Every proposal carries:- Diagnosis: what it noticed, in plain words
- Evidence: the numbers behind it
- Confidence:
high,medium,loworinsufficient_data - The complete flow it would create or change, already checked against the same rules as a hand-written one
Users who add 3+ connections in week 1 retain at 4.2x. 61% never reach 3;
median drop-off is at connection #1 with no second within 48h.
Proposed: trigger on connection_added with no second one in 48h, send at
most once a week.
It is built to say “insufficient data” when that is the truth. With modest
traffic, most differences between flows aren’t significant yet, and a
confident story invented from noise would be worse than no proposal.