Feature pitched as if every customer has rich history
New accounts have sparse records, a handful of events, or fragmented data across systems. Without a cold-start plan, Day 1 output is garbage regardless of model quality.
Failure mode
Feature gets abandoned in the first week before data accumulates
The data exists but rights are not verified
Terms, DPAs, enterprise procurement requirements, residency constraints, and provider retention policies can all change the design. Not knowing is not a low-severity footnote.
Failure mode
Enterprise deal blocked by legal or security review post-launch
Activity metrics substituted for output quality
Clicks, impressions, and session duration are not quality signals. Without task-specific evals and correction loops, the team cannot tell whether the AI is useful or just used.
Failure mode
Feature ships with no way to measure whether output is correct
Observability treated as operations work for later
Without failure logging, the team cannot determine whether the problem is the model, the data slice, the prompt, or the customer segment. Silent failure looks like adoption data.
Failure mode
Poor segments underperform for months before anyone identifies why