Agent #003 · ABM
Account Fit Scoring Agent: governance & failure modes
What this agent is built not to do, and the specific ways it can fail if misconfigured.
Governance notes
- Weights and tier thresholds change only through a reviewed, versioned config commit — never at runtime, and never for a single pet account.
- The agent never writes to sales-owned fields: deal stage, deal owner, and forecast category are read-only inputs, not outputs.
- Every score stores the config version it was computed under, so a score change is always traceable to a specific, dated config edit.
- Score distribution is checked on every run: if more than roughly 60% of accounts land in a single tier, that’s treated as a miscalibration signal, not a result to trust.
Failure modes
- Flat scoring: if every account lands in the same mid-band tier, the weights are too flat or the thresholds are miscalibrated — caught by the score-distribution check on every run.
- New accounts scored to zero: without the insufficient_data path, brand-new accounts would look like poor fits simply because they haven’t accumulated engagement yet.
- Silent score drift: a score that changes without an explanation is a red flag — every write carries its model/config version so changes are diffable.
- Signal double-counting: readiness and fit sub-scores are configured to draw from mutually exclusive signal sets so a single event doesn’t get counted twice.