Financial Services
Rewrote a $5B AUM wealth manager's CRMA security predicate model in a 3-week sprint
Industry
Wealth management, US-based, $5B AUM
Engagement
Sprint ($5,999)
Headline outcome
11 of 13 predicate-failure patterns fixed; dashboard load time cut from 14s to 4s
The customer ran 47 dashboards across 9 advisor-pod segments. Their security predicate had been authored in 2017 around UserRole.RoleId hierarchy and accumulated CASE-branch complexity over 7 years. By 2024 the predicate had 23 CASE branches and was the top-cited cause of dashboard load latency. Audit identified 11 of our 13 known predicate failure patterns. Sprint scope: rewrite the predicate using territory-driven access plus an indexed join column, simplify the CASE structure to 4 branches, deploy with parallel A/B testing for 5 days, then cut over.
Pattern findings
- 01Inherited-role hierarchy assumption: predicate broke whenever an advisor was promoted
- 02Non-indexed predicate join column: full-table-scan timing under high concurrency
- 03Static roleId allow-list: new advisor pods (added Q3 2023) silently invisible
- 04CASE-based predicate complexity: 23 branches, brittle on every reorg
Healthcare technology
Caught a 92%-accurate Einstein Discovery churn model that was leaking and rebuilt at 71% real
Industry
Healthcare SaaS, US, 800-employee
Engagement
Audit + follow-on Sprint
Duration
5 days audit + 4 weeks rebuild
Headline outcome
Replaced unreliable model with honest 71% accuracy; saved $1.2M in misrouted retention spend over 6 months
A large healthcare-SaaS sales org used Einstein Discovery to predict mid-market customer churn. The model card showed 92% accuracy. Sales operations had been routing retention spend on its predictions for 14 months. Our audit found three target leakages: Account.Id was a feature (high-cardinality categorical leakage), Stage_Change_Date was included (date-feature trap with post-prediction information), and 73% of records had Renewal_Risk_Manual_Score populated by the customer success team retroactively after the renewal event. The rebuilt model excluded all three, scored honestly at 71%, and the resulting retention spend reallocation surfaced $1.2M in budget that had been chasing false positives.
Pattern findings
- 01High-cardinality categorical leakage (Account.Id directly as feature)
- 02Date-feature trap (Stage_Change_Date populated post-prediction event)
- 03Retroactive feature population (CS team scored renewal risk after the renewal)
- 04Discovery silently dropped 4 features above 70% null density without flagging
Manufacturing
Saved a global industrial customer 11 hours/day of nightly Recipe runtime by reverting to dataflow
Industry
Industrial manufacturing, global, 30,000-employee
Engagement
Sprint ($5,999)
Headline outcome
Cut nightly data prep from 14h to 3h; eliminated 11h of org compute resource pressure
The customer migrated from Wave dataflow JSON to CRMA Recipes in 2022 because Salesforce's docs recommended it. By 2024 their nightly data prep was running 14 hours, regularly missing the 6 AM dashboard refresh deadline. Audit found three datasets where Recipe was the wrong tool: a 90M-row Account dataset where the Recipe was full-refreshing nightly because the smart-transform incremental detection was misfiring, a multi-source revenue rollup where append + augment via dataflow JSON was 4x faster than Recipe equivalents, and a window-function-heavy ARR-by-cohort dataset where Recipe COMPUTE_RELATIVE was inefficient. Sprint scope: revert these three datasets to dataflow JSON, document the decision matrix, train the in-house team.
Pattern findings
- 01Smart-transform incremental detection misfiring on rename/recreate cycles
- 02Recipe append + augment slower than dataflow JSON for multi-source aggregation
- 03Recipe COMPUTE_RELATIVE inefficient for high-cardinality window functions
- 04Customer team had no decision matrix for when Recipe wins vs Dataflow wins
Consumer packaged goods
Anthropic-readiness audit on a 2017 Wave implementation surfaced 8 of 10 datasets needing rebuild
Industry
CPG, multinational, $40B revenue
Engagement
Audit ($999) + planning consultation
Headline outcome
Tier-1 dataset readiness scored at 18/100. Six-month rebuild roadmap delivered with phased priority.
The customer launched Wave Analytics in 2017 to power their distributor sales analytics across 40 markets. Implementation accumulated layers across 7 years. Following the October 2025 Salesforce-Anthropic partnership announcement, the customer wanted to know whether their existing CRMA could be exposed to Claude via Agentforce 360. Audit scored each of their 10 priority datasets against our 10-dimension compatibility rubric. Result: 8 of 10 datasets scored below 50/100 (needing significant re-architecture), 2 scored 60-69 (re-architect 1-2 dimensions). Top failures: high-cardinality Distributor_Code categorical, Wave-era security predicates that confused multi-step Claude reasoning across regional boundaries, auto-exposed Recipe schemas that broke on column rename. Audit deliverable: 12-page readiness report plus 6-month phased rebuild roadmap with engineering effort estimates.
Pattern findings
- 01High-cardinality categoricals (Distributor_Code, Account.Id) leaking target signal
- 02Wave-era security predicates failing under multi-step agent reasoning
- 03Auto-exposed Recipe schemas: 5 of 10 datasets broke when columns renamed
- 04Trust-boundary classification posture undefined: Claude couldn't read most datasets
- 05Recipe vintage: 6 of 10 datasets still on Wave dataflow JSON from 2017