Recommendations moved into Salesforce through API and microservice integration instead of depending only on a standalone surface.
Anonymized enterprise AI case study
AI Seller Workflow Intelligence System
I reframed a dashboard request into a seller-facing workflow intelligence system. Commercial teams needed a faster way to prepare, prioritize, and act with a visible evidence trail.
Shipped enterprise systemAnonymized workflow walkthrough
- Role
- Senior PM
- Users
- 1,600 NA/EU
- Portfolio
- $16B+ fees
- Public proof mode
- Walkthrough
- Decision
- Embed seller guidance in Salesforce and use hybrid batch-and-cache inference.
- Why
- Account teams needed a recommended action with evidence inside their existing workflow.
- Result
- 1,600 NA/EU users, 89% less prep time, and roughly 99% lower inference cost. The public walkthrough does not claim a reproducible latency benchmark.
What this provesStaff PM case
A reporting ask became a workflow intelligence product.
This is my clearest Staff/Principal PM case: a stakeholder request for reporting became a workflow intelligence product with a defined wedge, embedded distribution, deterministic system controls, AI synthesis, measurable adoption, and operating economics.
Hybrid batch/cache inference reduced annual inference cost roughly 99% while keeping the account-team experience fast.
Quick readDecision
Commercial teams needed better decisions, not more reports.
Account teams had dashboards. They lacked a consistent way to turn fragmented seller signals, playbooks, and regional constraints into a confident next action. A reporting surface would have moved the burden of synthesis back to the user.
Separate the stated request from the real workflow problem: preparation burden, evidence discovery, and inconsistent recommendation quality.
Start with call preparation and seller prioritization, then expand from insight consumption into embedded guidance.
AI could synthesize and explain. Deterministic services owned permissions, eligibility, regional routing, freshness, and auditability.
Success needed to show up in workflow adoption, decision speed, recommendation quality, and operating economics.
ArchitectureSystem
A hybrid AI architecture made the strategy shippable.
The system separated deterministic workflow control from AI synthesis so teams could keep the product fast, explainable, and cost-aware.
Frontend
Workflow surface
Seller context, recommended actions, explanation trails, and feedback capture.
Backend
Orchestration
Permissions, seller-state checks, retrieval, cache lookups, inference, and response assembly.
Data
Seller signals
Structured account data combined with curated policy and playbook context.
AI
Selective generation
Retrieval-grounded synthesis while eligibility, ranking, and routing stay deterministic.
TradeoffsJudgment
The important decisions were about where the product should be deterministic, intelligent, or embedded.
A reporting page would still ask account teams to synthesize the answer themselves.
Pre-computation and retrieval kept the experience responsive and cost-aware.
Recommendations needed evidence sufficiency, uncertainty boundaries, and policy grounding.
Adoption depended on appearing where account teams already made decisions.
Anonymized evidenceArtifact
Confidential work still needs visible proof.
I show the shape of the decision without exposing customer data, internal systems, or proprietary screens. The artifact below captures the product move: seller signals become recommended actions only after the system checks policy, eligibility, freshness, and evidence sufficiency.
Signal
Seller context
Account data, playbook intent, regional constraints
Controls
Deterministic gates
Permissions, eligibility, policy, freshness, audit trail
Output
Recommended next action
Evidence, follow-up path, feedback capture
Evidence contractWorkflow walkthrough
The public route explains the workflow without simulating AI.
Inputs select an authored, anonymized workflow state. Run node scripts/sync-portfolio-evidence.mjs to regenerate the published artifact. The route contains no generated output, confidence score, latency benchmark, proprietary data, or provider call, and the manifest makes that boundary inspectable.