Resolve the need, not just the message.
Shipped a GPT-based sales and support assistant with retrieval, memory, and tool skills, measured on resolution, customer satisfaction, and escalation — not demo quality.
- 56%Contacts self-served
- 3.1 → 4.44CSAT
- −34%Escalations
02 / Operating system
Conversational AI resolution loop
- 01Intent
Classify the customer need and urgency.
- 02Retrieve
Ground the response in approved context.
- 03Act
Invoke approved tools or route to a person.
- 04Evaluate
Measure resolution, CSAT, and escalation.
03 / Problem
The constraint behind the outcome.
A production assistant needed to resolve real customer needs without turning a persuasive demo into an uncontrolled support surface.
04 / Decisions
The product choices that shaped the system.
- 01
Separate intent, retrieval, action, and evaluation.
- 02
Treat memory and tools as governed capabilities.
- 03
Optimize for resolution and safe escalation rather than response fluency.
05 / Tradeoffs
The boundaries that made it operable.
- Tighter grounding reduces creative flexibility.
- Human escalation protects quality but limits full automation.
06 / Evidence
What the evidence can—and cannot—support.
AI performance metrics are owner-supplied portfolio outcomes. Architecture references describe the system pattern without disclosing internal domains or implementation details.
- Owner-reported outcomeContacts self-served
A production GenAI assistant self-serves 56% of customer contacts.
- Neel Jaiswal
- Owner-reported outcomeCSAT
Assistant CSAT increased from 3.1 to 4.44.
- Neel Jaiswal
- Owner-reported outcomeEscalations
The assistant reduced escalations 34%.
- Neel Jaiswal
- Owner-reported outcomePublished evidence
The production assistant combines retrieval-augmented generation, memory, and tool skills.
- Neel Jaiswal