ASK to Multi-Agent
Employee Intelligence
Built 0→1 inside a 10,000-person org with no mandate, no budget, no prior AI infrastructure. Today it handles 40,000+ queries a month across HR, IT, and Finance to autonomously.
enterprise-wide
vs ₹850 manual
post-resolution
vs 30 to 35 min SLA
88% fully automated
avg 30 to 35 min SLA
avg 20 second resolution
Monthly saving: 60% of 40,000 queries deflected = 24,000 resolved autonomously. (₹850 − ₹17) × 24,000 = ₹2 crore/month in operational cost avoidance. This number is what convinced Finance to invest in the enterprise platform layer.
ASK wave-based roadmap to from RAG foundation to agentic enterprise platform
Praveenkumar
Pain: 3 different portals for 3 different questions. Never knows who to contact first.
Needs: One place to ask anything. Self-service for leave, IT assets, travel policy.
Aishwarya
Pain: Answers the same 20 questions 50 times a week. Can't focus on strategic HR work.
Needs: Repetitive queries automated. Escalation only for genuinely complex issues.
Dimple
Pain: Afraid to bother senior colleagues. Onboarding doc is 200 pages.
Needs: Ask anything without judgment. Instant answers on benefits, IT, code of conduct.
Why this is agentic, not a chatbot: A chatbot retrieves and responds. ASK understands intent, decomposes cross-domain queries, routes to specialised agents in parallel, maintains state across turns, gates on confidence, and escalates with full context to all invisibly to the user.
- User query received → Supervisor agent runs intent classification
- Confidence scored per domain (HR / IT / Finance / Cross-domain)
- Below 0.65 confidence on primary domain → clarification node: one targeted question to resolve ambiguity
- Above 0.65 → route to domain agent(s). Cross-domain queries dispatched in parallel
- Domain agent retrieves from its dedicated vector store (RBAC-filtered per user's department)
- Vertex AI cross-encoder re-ranks retrieved chunks · LLM generates response
- Response delivered · feedback collected · RLHF loop updates document ranking
Why RAG over fine-tuning: HR/IT/Finance knowledge bases update continuously to new policies, org changes, compliance updates. Fine-tuning requires 2 to 3 week retraining + validation per update. RAG updates in hours. Knowledge currency beats marginal accuracy gain from fine-tuning for this use case.
| Trigger | Threshold | Behaviour | Rationale |
|---|---|---|---|
| Low domain confidence | 0.65 | Clarification node to one targeted question to resolve ambiguity first | Better to clarify than confidently answer wrong. False confidence is the worst failure mode. |
| Low confidence after clarification | <0.65 | Human handover with full query + context + agent path exported | Human agent never starts cold to they see everything ASK tried |
| Policy exception request | to | Direct route to HR/IT manager · flagged as exception | Keeps high-judgment decisions with humans. ASK handles volume. |
| PII or complaint detected | to | Immediate human route + flagged in audit log | Risk mitigation. Sensitive queries never auto-resolved. |
Active Directory integration. Engineers see engineering policies. Finance sees finance data. Cross-department knowledge is architecturally separated to no accidental leakage.
Every query logged with: original intent, domain classification, agent path, confidence at each step, response generated, whether human handover was triggered. Full compliance trail.
Standardised agent interface so other teams can build domain agents on ASK infrastructure without owning orchestration. Two additional teams building on the platform.
Deployed entirely inside Microsoft Teams to existing auth, existing habit, zero adoption friction. The biggest barrier to enterprise AI adoption is the new login. Eliminated architecturally.