System Architecture · v2

Customer Engagement
Single Timeline Workflow

One timeline. Business outcome layer, agentic intelligence, client-defined skills, and governance checks activate at specific moments in sequence. Read top to bottom: this is the life of a single customer interaction.

Business
Agentic Intelligence
Governance
Quality Gate (fail-closed)
Human Approval (conditional)
Skills Engine

Incoming Customer Message

A customer reaches out via chat or email. This single event starts the timeline for this interaction.

↓
Capture Intent & Load Context
Pull in the customer's full conversation history and identify the core ask, so every subsequent step starts with complete context.
Business
↓
Secure the Record & Lock Concurrency
Log the inbound message immediately — no message is ever lost, even under high load. A concurrency lock prevents overlapping duplicate responses for the same conversation.
Governance
↓
⛔ Quality Gate A — Inbound Safety Screen
Before any AI work begins, the message is evaluated for policy compliance and harmful content. This gate is fail-closed.
✓ passes → AI engagement begins ✗ blocked → pipeline halts, safe fallback sent
↓
Sense — Interpret & Assess
The AI reads the message in full context: sentiment, urgency, intent category, and relevant business knowledge. This assessment shapes the entire response strategy and determines whether action-intent is present — the signal that activates the Skills Engine.
Agentic
↓
↺ Agentic Reasoning Loop — iterates until confident (max 5 cycles)
1
Reason
Form a response hypothesis. Identify what information or actions are still missing to answer with confidence.
2
Act — with Skills
Execute business actions: look up order status, retrieve customer records, apply account tools, pull knowledge — and invoke client-defined Skills when action-intent is confirmed.
⚙
Skills Engine
activates on action-intent only
Skills Lookup Agent
Runs semantic search over the client's skill registry. Returns a ranked shortlist of relevant skills with their API contracts to the Decider.
Decider Agent
Evaluates the shortlist against current context. Selects the correct skill, validates the risk tier, and triggers invocation or escalation.
intent confirmed
→
semantic lookup
→
skill selected
→
risk tier check
→
API invoked
Skill risk tiers
● low — auto-execute
● medium — confirm with customer
● high → gate C / human queue
⛔
Quality Gate C — Pre-Action Guard: Skills with side effects (refunds, reprints, cancellations, account changes) pause before API execution. High-risk skills route to the Human Approval Queue. This gate sits inside the loop, before any irreversible action is taken.
Example client-defined skills
issue reprint check order status apply discount code cancel subscription process refund update delivery address escalate to team
Skill response contract: every skill API response is normalised to a standard envelope {status, data, summary} before re-entering the reasoning loop — preventing inconsistent API outputs from degrading agent reasoning.
3
Evaluate
Is the answer complete, confident, and policy-safe? Were all required skills executed successfully? If not, the loop returns to Reason with new evidence.
4
Adapt
Revise the strategy and response wording based on what the loop has learned, including skill execution outcomes. Prepare the final draft.
↺ If Evaluate finds gaps or low confidence, the loop returns to Reason with new evidence — up to 5 cycles. Skills are only invoked when action-intent is confirmed at Sense; pure information queries skip the Skills Engine entirely.
↓
⛔ Quality Gate B — Outbound Compliance Screen
Final check on the drafted response: policy alignment, tone quality, and relevance to the customer's latest message. The system never auto-sends an unreviewed response past this gate.
✓ passes → routes to delivery ✗ blocked → held in human approval queue, no auto-send
↓
Commit Response & Route to Owner
Finalise the response and determine delivery path — automatic send for standard cases, or assignment to the agent queue based on risk level and channel type.
Business + AI
↓
Conditional path — escalation, exception, high-risk skill, or above risk threshold
👤 Human Approval Queue
Complaints, sensitive cases, policy exceptions, and high-risk skill invocations are held for a human agent to review and approve before any response or action is executed.
↓
Deliver Response to Customer
Send the approved response via the appropriate channel — chat widget, email reply, or other configured channel.
Business
↓
Advance Lifecycle State & Audit Log
Move the customer's lifecycle stage forward: Awareness → Intent → Commitment → Retention → Advocacy. Every decision, routing choice, skill invocation, and AI action is logged for QA, compliance, and future personalisation.
AI + Governance
Interaction complete. Record updated, skills log written, outcome logged. The next interaction begins with richer context than the last.
Speed KPI
First Response Time (FRT)
Quality KPI
Resolution Rate / Re-contact Rate
Risk KPI
Escalation Accuracy + Policy Incidents
Experience KPI
CSAT / Sentiment Trend
Action KPI
Skill Invocation Rate + Success Rate