Building a Multi-Agent AI Customer Support Assistant on AWS with Strands Agents SDK and Amazon Bedrock

14:59Claude responded: A five-agent customer support pipeline on AWS that automates intent classification, semantic KB retrieval, AI response generation with auto-actions, sentiment-…A five-agent customer support pipeline on AWS that automates intent classification, semantic KB retrieval, AI response generation with auto-actions, sentiment-triggered human escalation via Amazon Connect, and post-resolution CSAT follow-up.

sweta jha

Published Jun 26, 2026

Last modified Jun 26, 2026

Building a Multi-Agent AI Customer Support Assistant on AWS with Strands Agents SDK and Amazon Bedrock

GitHub: https://github.com/its-me-sweta/Customer_Supprt_AI_Assistant.git

Introduction

Customer support is one of the highest-volume, highest-stakes workflows in any product company. Every ticket carries a customer's frustration, expectation, or urgency — and the speed and quality of the response directly drives retention and CSAT scores.

In this post I'll walk through a production-ready, five-agent AI pipeline that handles the full customer support lifecycle — from the moment a ticket arrives to the CSAT survey and profile update — using:

Strands Agents SDK with @tool decorators
Amazon Bedrock Nova Pro for AI response generation
Amazon Comprehend for sentiment analysis and NER
FAISS + Amazon Titan Embed v2 for semantic KB search
Amazon OpenSearch Serverless for full-text KB fallback
Amazon Connect for live human agent handoff
Amazon DynamoDB for pipeline state and customer 360 profiles
Amazon SES + SNS for resolution emails and alerts
Amazon EventBridge Scheduler for proactive follow-ups
AgentCore Runtime for serverless deployment

Architecture Overview

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
                   ┌───────────────────────────────────┐
                   │     FastAPI / REST Layer           │
                   │   POST /support/ticket             │
                   └─────────────────┬─────────────────┘
                                     │
                   ┌─────────────────▼─────────────────┐
                   │       Orchestrator Agent           │
                   │  Triage · Route · Audit · State    │
                   └──┬──────────────────────────┬─────┘
                      │                          │
         ┌────────────▼────┐         ┌───────────▼──────────┐
         │ Intent &        │         │  Knowledge Retrieval  │
         │ Sentiment Agent │         │  Agent                │
         │ Comprehend      │         │  FAISS + Titan Embed  │
         └────────────┬────┘         └───────────┬──────────┘
                      └──────────┬───────────────┘
                                 │
                   ┌─────────────▼───────────────────┐
                   │       Resolution Agent           │
                   │  Nova Pro · Auto-actions · HITL  │
                   └─────────────┬───────────────────┘
                                 │
              ┌──────────────────┼──────────────────┐
              │                  │                  │
   ┌──────────▼──────┐  ┌────────▼───────┐  ┌──────▼──────────────┐
   │ Escalation      │  │ Auto-resolved  │  │ VIP / P1 Fast-track │
   │ Agent           │  │                │  │                     │
   │ Connect · SNS   │  │                │  │                     │
   └──────────┬──────┘  └────────────────┘  └─────────────────────┘
              │
   ┌──────────▼──────────────────────────────────────────┐
   │              Follow-up Agent                        │
   │  SES · CSAT Survey · Scheduler · Customer 360       │
   └─────────────────────────────────────────────────────┘

The Five Agents

Agent	Role	AWS Services
Orchestrator Agent	Triages, routes, and maintains state	DynamoDB, SNS
Intent & Sentiment Agent	Classifies intent, detects sentiment	Comprehend, DynamoDB
Knowledge Retrieval Agent	Semantic KB search and article ranking	Titan Embed v2, OpenSearch, FAISS
Resolution Agent	Generates responses + executes actions	Bedrock Nova Pro, SES, DynamoDB
Escalation Agent	Routes to human agents with SLA timers	Connect, SNS, DynamoDB
Follow-up Agent	CSAT surveys, proactive check-ins, profiles	SES, Scheduler, DynamoDB

Pipeline Deep Dive

Stage 1 — Intent & Sentiment Agent

The first thing the pipeline does is understand the customer — not just what they're asking, but how they're feeling.

Sentiment via Amazon Comprehend:

1
2
3
4
5
6
7
8
9
10
11
12
@tool
def detect_sentiment(text: str, language_code: str = "en") -> dict:
    """Analyzes sentiment using Amazon Comprehend."""
    response = comprehend.detect_sentiment(Text=text[:5000], LanguageCode=language_code)
    scores = response["SentimentScore"]
    # Compute scalar: Positive pulls toward +1, Negative toward -1
    scalar = scores["Positive"] - scores["Negative"]
    return {
        "sentiment_label": response["Sentiment"],
        "sentiment_score": round(scalar, 4),
        "escalation_flag": scalar <= -0.6,   # immediate escalation threshold
    }

Intent classification uses keyword scoring across 8 categories: billing, technical, account, shipping, feature_request, complaint, compliment, legal.

VIP check queries DynamoDB for enterprise tier or LTV ≥ $10,000 customers, giving them priority routing throughout the pipeline.

Stage 2 — Knowledge Retrieval Agent

The Knowledge Agent uses a two-tier RAG architecture:

Tier 1 — Dense retrieval (primary):

1
2
3
4
5
6
7
8
9
10
embedding = get_embedding(query)   # Amazon Titan Embed v2 → 1024-dim vector
# kNN search on OpenSearch Serverless index
search_body = {
    "query": {
        "bool": {
            "must": [{"knn": {"content_vector": {"vector": embedding, "k": 3}}}],
            "filter": [{"term": {"intent_category": intent}}],
        }
    }
}

Tier 2 — BM25 keyword fallback:
If the top dense retrieval score is < 0.65, the agent falls back to DynamoDB scan for keyword matching — ensuring coverage even for edge-case queries.

Similar ticket lookup searches resolved historical tickets for the same intent + error code combination, giving the agent real resolution patterns to draw from.

Every KB hit is logged to support-kb-analytics for continuous improvement — so articles that consistently fail to resolve tickets get flagged for review.

Stage 3 — Resolution Agent

The Resolution Agent is the heart of the pipeline. It decides whether to auto-resolve, act, or escalate.

Confidence scoring:

1
2
3
4
5
6
7
8
9
10
11
@tool
def check_resolution_confidence(kb_score, intent_confidence, is_vip, priority) -> dict:
    # VIP and P1/P2 tickets need 80% confidence for auto-resolve (vs 65% standard)
    threshold = 0.80 if (is_vip or priority in ("P1","P2")) else 0.65
    composite = (kb_score * 0.60) + (intent_confidence * 0.40)
    return {
        "resolution_confidence": composite,
        "recommended_action": "auto_resolve" if composite >= threshold
                              else "resolve_with_followup_offer" if composite >= 0.50
                              else "escalate_to_human"
    }

Automated actions the agent can execute without human involvement:

Action	Auto-limit	Tool
Refund	≤ $500	`execute_refund`
Courtesy credit	≤ $50	`apply_account_credit`
Password reset	Always safe	`reset_customer_password`
Account unlock	If no fraud flag	`unlock_account`
Order status update	Always safe	`update_order_status`

Response generation uses Bedrock Nova Pro directly, with tone calibrated to the customer's sentiment:

1
tone = "warm and reassuring" if sentiment_label == "NEGATIVE" else "friendly and helpful"

Stage 4 — Escalation Agent

The Escalation Agent handles the human handoff with precision. It doesn't just "route to a human" — it selects the right queue, finds the right agent, sets an SLA timer, and generates a briefing.

Queue routing:

1
2
3
4
5
6
7
QUEUE_MAP = {
    "billing":  "billing_specialists",
    "legal":    "legal_compliance",
    "technical":"technical_tier2",
    "account":  "account_security",
}
# VIP customers always route to "vip_support" regardless of intent

SLA timers:

Priority	SLA
P1	15 minutes
P2	2 hours
P3	8 hours
P4	24 hours

Case summary (5-bullet handoff note generated for the human agent):

1
2
3
4
5
6
7
CASE SUMMARY — TKT-A1B2C3D4 [VIP CUSTOMER — PRIORITISE]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Customer:     Jane Smith (VIP)
Issue type:   Billing
Frustration:  High (score: -0.72)
AI attempted: Yes — no resolution achieved
Recommended:  Apologise first, then investigate account directly.

Live chat escalations transfer directly to Amazon Connect for seamless agent takeover.

Stage 5 — Follow-up Agent

The Follow-up Agent closes the loop on every ticket — whether AI-resolved or human-escalated.

CSAT survey timing:

Chat tickets: sent immediately after resolution
Email tickets: scheduled 30 minutes post-resolution via EventBridge Scheduler
P1 tickets: scheduled 24 hours later with a personal outreach note

1
2
3
4
5
scheduler.create_schedule(
    Name=f"csat-{survey_id}",
    ScheduleExpression=f"at({send_at.strftime('%Y-%m-%dT%H:%M:%S')})",
    Target={"Arn": "arn:aws:lambda:...:function:send-csat-email", ...},
)

Customer 360 profile is updated in DynamoDB on every interaction:

Last interaction date and channel
Issue history and resolution type
Sentiment trend over time

KB feedback loop — if the resolution confidence was < 0.80, the KB article is flagged for review. This creates a self-improving knowledge base over time.

Escalation Decision Tree

1
2
3
4
5
6
7
8
9
10
11
Incoming ticket
    │
    ├─ Sentiment <= -0.6 OR intent=legal/data_breach → IMMEDIATE ESCALATION (P1/P2)
    │
    ├─ is_vip = True → VIP fast-track queue, confidence threshold = 0.80
    │
    ├─ KB confidence >= threshold → Auto-resolve + follow-up
    │
    ├─ KB confidence 0.50–0.65 → Resolve with "reply if unresolved" offer
    │
    └─ KB confidence < 0.50 after 2 attempts → Human escalation

AWS Infrastructure Summary

Resource	Purpose
Amazon Bedrock Nova Pro	AI response generation
Amazon Comprehend	Sentiment analysis, NER, language detection
Amazon Titan Embed v2	Semantic embeddings for KB search
OpenSearch Serverless	Dense + BM25 KB retrieval
DynamoDB (6 tables)	Tickets, customers, KB, cases, agents, analytics
Amazon Connect	Live chat/call human agent handoff
Amazon SES	Resolution emails, CSAT surveys
Amazon SNS	Escalation alerts, supervisor notifications
EventBridge Scheduler	CSAT surveys, proactive follow-ups
AgentCore Runtime	Serverless agent deployment

Key Design Patterns

Dual-threshold confidence — VIP and P1/P2 tickets need 80% confidence for auto-resolve vs 65% for standard tickets. Never compromise resolution quality for high-value customers.

Auto-action limits — Hard limits on automated refunds ($500) and credits ($50) prevent runaway automation while keeping routine resolutions instant.

KB feedback loop — Every resolution outcome feeds back into the KB analytics. Articles that fail to resolve tickets get flagged automatically.

Sentiment-calibrated tone — The Resolution Agent adjusts its communication style based on the detected sentiment score, not just the intent category.

Tiered SLA with escalation — SLA timers start the moment a ticket escalates. Supervisors are notified for P1 and VIP tickets automatically.

Try It Yourself

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
git clone https://github.com/its-me-sweta/Customer_Supprt_AI_Assistant.git
cd customer-support-assistant
pip install -r requirements.txt
uvicorn main:app --reload

# Submit a support ticket
curl -X POST http://localhost:8000/support/ticket \
  -H "Content-Type: application/json" \
  -d '{
    "customer_id": "CUST-001",
    "customer_name": "Jane Smith",
    "customer_email": "jane@example.com",
    "channel": "email",
    "message": "I was charged twice for my subscription this month. This is unacceptable.",
    "product": "Premium Plan"
  }'

Built with Strands Agents SDK · Amazon Bedrock Nova Pro · AgentCore Runtime

# agentic-ai # agents # aws-community-builders

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.

Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article

Builder Center

Building a Multi-Agent AI Customer Support Assistant on AWS with Strands Agents SDK and Amazon Bedrock

sweta jha

Building a Multi-Agent AI Customer Support Assistant on AWS with Strands Agents SDK and Amazon Bedrock

Introduction

Architecture Overview

The Five Agents

Pipeline Deep Dive

Stage 1 — Intent & Sentiment Agent

Stage 2 — Knowledge Retrieval Agent

Stage 3 — Resolution Agent

Stage 4 — Escalation Agent

Stage 5 — Follow-up Agent

Escalation Decision Tree

AWS Infrastructure Summary

Key Design Patterns

Try It Yourself