Marketing ops · Production

Skai builds Celeste with Amazon Bedrock Agents to reduce report generation time by 50–90%

The problem

Skai's customers were spending up to 1.5 days per week manually preparing static reports, could not intuitively query complex advertising datasets, and lacked dynamic visualization tools — critical business questions remained hidden in siloed data that required technical expertise to access.

Workflow diagram · grounded in source
1
User submits natural language query
trigger
“Customer Experience UI (CX UI) – The frontend interface that users interact with to submit questions and view responses”
2
Bedrock Agent orchestrates tool invocation
ai_action
“Amazon Bedrock agent – An orchestrator that receives queries from Chat Executor, determines which tools to invoke based on the query, and manages the tool invocation process”
3
Tool API queries customer data
integration
“Tool API – A custom API that receives tool invocation requests from the Amazon Bedrock agent and queries the customer data”
4
RAG-based memory retrieval
ai_action
“Our solution uses dynamic session chunking to split conversations while retaining key context, and employs Retrieval Augmented Generation (RAG)-based memory retrieval”
5
Claude 3.5 Sonnet V2 generates response
ai_action
“Anthropic's Claude 3.5 Sonnet V2 – The FM that generates natural language responses. The model generates queries for the API and processes the structured data returned by tools. It creates coherent, contextual answers for users.”
6
Insights and recommendations output
output
“the assistant generates comprehensive insights and case studies while providing actionable recommendations on campaign activity, complete with detailed analytical approaches and ready-to-present stakeholder materials”
Reported outcome

Deploying Celeste on Amazon Bedrock Agents delivered 50% faster report generation, 75% faster case study creation, 80% faster QBR composition, and 90% faster report-to-recommendation time, turning workflows that once took days or weeks into processes completed in minutes.

Reported metrics
Report generation time50% Faster
Case study generation time75% Faster
QBR composition time80% Faster
Report to recommendation time90% Faster
Show all 11 reported metrics
Report generation time50% Faster
Case study generation time75% Faster
QBR composition time80% Faster
Report to recommendation time90% Faster
Median response latency reduction30%
Average response latency (baseline)136
Average response latency (optimized)44
POC to production timeline reduction50%
Uptime during customer demonstrations99.9%
Weekly report preparation time (baseline)1.5 days a week
Hours saved per usersaving hours of time
Reported stack
Amazon Bedrock AgentsAmazon BedrockCelesteAnthropic's Claude 3.5 Sonnet V2Amazon CloudWatchAmazon CloudWatch Logs InsightsIAMAmazon NovaRAGWatchDogMeta's Llama
Source
https://aws.amazon.com/blogs/machine-learning/skai-uses-amazon-bedrock-agents-to-significantly-improve-customer-insights-by-revolutionized-data-access-and-analysis?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Deploying Celeste on Amazon Bedrock Agents delivered 50% faster report generation, 75% faster case study creation, 80% faster QBR composition, and 90% faster report-to-recommendation time, turning workflows that once…

What tools did this team use?

Amazon Bedrock Agents, Amazon Bedrock, Celeste, Anthropic's Claude 3.5 Sonnet V2, Amazon CloudWatch, Amazon CloudWatch Logs Insights, IAM, Amazon Nova, RAG, WatchDog.

What results were reported?

Report generation time: 50% Faster; Case study generation time: 75% Faster; QBR composition time: 80% Faster; Report to recommendation time: 90% Faster (source-reported, not independently verified).

How is this marketing ops AI workflow structured?

User submits natural language query → Bedrock Agent orchestrates tool invocation → Tool API queries customer data → RAG-based memory retrieval → Claude 3.5 Sonnet V2 generates response → Insights and recommendations output.

Related marketing ops cases