Customer support · Production

Boti: Agentic AI assistant on Amazon Bedrock transforms citizen access to Buenos Aires government procedures

The problem

Citizens of Buenos Aires faced difficulty navigating the city's complex bureaucratic landscape of over 1,300 government procedures, each with its own logic, nuances, and exceptions, making it hard to find the right procedure quickly.

First attempt

Standard retrieval-augmented generation approaches struggled to disambiguate similar government procedures, and the mixture of chunks they returned increased the likelihood of generating incorrect responses.

Workflow diagram · grounded in source
1
Citizen submits question
trigger
“The process begins when a user submits a question. In parallel, the question is passed to the input guardrail system and government procedures agent.”
2
Input guardrail classifies query
validation
“The input guardrail uses a custom LLM classifier to analyze incoming user queries, determining whether to approve or block requests based on their content”
3
Route approved or blocked query
routing
“If triggered, it stops graph execution and redirects the user to ask questions about government procedures. Otherwise, the agent continues to formulate its response.”
4
Reasoning retriever fetches procedures
ai_action
“An LLM follows a chain-of-thought (CoT) process in which it compares the user query to the retrieved summaries. It discards irrelevant procedures and reorders the remaining ones based on relevance.”
5
Sentiment-based prompt routing
routing
“we incorporated sentiment analysis into our knowledge base as metadata. This allows our system to route to different prompt templates. Sensitive topics are directed to prompts with reduced emoji usage and more empathetic language, wherea…”
6
Generate response in Rioplatense Spanish
output
“the agent responds using Boti's distinctive communication style: concise, helpful messages in Rioplatense Spanish”
Reported outcome

The agentic system achieves up to 98.9% top-1 retrieval accuracy, a 12.5–17.5% improvement over standard RAG methods, blocks 100% of harmful queries, and Boti facilitates more than 3 million conversations each month.

Reported metrics
Monthly conversations handledmore than 3 million conversations each month
Government procedures on city websiteover 1,300
Top-1 retrieval accuracy98.9%
improvement over standard RAG methods12.5–17.5%
Show all 8 reported metrics
monthly conversations handledmore than 3 million conversations each month
government procedures on city websiteover 1,300
top-1 retrieval accuracy98.9%
improvement over standard RAG methods12.5–17.5%
harmful queries blocked100%
voseo usage accuracy98%
periphrastic future usage accuracy92%
retrieval accuracy improvement from embedding summaries vs document sections7.8–15.8%
Reported stack
Amazon Bedrock Converse APIAmazon Titan Text Embeddings v2Cohere Multilingual v3Claude 3.5 SonnetClaude 3 SonnetHaiku 3WhatsApp
Source
https://aws.amazon.com/blogs/machine-learning/meet-boti-the-ai-assistant-transforming-how-the-citizens-of-buenos-aires-access-government-information-with-amazon-bedrock?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The agentic system achieves up to 98.9% top-1 retrieval accuracy, a 12.5–17.5% improvement over standard RAG methods, blocks 100% of harmful queries, and Boti facilitates more than 3 million conversations each month.

What tools did this team use?

Amazon Bedrock Converse API, Amazon Titan Text Embeddings v2, Cohere Multilingual v3, Claude 3.5 Sonnet, Claude 3 Sonnet, Haiku 3, WhatsApp.

What results were reported?

Monthly conversations handled: more than 3 million conversations each month; Government procedures on city website: over 1,300; Top-1 retrieval accuracy: 98.9%; improvement over standard RAG methods: 12.5–17.5% (source-reported, not independently verified).

What failed first in this deployment?

Standard retrieval-augmented generation approaches struggled to disambiguate similar government procedures, and the mixture of chunks they returned increased the likelihood of generating incorrect responses.

How is this customer support AI workflow structured?

Citizen submits question → Input guardrail classifies query → Route approved or blocked query → Reasoning retriever fetches procedures → Sentiment-based prompt routing → Generate response in Rioplatense Spanish.

Related customer support cases