Can AI Handle 90% of Data‑Protection Tasks? Multi‑Agent Assistant Cuts Time 70%

Overview

Generative AI is reshaping enterprise‑customer interactions. Druva, in partnership with Amazon Web Services, has created a multi‑agent AI assistant that leverages Amazon Bedrock large language models to provide a conversational interface for data‑management, security insights, and operational support.

Architecture

The system’s core is a supervisory Agent that coordinates multiple specialized sub‑agents:

Data Agent uses GET APIs to retrieve backup schedules, status, and related details.

Assist Agent draws on a knowledge base of API documentation, manuals, and FAQs to offer best‑practice guidance and step‑by‑step troubleshooting.

Operation Agent executes actions such as initiating backups or modifying protection policies via POST APIs.

When a user submits a natural‑language query, the supervisory Agent analyses the request, determines its business attribute, and routes it to the appropriate sub‑agent. Dynamic API selection is powered by the Amazon Bedrock Knowledge Base, which semantically ranks relevant APIs; the LLM then parses the top‑K candidates to choose the optimal API and configure its parameters.

Key Benefits

Simplified user experience through natural‑language interaction.

Intelligent fault diagnosis that pinpoints root causes of backup failures and suggests personalized remediation steps.

Optimized policy management that guides users through creation, modification, and enforcement of data‑protection policies, reducing human error.

Proactive support that continuously monitors the environment, identifies potential issues, and offers preventive guidance.

Scalable operations capable of handling massive volumes of customer inquiries while freeing support teams to focus on complex, strategic tasks.

For example, a global financial services firm managing over 500 servers reduced backup‑failure investigation time from several hours to a few minutes by asking, “Why did my backup fail last night?” and receiving an analysis that identified a policy‑update conflict in the European data center.

Evaluation Methodology

The assistant was subjected to a rigorous three‑layer testing regime:

Unit tests for each component (individual agents, data extraction module, API‑selection workflow).

Integration tests to verify seamless communication and data/control flow among agents.

System tests that simulate real‑world user workflows to assess overall functionality, performance, and user experience.

Evaluation Results

Dynamic tool selection proved critical; selecting the correct API determines end‑to‑end success. Benchmarking against a Sonnet 3.7 baseline showed that lightweight models such as Amazon Nova Lite and Haiku 3 could select the right API but struggled with accurate parameter parsing. Model‑selection accuracy figures were:

Amazon Nova Micro – 81%

Amazon Nova Lite – 88%

Amazon Nova Pro – 93%

Haiku 3 / Haiku 3.5 / Sonnet 3.5 – 91‑92%

Nova Pro achieved the best balance, with an average response time just over 1 second, whereas Sonnet 3.5 incurred ~8 seconds latency due to longer output (average 291 tokens vs. 86 tokens for Nova Pro). In end‑to‑end testing, domain experts rated the system 3.3 out of 5 on completeness, accuracy, and relevance—a solid score for an early‑stage prototype.

Impact

The assistant can complete roughly 90% of routine data‑protection tasks via natural language, cutting average resolution time by up to 70% and turning multi‑hour manual investigations into instant conversational insights. The underlying principles—agentic AI orchestration, dynamic API selection, and human‑in‑the‑loop safeguards—are applicable across industries.

Implementation Details

The solution relies on Amazon Bedrock AgentCore Runtime and AgentCore Gateway for agent orchestration. Interested readers can explore Amazon Bedrock Agents, the Bedrock Knowledge Base, and the fully managed AgentCore offering for similar implementations.