AI-Powered Conversational Assistant for Streamlined Home Buying Experience

rocket · Finance · 2025 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Home-buying and mortgage customer engagement

Rocket needed to make a complex homeownership journey available across web, mobile, servicing, origination, and broker contexts with personalized guidance, backend actions, and human support when needed.

Constraints:

Architecture

Rocket AI Agent is a centralized web/mobile chat capability backed by Amazon Bedrock Agents and eight domain-specific agents. The Rocket AI Agent API routes by static channel/property signals or LLM intent, agents decompose tasks with ReAct, use Bedrock Knowledge Bases/Kendra over Rocket content, invoke action groups for backend operations, enforce Guardrails, retain memory across turns, and escalate to mortgage bankers when confidence or keywords indicate human help is needed.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  web-mobile-chat-clients["web-mobile-chat-clients"]
  rocket-ai-agent-api["rocket-ai-agent-api"]
  domain-specific-agents["domain-specific-agents"]
  bedrock-knowledge-bases["bedrock-knowledge-bases"]
  bedrock-action-groups["bedrock-action-groups"]
  rocket-backend-services["rocket-backend-services"]
  bedrock-guardrails["bedrock-guardrails"]
  agent-memory["agent-memory"]
  human-banker-escalation["human-banker-escalation"]
  web-mobile-chat-clients --> rocket-ai-agent-api
  rocket-ai-agent-api --> domain-specific-agents
  domain-specific-agents --> bedrock-knowledge-bases
  domain-specific-agents --> bedrock-action-groups
  domain-specific-agents --> bedrock-guardrails
  domain-specific-agents --> agent-memory
  bedrock-action-groups --> rocket-backend-services
  bedrock-guardrails --> human-banker-escalation
  domain-specific-agents --> human-banker-escalation
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
rocket-ai-agent-api Unified API for chat clients that routes requests to the appropriate Bedrock-backed domain agent. domain-specific-agents, web-mobile-chat-clients
web-mobile-chat-clients Rocket mobile app and website chat surfaces where clients ask homeownership questions and start workflows. rocket-ai-agent-api
domain-specific-agents Eight job-scoped agents for loan origination, servicing, broker support, payments, refinance, and related mortgage workflows. rocket-ai-agent-api, bedrock-knowledge-bases, bedrock-action-groups, bedrock-guardrails, agent-memory
bedrock-knowledge-bases Retrieval over Rocket Learning Center, FAQs, compliance documents, and proprietary servicing workflows using Amazon Kendra internally. domain-specific-agents
bedrock-action-groups Secure backend operations such as lead submission, filling preapproval forms, scheduling payments, and other self-service actions. domain-specific-agents, rocket-backend-services
rocket-backend-services Rocket proprietary systems and APIs for mortgage, servicing, broker, and payment operations. bedrock-action-groups
bedrock-guardrails Responsible-AI boundary layer that blocks off-topic or unsafe language and keeps the experience in scope. domain-specific-agents, human-banker-escalation
agent-memory Maintains context across multiple conversation turns for personalized, natural interactions. domain-specific-agents
human-banker-escalation Live support path triggered by confidence thresholds, keywords, or client preference for complex/sensitive mortgage needs. bedrock-guardrails, domain-specific-agents

Data flow

A client starts chat in Rocket's mobile app or web page. The Rocket AI Agent API chooses a domain agent by static origin or LLM-based intent identification; the agent breaks the request into subtasks, retrieves from Knowledge Bases, executes action groups such as lead submission or payment scheduling, applies Guardrails, and either responds or transitions the session to a live banker when the interaction becomes uncertain, complex, or sensitive.

“The AI Agent API routes the request to the correct Amazon Bedrock agent based on static criteria, such as web or mobile property that the client entered the chat through, or the use of LLM-based intent identification”article, stated
“The agent breaks the task into subtasks, determines the right sequence, and executes actions and knowledge as it works”article, stated

Why this architecture

Rocket chose a unified interface over narrow app-specific bots, but kept each agent tightly scoped to avoid monolithic drift. Bedrock managed agents, Knowledge Bases, action groups, guardrails, memory, and cross-Region inference give the system retrieval, action, safety, continuity, and burst resilience without making every mortgage case fully automated.

“assigning each agent a tight scope of 3–5 actions led to more maintainable, testable, and high-performing agents.”article, stated
“Escalation isn’t failure, it’s a critical part of user trust.”article, stated

Models used: foundation models via Amazon Bedrock Agents

Scale

Alternatives considered (roads not taken)

One broad monolithic mortgage agent

Rejected because: Rocket found tight 3-5 action scopes per domain agent were more maintainable, testable, and performant.

“Rocket found that assigning each agent a tight scope of 3–5 actions led to more maintainable, testable, and high-performing agents.”article, stated

AI-only handling when the conversation becomes uncertain or sensitive

Rejected because: Rocket proactively transitions to live support or offers escalation to avoid frustrating loops and preserve trust.

“Rocket AI Agent proactively transitions the session to a live support agent or gives the user the option to escalate.”article, stated

Single-Region model inference for bursty mortgage traffic

Rejected because: Cross-Region inference avoids quota bottlenecks and improves latency/availability during launches or interest-rate shifts.

“During peak traffic windows such as product launches or interest rate shifts this architecture has allowed Rocket to avoid Regional service quota bottlenecks, maintain responsiveness, and increase throughput”article, stated

Patterns

multi-agent-orchestration (customer-journey)

Rocket routes clients through a unified interface to domain-specific agents for origination, servicing, broker support, and other functions.

“At the core of its architecture is a growing network of domain-specific agents currently eight each focused on distinct functions such as loan origination, servicing, or broker support.”article, stated

tool-use-orchestration (backend-actions)

The agent decomposes tasks, retrieves Rocket data, and invokes backend operations such as lead submission or payment scheduling through action groups.

“Agents use the FM to analyze a user’s request, break it into actionable steps, retrieve relevant data, and trigger downstream APIs to complete tasks.”article, stated

agent-memory (personalization)

Memory keeps conversational context across turns so clients get more natural and personalized assistance.

“Memory retention allows Rocket AI Agent to maintain contextual awareness across multiple turns, enhancing user experience with more natural, personalized interactions.”article, stated

guardrails-io-filtering (responsible-ai)

Bedrock Guardrails constrain the assistant to appropriate topic boundaries for Rocket's responsible AI goals.

“Supports Rocket’s responsible AI goals by making sure that the agent stays within appropriate topic boundaries.”article, stated

confidence-thresholding (handoff)

Rocket uses confidence scores and keyword triggers to decide when to transition from AI to live support.

“Rocket implemented uncertainty thresholds using confidence scores and specific keyword triggers to detect when an interaction might require human assistance.”article, stated

Tradeoffs

How to structure the assistant's intelligence.

Chose a unified interface backed by domain-specific agents over one broad monolithic mortgage agent.

Rationale: Specialized agents can keep tighter scope and still provide continuity through shared routing and client context.

“These agents work together behind a unified interface to provide seamless, context-aware assistance.”article, stated

How to combine self-service and expert guidance.

Chose AI assistance with seamless transitions to mortgage bankers over pure automation or always-human support.

Rationale: The AI handles common guidance and actions, while bankers remain available when the client or risk profile requires them.

“Seamless transitions to Rocket Mortgage bankers when human support is needed”article, stated

Gotchas (what bit them)

Broad agents are harder to test and maintain.

Root cause: A single agent with too many actions makes task ownership and resolution metrics blurry.

Resolution: Rocket scoped each agent to about three to five actions with documented interfaces and separately monitored resolution rates.

“Rocket found that assigning each agent a tight scope of 3–5 actions led to more maintainable, testable, and high-performing agents.”article, stated

Escalation loops can damage user trust.

Root cause: Sensitive or complex mortgage interactions may exceed what an AI assistant should resolve alone.

Resolution: Rocket treats graceful escalation as part of the product, using proactive transfer to live support or optional escalation.

“Escalation isn’t failure, it’s a critical part of user trust.”article, stated

Regional service quotas can bottleneck model serving during bursts.

Root cause: Traffic spikes from product launches or interest-rate shifts create unpredictable load.

Resolution: Rocket enabled cross-Region inference early to route requests to capacity and preserve responsiveness.

“During peak traffic windows such as product launches or interest rate shifts this architecture has allowed Rocket to avoid Regional service quota bottlenecks”article, stated

Evaluation

Rocket reports conversion, containment, handoff reduction, customer satisfaction, task completion, and agent task-resolution metrics.

Metrics: closed-loan conversion · transfer-to-care reduction · transfer-to-servicing reduction · customer satisfaction · task resolution rate

“Each agent’s capabilities use Amazon Bedrock action groups with well-documented interfaces and monitored task resolution rates separately.”article, stated
“Since launching Rocket AI Agent, we’ve seen transformative improvements across the client journey and internal operations”article, implied

Outcomes

Rocket uses 11 PB of data and deploys the assistant across the majority of its web pages, mobile apps, servicing portal, and third-party broker system.

“Clients who interact with Rocket AI Agent are three times more likely to close a loan compared to those who don’t.”article, stated
“Rocket saw an 85% decrease in transfer to customer care and a 45% decrease in transfer to servicing specialists.”article, stated