Guided Psychological Intake:
The client’s Agentic AI PoC on AWS
About the client
The client sought to validate the core of an AI-assisted psychological intake flow that can conduct a brief guided interview, recommend suitable professionals, and generate a structured preliminary report for therapist review.
Challenges
The client needed to validate a scalable and consistent way to support the initial psychological-intake process while preserving human oversight. The core workflow required a guided 10-15 question interview, context-aware clarification, a Top 3 professional recommendation with rationale, and a structured preliminary report for therapist review. Without a validated agentic core, the organization would remain dependent on manual intake, matching, and report preparation and would lack a proven foundation for later capabilities such as scheduling, notifications, and patient-facing portals. We therefore focused the engagement on validating an Agentic AI workflow that could guide a multi-step interaction, reason over responses, invoke actions, recommend an appropriate next step, and generate a persistent report without replacing professional judgment or issuing a diagnosis.
Project Goals & Objectives
-
Primary Goal: Build and validate a conversational psychological-intake agent on Amazon Bedrock within a 100-hour PoC.
-
Conduct a guided 10-15 question intake and allow natural-language follow-up questions for clarification.
-
Use the client documentation and guidance to support the conversation.
-
Recommend a Top 3 list of professionals and explain why each is appropriate.
-
Generate a structured preliminary therapist report when a professional is selected and store it in Amazon S3.
-
Compare more than one Bedrock model during the PoC and document model trade-offs.
-
Validate the solution only with simulated/internal conversations; the agent must not replace a professional or issue a diagnosis.
The Solution
We implemented the PoC with Amazon Bedrock Agents to orchestrate the guided questionnaire, conversational flow, professional recommendation, and report-generation workflow. The implemented engagement used Haiku 4.5 through Amazon Bedrock. Bedrock Agent Action Groups invoke AWS Lambda functions to read the file-based professional catalog and generate and store reports in Amazon S3. We intentionally validated the solution from the Amazon Bedrock console and kept the PoC focused on the core agent behavior, avoiding unnecessary production components, external patient routing, scheduling integrations, or a dedicated relational database during this phase.
Designation Definition Fit
This reference directly meets the AWS Agentic AI specialization scope because we implemented a goal-oriented, multi-step agent workflow rather than a simple conversational assistant. The Bedrock Agent guides the intake, reasons over the user’s responses and the client’s guidance, invokes actions to access the professional catalog, recommends a Top 3 set of professionals with rationale, and creates a persistent structured report in Amazon S3 after a professional is selected. Human review and the explicitly non-diagnostic scope provide controlled oversight for the autonomous workflow. The solution demonstrates managed agent orchestration, multi-step reasoning, tool use, action execution, and human-in-the-loop validation.
Key Technologies Implemented
🔹Amazon Bedrock Agents - managed orchestration of the questionnaire, conversation, recommendation, and report workflow. | 🔹Haiku 4.5 on Amazon Bedrock - foundation model used in the implemented engagement. | 🔹Bedrock Agent Action Groups - invoke functions for professional-catalog access and report persistence. | 🔹AWS Lambda - lightweight action execution backend. | 🔹Amazon S3 - generated report storage, file-based professional catalog, and optional RAG documentation storage. | 🔹AWS IAM - service-layer access control. | 🔹AWS CloudTrail - audit logging confirmed as implemented. | 🔹Amazon CloudWatch - operational logging and basic monitoring confirmed as implemented. | 🔹 AWS KMS - proposed encryption for the S3 report bucket in the technical proposal.
Technical Solution & Justification
The architecture intentionally minimizes infrastructure for a 100-hour internal PoC. Bedrock Agents provides the managed agent runtime, Lambda executes actions, and S3 stores reports and reference files. Testing through the Bedrock console removes the need for a custom frontend during validation. Responsible AI controls include simulated conversations, human review of generated reports, an explicit statement that the agent does not replace a professional, and a prohibition on diagnosis. No external APIs, persistent memory database, MCP/A2A, or production patient-routing integrations are evidenced.
Architecture Highlights
-
Guided multi-step workflow: 10-15 intake questions with natural-language clarification.
-
Recommendation action: Top 3 professionals are selected with an explanation of the match.
-
Persistent business artifact: selection of a professional triggers generation of a structured therapist report stored in S3.
-
Controlled validation: simulated/internal conversations and human review are used before any future production evolution.
Solution Optimality & Alternative Approaches
We intentionally selected a simplified serverless architecture to validate the agent core within the fixed 100-hour PoC. By focusing on the interview-recommendation-report workflow first, we avoided introducing scheduling, portals, multi-channel delivery, or a relational database before proving the central agent behavior. This approach reduced implementation overhead and created a reusable foundation on which a future MVP can add those capabilities without rebuilding the core workflow.
Business Impact & Key Performance Indicators (KPIs)
We measured the following KPIs to validate the business impact of the implemented solution:
-
KPI 1 - Intake and Preliminary Report Preparation Efficiency: Baseline: the manual intake and preliminary report-preparation effort was normalized to 100%. Target: reduce the combined effort required to collect structured intake information and prepare a therapist-ready preliminary report by at least 30%. Actual result: internal PoC validation showed an approximately 35% reduction in the effort required to reach a review-ready report. Measurement method: we compared completion effort for simulated manual intake/report preparation with equivalent simulated conversations processed through the agent and reviewed by the team.
-
KPI 2 - Professional Recommendation Quality: Baseline: initial manual/simulated matching scenarios produced approximately 65% first-pass agreement with the preferred professional shortlist. Target: achieve at least 80% reviewer agreement with the agent’s Top 3 recommendations and rationale. Actual result: PoC validation achieved approximately 84% first-pass reviewer agreement. Measurement method: The client reviewers evaluated simulated conversations and compared the agent’s Top 3 recommendations and explanations with the professionals they considered appropriate for the same scenarios.
Unmet Goals & Lessons Learned
The PoC confirmed that a constrained, human-reviewed workflow is the appropriate first step for this use case. Simulated conversations, model comparison, prompt and rule adjustment, report review, and client feedback allowed us to refine the core behavior while maintaining a clear non-diagnostic boundary. The engagement also showed that validating the central intake, recommendation, and report flow before adding scheduling or patient-facing production integrations reduces complexity and provides a safer path toward a future MVP.
Get in touch
Take your business to the next level.