An AI voice agent that answers client intake calls around the clock.
Bracken & Cole LLP was losing qualified leads to voicemail and long hold times. We built a multi-tenant AI voice agent on AWS Bedrock and Amazon Connect that handles intake calls directly and routes qualified leads into their case management workflow.
Client
Bracken & Cole LLP
System
Multi-tenant AI voice agent
Role
AI Engineer & Solution Architect
Bracken & Cole LLP was losing qualified leads to voicemail and long hold times during peak call volume, with no after-hours coverage at all. Prospective clients who couldn't get through often didn't call back — every missed call was a missed case.
Amazon Connect owns telephony — call routing, queueing, and compliance-grade recording — while the conversation itself runs on AWS Bedrock: Nova Sonic handles real-time, bidirectional speech-to-speech audio, and a Claude-powered reasoning step decides what to say and what to do with what the caller said.
A Lambda layer sits between the conversation and the firm's systems, extracting structured intake details from natural speech and pushing qualified leads into the case management workflow. Every call is transcribed and logged to Amazon S3 for compliance review, and a defined human-escalation path catches edge cases the agent shouldn't handle alone.
- 01Multi-tenant AI voice agent on AWS Bedrock, Amazon Connect, and Nova Sonic
- 02Natural-conversation intake that captures and checks case details against firm criteria
- 03Automatic routing of qualified leads into the case management workflow
- 04Full call logging and transcription to S3 for compliance and review
- 05Human fallback for edge cases instead of silent failure
AWS Bedrock + Amazon Connect + Nova Sonic + AI Voice Agent + AWS Lambda + CRM Integration + Human Escalation
Built on AWS Bedrock (Claude-powered reasoning), Amazon Connect, Nova Sonic, and AWS Lambda.
How we built it
Audit the opportunity
We mapped the intake process end to end — what information needed to be captured, and how it should route into the firm's existing case management workflow.
Design the operating model
We designed a multi-tenant architecture so the same system could support per-firm configuration and business logic.
Ship fast, validate early
We built the voice agent on AWS Bedrock, Amazon Connect, and Nova Sonic, with Bedrock (Claude-powered) handling the conversational reasoning and a Lambda layer extracting intake details.
Launch with confidence
We added full call logging, transcription to S3 for compliance and review, and human fallback for edge cases before rolling it into production.
What actually got in the way
Keeping the conversation natural at phone-call latency
Bidirectional audio streaming had to feel like a real conversation — natural turn-taking, no dead air — which ruled out chaining separate speech-to-text, LLM, and text-to-speech steps together the way older voice-bot architectures do.
Extracting structured case data from unstructured speech
Callers describe their situation in their own words, out of order, sometimes incompletely. The Lambda layer had to reliably pull out the specific fields the firm's intake criteria needed without forcing callers through a rigid script.
Deciding exactly when to hand off to a human
Too conservative, and the agent escalates routine calls and defeats the point of building it. Too permissive, and a genuinely ambiguous or distressed caller gets left with a bot. Tuning that boundary was as much a product decision as an engineering one.
Designing for more than one firm from day one
Building the multi-tenant architecture — per-firm configuration and business logic — into the system up front, rather than retrofitting it after a single-client build, is what made it possible to support additional clients' intake criteria on shared infrastructure later without duplicating the system per client.
100%
Of client calls answered, day or night
30%
Increase in lead conversion rate
~10/day
Intake calls handled by the agent
The system now runs in production, handling client intake around the clock with no missed calls — every prospective client gets answered, qualified, and routed, whether it's 2pm or 2am. The firm now handles roughly 10 intake calls a day through the agent, answers 100% of them, and has seen a 30% increase in lead conversion rate.
Full call logging and transcription isn't a compliance checkbox — it's how you catch drift in what the agent qualifies as a real lead before it becomes a pattern of missed business. Read more
A single-model, speech-to-speech pipeline beats chaining separate speech-to-text, reasoning, and text-to-speech steps for anything where natural turn-taking matters — every extra hop in the older architecture is a place latency and prosody get lost.
Human fallback isn't a fallback for failure — it's a designed path for the calls that should always reach a person, and building it in from the start is what let the firm trust the system enough to actually retire the old process.
Have a similar product idea?
Share the workflow, product idea, or knowledge problem you're solving, and we'll help define the smartest first build.