Armakuni Contact us
← All case studies

They asked for a new phone system.
The real problem was everything after the call.

A U.S. benefits platform verified every employee's dependents by hand: a contractor called, asked for documents, waited, checked and typed the result. We moved its contact center to Amazon Connect and built AI agents, on voice and chat, that guide the employee, collect the document, check it against the rules and write the result back, with a clear path to a person for exceptions.

Before

Every verification went through a person, one employee at a time.

The phone and the chat couldn't do the work.

After
$500K+a year saved
70%faster document processing
40%more consistent eligibility checks

As the company reported.

Every verification went through a person, one employee at a time

When an employee adds a spouse or child to a benefits plan, the company's platform has to confirm the dependent qualifies: a marriage certificate, a birth certificate, a government ID. The documents existed and the rules were clear. A contractor still had to move each case along by hand.

Someone called the employee, explained what to send, waited for it, inspected it and updated the record. If something was wrong or missing, the loop started again. The number of verifications done each week was set by how many contractors were working.

The phone and the chat couldn't do the work

The existing contact center and chat tools could answer basic questions, but they couldn't send an upload link during a call, see a document's status or check eligibility. Every real question ended up with a person, and a question that should have taken seconds could take days.

People also applied the same rules differently. Some documents passed that shouldn't have, and some were rejected with no clear reason recorded. In benefits administration, that's a compliance problem, with protected health and personal information on every record. Employers couldn't see unfinished verifications until deadlines were close.

The brief was a phone system. The fix was the workflow

The company came to replace its phone platform. Mapping the workflow showed the phone system wasn't the bottleneck: the verification process behind it was. So we connected the conversation, the document and the system of record into one service, and moved the phones as the foundation for it.

How we built it

5 decisions, in the order we made them. Each one says what we did and why, with what it did drawn beside it.

  1. Get the architecture right before building
  2. Prove the whole call, not one piece of it
  3. Retrieval answers questions. Rules decide the case
  4. The phone system moved in a weekend
  5. Employers see unfinished verifications before the deadline

Get the architecture right before building

Built withConnectLexBedrockTextractIAM
The whole path, before building The call1The upload2The extraction3The rules check4 then build what the map showed
The engineering

Building the wrong thing first is expensive when every record holds personal health information. The assessment ran workshops across inbound voice, the web portal and chat, mapped the whole flow, and found the gaps in requirements, the undefined business rules and the integration points. It assessed Amazon Connect, Amazon Lex, Bedrock and third-party tools against the company's own voice, chat and document-verification requirements, and documented the recommendation against the alternatives. It recommended Amazon Textract for entity extraction, Bedrock where a case needs more than rule-based matching, and Amazon Connect for voice and chat, with data flow, PHI and PII security, IAM and SSO, scalability and the OCR strategy designed in from the start.

Prove the whole call, not one piece of it

Built withTwilioNova SonicBedrockS3StreamlitTextract
An SMS sends an upload linkThe employee uploads itTextract and Comprehend read itRules compare it with recordsThe result, in the same call
The engineering

The six-to-seven-week proof of value tested the full interaction on real employee records and document types, with inbound calls through Twilio and Amazon Nova Sonic on Bedrock for speech-to-speech conversation. The AI agent explained the required documents, kept context through follow-up questions, and triggered the next action when the employee was ready:

  • An SMS sent a presigned S3 upload link, so the employee could submit evidence during the interaction.
  • A Streamlit interface accepted the document and showed its processing status.
  • Textract and Comprehend extracted names, dates and relationship information.
  • Rule-based checks compared the extracted fields with the reference employee records.
  • The interface and the conversation returned the verification result, in the same interaction.

The production build replaced Streamlit with a React interface embedded in the benefits platform, so the work stays inside the application employees already use. We proved the whole call, because a single step on its own can't show whether the business case holds.

Retrieval answers questions. Rules decide the case

Built withOpenSearchBedrockFastAPIECSFargateLambda
Each case Answers: retrievalDecides: rulesOdd case: person the rules decide, never the model
The engineering

The most important design line is between explaining and deciding. Curated eligibility guides, FAQs and document requirements were chunked and embedded into an OpenSearch knowledge base with Bedrock embedding models, and the chat AI agent retrieves from it to answer questions. Application rules and record matching decide the document outcome, never the model.

The production architecture uses FastAPI services on ECS/Fargate under auto-scaling, with direct APIs to the benefits application. Voice and chat share one processing path:

  • Either channel can start an upload, and Lambda coordinates extraction and rule-based validation, including required fields, identifier format and signature checks.
  • DynamoDB records the extracted data and verification state, and an API pushes the outcome into the benefits system of record, with no separate data entry.
  • A document that needs review appears with its original submission, its extracted fields and the reason for the exception, so the reviewer sees exactly what failed instead of starting from an unexplained status.

Why: a model can explain a rule well and still apply it inconsistently, while rules in code apply the same way to every document, which is where the 40% gain in consistency came from.

The phone system moved in a weekend

Built withConnectS3TwilioOpenSearchBedrock
Old phone systemAmazon Connect Contact flowsContact flowsVoice AI agentVoice AI agentChat AI agentChat AI agent moved over a weekend
The engineering

The Connect migration covered number porting, contact flows, routing rules, agent profiles, SSO and outbound permissions. Historical contact-center and chatbot records were archived in S3. The cutover happened over a single weekend: on Monday morning the team opened the same tools, running on a different platform underneath.

The phased deployment moved from the Twilio-based voice integration to Amazon Connect, keeping the containerized backends, the OpenSearch knowledge base and the shared AWS delivery pipeline:

  • Amazon Connect receives the call and routes it through contact flows to the Bedrock voice application, over SIP.
  • The AI agent handles FAQs, upload guidance and verification-status questions with the connected workflow.
  • Low-confidence conversations and explicit requests for help transfer to a live agent, with the call context in the Connect workspace.
  • The replacement chat AI agent uses the same knowledge base and verification pipeline, and escalation keeps the whole conversation history for the receiving team.

Employers see unfinished verifications before the deadline

Built withKinesisGlueBedrock
SubmissionsEscalationsNear the deadline Employer seen before the deadline
The engineering

We brought operational reporting into the employer portal, next to the verification:

  • Contact Trace Records stream through Kinesis and Glue into reporting data.
  • Bedrock transcript analysis supports sentiment and compliance review.
  • The dashboard shows submission volumes, escalation rates and searchable transcripts, so an employer can look at the interactions behind the totals.

One verification, end to end

  1. An employee calls about adding their spouse. Amazon Connect routes the call over SIP to the Nova Sonic AI agent.
  2. The AI agent explains what's needed, using the knowledge base, and sends an upload link by SMS.
  3. The employee photographs the marriage certificate and uploads it inside the benefits app.
  4. Lambda sends it to Textract, which extracts names, dates and the relationship.
  5. The rules check it against the employee record: required fields, identifier format, signature.
  6. It passes, and the result goes to the system of record, all before the call ends. If it fails, a reviewer gets it with the exact mismatch.

What we'd tell another CTO

1Look behind the system you were asked to replace. The brief was a phone system, and the bottleneck was the verification process behind it.
2Let retrieval explain and let rules decide, and record the reason for every exception, so a reviewer starts from the mismatch rather than from an unexplained status.

What else changed

As the company reported.

  • Every inbound call answered by the AI agent first, with a person for the cases that need one.
  • Employers watching their program live, with sentiment, escalations, submissions and transcripts in their portal.
  • The phones moved to Amazon Connect over one weekend.

What comes next

Outbound calling, advanced fraud detection and multi-language document support, then Amazon Connect and Bedrock extended into the company's other claims workflows. The infrastructure is in place, so the work is adding the use cases.

How we'd do it today

This project came before our current accelerators. Here's how we'd approach the same work now.

The phone move would start with FlowShift, which reads the old platform's hours, queues, routing, prompts, flows and help articles and rebuilds them for Amazon Connect after a dry run that gives every item a verdict (for a typical Genesys Cloud setup, our estimate is that 75–80% converts automatically), and the Compliant Connect Accelerator, an eight-step wizard, would build Connect in the company's own account in about 25 to 45 minutes, with each control mapped to its rule, HIPAA included. The first AI workflow, built on AK AI Factory, would be a Guaranteed AI Outcome, signed against cost per verification, with a go-or-stop decision at week 4 and our fee riding on the result.

Amazon Connect Guaranteed AI Outcomes Modernize with AI

More work like this

  1. Amazon Connect4 weeksA Medicare enrollment contact centerto the first honest baseline of the contact centerIts own leaders rated the platform 4 out of 10. Medicare enrollment season was coming anyway.600 people on a steadier platform3 Medicare Advantage health-plan clientsOn AWS Connect · Contact Lens · Lambda · DynamoDB
  2. Amazon ConnectProvedA U.S. benefits technology companyan AI agent on Amazon Connect can carry the 10–15-minute enrollment callIn enrollment season, they turned customers away. There weren't enough counselors to take the calls.100,000+ enrollment calls a year10–15 minutes for each of those callsOn AWS Connect · Nova Sonic · Lambda · DynamoDB
  3. Amazon Connect2,200+A contact-center outsourceraudio files converted and liveTheir move to Amazon Connect was stuck behind 2,500 call scripts. So we built the thing that converts them.2,500+ call scripts with an engine to convert them4 shared contact flows instead of hundredsOn AWS Connect

All case studies →

Have a problem like this one?

A free workshop: a few hours with the architects who'd do the work, on your own systems. No cost.

Book a free workshop