Armakuni Contact us
← All case studies

In enrollment season, they turned customers away.
There weren't enough counselors to take the calls.

A U.S. benefits technology company enrolls employees by phone, over 100,000 calls a year, each a 10 to 15-minute conversation with a licensed counselor. At peak, demand outran the counselors. We proved an AI agent on Amazon Connect can run that consultative call: verify the caller, retrieve the right employer's plans, explain the trade-offs and hand over to a person when needed. Then we planned the move of the whole contact center onto Amazon Connect, and the build is underway.

Before

A benefits choice is a conversation, not a form.

The contact center was split in pieces.

After
Provedan AI agent on Amazon Connect can carry the 10–15-minute enrollment call
  • A migration roadmap from Genesys to one Amazon Connect contact center.

A benefits choice is a conversation, not a form

The company sits between insurance carriers, brokers and employers, and supports hundreds of employer groups across the United States. Its employees choose their coverage over the phone. A counselor verifies who's calling, asks about their family and their needs, explains the plans their employer offers and walks them through the choice.

Those conversations typically last ten to fifteen minutes. Staffing estimates put the counselor cost at $18 to $32 a call, based on daily rates and calls handled per counselor. At open enrollment, demand climbed past 500 calls a day, more than the counselors could take, and the business was turning work away.

And the contact center was split in pieces

An acquisition had added a payment-services operation serving more than 15,000 insurance agents to an already fragmented contact-center environment. Voice, chat, SMS and email ran through a legacy Genesys contact center with little self-service. People looked up accounts by hand, took notes during calls and wrote summaries afterwards.

Can an AI agent hold the whole call?

Before committing to AI enrollment, the company needed more than answers to simple questions. It needed to know whether an AI agent could run a compliant, consultative benefits call from start to finish: verify identity, understand the employee's situation, find the right plan documents, compare the options fairly and hand over to a person with the whole context.

How we built it

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

  1. Plan tables that keep their meaning
  2. Verify first, then discover, then explain
  3. From one platform's call flows to another's

Plan tables that keep their meaning

A plan table, page by page Plan nameBenefitNetworkValue each value kept tied to its plan
The engineering

Benefits plans are hard for software to read. Deductibles, copays and out-of-pocket limits sit in tables that span several pages, and ordinary document processing loses which number belongs to which plan, benefit and network, which matters because an AI agent that quotes the wrong deductible has done harm. We built a table-aware document pipeline:

  • Page-level visual processing keeps the link between each value and its plan name, benefit type and coverage network.
  • Documents for ten employer groups were indexed with employer-specific isolation.
  • Retrieval runs only within the verified employer's context, so a caller never hears another employer's plans.
  • Answers cite the source document and section, and the workflow blocks any answer without a citation.

Why: in benefits, the only safe answer is one the caller could check in their own plan document.

Verify first, then discover, then explain

Built withConnectBedrockNova SonicOpenSearchS3DynamoDB
Verify the callerDiscover their needsExplain the plansA person, with the handover
The engineering

The proof of concept ran on Amazon Connect, in its HIPAA-eligible environment, and connected the voice conversation to that document library:

  • The AI agent verifies the caller and loads their employer context.
  • Needs discovery adapts to the employee's family circumstances, budget and questions as the call goes.
  • The application retrieves the relevant medical, dental, life and disability plan information and explains the trade-offs.
  • Escalation transfers the call to a person with a handover message and the conversation so far.
  • Recordings, transcripts and a compliance log form the interaction's audit trail.

The design uses Bedrock and Nova Sonic for the conversational AI, with OpenSearch, S3, DynamoDB, Lambda, Step Functions and API Gateway for retrieval and the application workflows.

From one platform's call flows to another's

Built withConnect
underwayGenesysAmazon Connect VoiceChatSMSEmail one shared agent workspace
The engineering

With the proof of concept done, we assessed what moving the whole contact center would take. The assessment examined the existing Genesys flows, queues, routing and IVR, alongside the home-grown account-servicing platform, its SQL Server data and the SharePoint knowledge repository. That produced a gap analysis, a target architecture for one omnichannel contact center and a migration roadmap.

The Connect implementation now underway is configured around:

  • Voice, chat, SMS and email routing in one shared agent workspace.
  • Self-service for the frequent payment and account questions.
  • Customer profiles, payment history and summaries of earlier interactions, on screen when a call arrives.
  • Live knowledge help and automatic interaction summaries, so nobody writes them up afterwards.

One enrollment call, end to end

  1. An employee calls during open enrollment. Amazon Connect answers, and the AI agent verifies who they are.
  2. It loads their employer's context, so only that employer's plans are in reach.
  3. It asks about their situation, for example a spouse, two children, a budget and a doctor they want to keep.
  4. It retrieves the plans from the table-aware index, keeping each deductible with its plan and network, and cites the document and section for every figure.
  5. It explains the trade-offs between the options, and logs the conversation for compliance.
  6. If the employee wants a person, or the AI agent isn't sure, the call goes to a counselor with a handover message and the whole conversation.

What we'd tell another CTO

1Test the hardest call rather than the easiest question; a consultative, compliant enrollment conversation was the real bar.
2Keep tables as tables, and treat the employer context as a security boundary rather than a search filter. A deductible that loses its plan name, or comes from another employer's plan, is worse than no answer at all, which is why uncited answers are blocked.

What else changed

  • A proof of concept showing an AI agent can carry the consultative enrollment call, from identity to plan guidance to handover, on Amazon Connect.
  • A table-aware plan library for ten employer groups, with employer isolation and citations required for every answer.
  • A gap analysis, a target architecture and a migration roadmap from Genesys to one Amazon Connect contact center.
  • The omnichannel implementation underway, with self-service, account context on screen and automatic summaries.

What comes next

The omnichannel Amazon Connect rollout, then AI enrollment in production, measured on completion rate, cost per call and escalation rate.

How we'd do it today

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

The Genesys move would start with FlowShift, which works out queue-based routing from the old skill-based setup and, for a typical Genesys Cloud setup, converts 75–80% automatically by our estimate. The AI counselor would be signed as a Guaranteed AI Outcome against enrollments completed without a counselor, with our engineers inside the business until it's live in 90 days or less, with guardrails that mask personal details and refuse advice the AI agent may not give, and an AK AI Factory golden set of real enrollment calls gating every change. For the counselors, Sonic would score every call from 0 to 100 and pick the calls worth coaching each night, and ComplianceGen would put the right words on screen during a live call, from the company's own compliance documents.

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 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
  3. Amazon Connect$500K+A U.S. benefits platforma year saved, as reportedThey asked for a new phone system. The real problem was everything after the call.70% faster document processing40% more consistent eligibility checksOn AWS Connect · Bedrock · Textract · Nova Sonic

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