Built for the call shape Atlanta actually has.

The reality of phones in Atlanta.

Industries we work with in Atlanta.

Four call-flow patterns that show up repeatedly in Atlanta engagements. Each has its own integration list, compliance bar, and surge shape.

Questions Atlanta operators ask before they build.

How does the voice agent handle a Delta IRROPS event when 80,000 callers hit the rebooking line in two hours?

IRROPS (irregular operations: weather cancellations, ground stops, equipment swaps) is the worst-case shape for an airline customer-service line, and Delta hub operations at Hartsfield-Jackson push that pattern several times a year. We design the runtime for that load. The agent runs on ECS Fargate behind an Application Load Balancer with horizontal autoscaling triggered by active-call count, not CPU, and per-instance session bounds that keep latency inside 900ms even when capacity is climbing. The PNR-lookup and rebook tool calls into Delta's res system run through a circuit breaker so a slow downstream does not cascade. Pre-warming triggers from a weather feed (NWS aviation advisories for ATL, JFK, LAX, LGA, and DTW hubs) so capacity is up 60 to 90 minutes before the rebook surge hits. Calls that exceed the agent's policy authority (involuntary downgrades, partner-airline rebooks, hotel vouchers above a threshold) route to the human queue with the PNR context and rebook attempts already captured. The agent does not pretend to be human, and the IRROPS script is the place where caller frustration is highest, so the disclosure stays explicit.

Can the agent integrate with WellStar, Piedmont, and Emory Epic Cadence at the same time for a metro-wide referral flow?

Multi-system Epic integration is common in metro Atlanta because the three big healthcare networks (WellStar with 11 hospitals, Piedmont with 22 hospitals across Georgia, Emory with 11 hospitals) all run Epic but as separate instances. We integrate through Epic App Orchard with the Patient Access and Scheduling endpoints against each system's interconnect with separate OAuth 2.0 credentials. The agent reads availability across the requested network, presents the next three slots, writes the appointment back atomically, and triggers the system's native SMS confirmation. For a referral flow that crosses networks (a Piedmont primary care referring to an Emory specialist), the agent captures the referral on the Piedmont side and creates a referral order plus a scheduling task on the Emory side, which a human navigator approves before the appointment locks. The build includes an Epic certification cycle with each system's analyst, and we budget 4 to 6 weeks for a three-system integration versus 2 to 3 weeks for a single system.

How does the agent handle Southern accents and code-switching from AAVE to standard American English?

Speech-to-text accuracy on Southern English varies by provider. We test against recorded Atlanta caller samples before we ship. Deepgram Nova-2 has held above 92 percent word-accuracy on our Atlanta test set across the range from rural-Georgia Southern to Inman Park standard American to AAVE-feature-heavy speech. AssemblyAI Universal is close. Whisper through OpenAI lags by 4 to 6 points on the same set. For UPS driver-support calls where ambient car noise stacks on top of accent variation, we add noise-suppression on the inbound audio stream (Krisp or a custom RNN-based filter) before STT, which recovers 3 to 5 points. The bigger risk is in slot-filling for proper nouns: Atlanta neighborhood names (Cabbagetown, Kirkwood, Reynoldstown) and Spanish-derived street names trip generic NER models. We hand-tune the named-entity extraction layer with a custom gazetteer of Atlanta locations and Georgia place names, which closes the gap.

What does compliance look like for Equifax FCRA dispute intake or insurance claims calls in Georgia?

FCRA dispute intake at Equifax (or Experian, TransUnion) has specific Federal Trade Commission rules about how the caller is identified, what disclosures the agent has to read, and how the dispute record is captured for the 30-day investigation window. We script the agent's opening disclosure with the legal team and require an explicit consent capture before the agent collects any PII. The dispute record is written to the consumer-relations system with the agent's transcript attached as an evidence artifact, which the human dispute analyst reads during investigation. For insurance claims intake in Georgia, the Department of Insurance rules require call recording for first-notice-of-loss calls; we record to an S3 bucket in the customer's AWS account with KMS keys the customer controls and CloudTrail audit logs on every access. Both flows route to a human at the first sign of fraud, dispute escalation, or caller distress.