Our read
Replicant sits between the phone line and your contact center processes, and its customer stories show it on narrow, high-volume tasks: payments, scheduling, ride status and cancellation. Public material documents voice, chat and SMS plus the security controls. It does not explain how the telephony and backend connections work, so have the demo prove both before you widen scope.
Where the agent sits on the telephony side
Replicant describes AI agents that handle customer service conversations over voice, chat and SMS for contact centers. The catalogue record lists integrations with Genesys, Cisco, Five9, Talkdesk, NICE inContact and Ujet, plus Salesforce, Zendesk, ServiceNow, Stripe, Shopify and SAP. That is a list of names, not a description of how a call passes from one system to another.
NJ TRANSIT shows the intended posture. Its paratransit service, Access Link, did not replace its phone system: the general manager says a solution that complements what the agency already does was easier to justify than redoing the whole phone setup. The AI took its two highest-volume call types, ride status and ride cancellation, and live agents step in for harder cases. Sunrun's story says only that the rollout was integrated into existing systems, with no detail on the mechanism.
Business data and what happens to the caller's intent
The vendor says agents can be built from a company's existing conversation data and combine large language models with deterministic reasoning for business decisions. The sources do not say which rules a customer configures or who edits them.
ECSI gives the clearest picture of this step. Its agents authenticate the caller up front, capture the intent, then resolve or route the call across three seasonal products: tax forms, refunds and loan servicing. When a call is escalated, the human receives a summary and does not restart the conversation. Fanatics describes the same handoff: when a person is needed, the call is routed with context already captured.
Execution: what the customer examples show
Sunrun started with payment calls, then added autopay enrollment, after-hours support and Spanish. Bulwark Pest Control automated scheduling and payments first, built custom call flows, routed callers to specialists and A/B tested call introductions. Its case study also mentions chat and SMS as an expansion, but it is inconsistent on whether that has launched, so treat it as unconfirmed. For both companies, the work for human agents shifted toward complex cases.
The figures are vendor-published or customer-reported, from different businesses and metrics, so treat them as illustrations rather than a benchmark. Sunrun reports that half of payment-related calls are resolved by AI. NJ TRANSIT reports 12,000 automated calls a week and a 45 percent resolution rate, and says supervisors now coach agents instead of helping with the queue. None of the stories names which backend system carried out a payment or a cancellation.
Supervision, access and data handling
Replicant's security page says the platform captures and analyzes every conversation across channels and agents, for traceability and auditability. It also describes automatic removal of sensitive passages, such as personal and payment data, from transcripts, logs and QA analytics, plus granular permissions with audit trails, role-based access control and multi-factor authentication. The vendor lists SOC 2 Type 2, PCI DSS and HIPAA and states they are independently validated.
On training, the catalogue record says customer data is not used to train models by default. The security page words it more strongly: customer data is never retained, reused or used for training. Since the two wordings differ, get the exact commitment, and any exceptions, written into the contract.
Fanatics shows oversight in use. After adopting Conversation Intelligence, it found that roughly a third of calls had incomplete or inaccurate wrap codes, and every interaction is now tagged automatically. It was still working on pushing those insights into its CRM, so this is a reporting step, not an automated action.
Before you extend: editorial suggestions
These are our recommendations, not vendor capabilities. The customers here began narrow, though not identically: NJ TRANSIT with two call types, Sunrun with payments, Bulwark with scheduling and payments. Do the same with your most repetitive call type and ask to see it end to end in your environment.
Ask which telephony platform carries the call and how it transfers to a person. Ask which backend operations the agent can read and which it can write. Check that the person receives context, and that removal of sensitive passages still leaves enough material for QA. Add chat and SMS once the first call type behaves as you expect.
Method and limits
2026-11-09
This dated snapshot counts distinct companies in public customer evidence. CONFIRMED means a case study, named testimonial or press source; DECLARED means the company itself told usedby; OBSERVED means a vendor listing; SIGNAL means a job post. All four count, but their strength differs. Removed and unlevelled lines are excluded. Examples are a capped selection verified within 90 days. Vendor facts are sourced records checked within 60 days. Coverage is uneven and missing evidence is unknown. These figures measure neither market share nor product performance; capture dates are not adoption dates.
