Conversational AI is software that holds a back-and-forth conversation to get something done. Most lists of examples stop at Siri, Alexa, and website chat widgets. Those are real, but if you’re deciding whether to put an agent on your own phone lines, they won’t tell you much. You want to know who runs one in production, what it handles, and what changed after it went live.
Here are twelve deployments that answer those questions. Every one runs on Cartesia’s voice models, because those are the ones we can show you numbers for. Each links to the full customer story. The pattern across them matters more than any single example, so there’s a short section on that at the end.
The twelve examples at a glance
| Company | What the conversational AI does | What changed |
|---|---|---|
| Lorikeet | Resolves support calls for regulated businesses | 21,000+ calls in nine days during a benefits crisis, at 100% uptime |
| Replicant | Answers contact-center calls for brands like Fanatics | Calls answered instantly; containment improved within weeks |
| Assort Health | Books and moves patient appointments | Bookings up 7%, patient wait time down 89% |
| SuperDial | Calls insurers for benefits checks and claim status | 4x billing-team productivity |
| Arini | Answers the phone for dental practices | More than 2x faster responses than its previous provider |
| EliseAI | Schedules apartment tours and handles resident requests | Tour conversions up 5.3% after a regional-voice test |
| Fundamento | Calls back loan applicants who stall mid-application | 42% of dropped-off applicants came back and finished |
| 2X Solutions | Makes and answers sales calls for lead-heavy businesses | Abandoned calls under 1.25%, well inside the 3% legal limit |
| Thoughtly | Calls and qualifies leads straight from the CRM | Sub-200ms end-to-end latency on live calls |
| Bolna | Runs support and sales calls in Indian languages | 500,000+ calls a day; ~40% less engineering time spent on latency fixes |
| Tavus | Video avatars that hold a live conversation | Under one second end to end |
| Ego | Game characters players can talk to | Spoken NPCs and companions in a Valheim mod |
Customer support
Lorikeet: a crisis line stood up in a day. Lorikeet builds support agents for regulated businesses in financial services, healthcare, insurance, and travel. When 300,000+ families suddenly had their SNAP food benefits moved to GiveCard, call volume spiked overnight. Lorikeet had voice agents answering within 24 hours. In the first nine days, they took more than 21,000 calls, with hundreds running at once, in English, Spanish, and Mandarin, at 100% uptime. The agents verified callers and troubleshot card problems with them. Callers often thanked the agent and told it to have a nice day, even though the agent said up front that it was AI.
The lesson isn’t that the agent was clever. It’s that the job was narrow (card problems), the data was available (account lookups), and the voice was calm enough that stressed callers stayed on the line.
Replicant: the high-volume middle of a contact center. Replicant automates customer conversations for enterprises including DoorDash, AAA, and Fanatics. For Fanatics, that means absorbing seasonal call spikes without adding headcount. Since the switch, calls are answered instantly, and the human team spends its time on the calls that need judgment. Replicant reports 99.99% uptime, and it needs that: when sudden weather floods a roadside-assistance line, an agent that slows down under load is worse than no agent.
If support is your first use case, the contact center automation guide covers how to pick which calls go first.
Healthcare
Assort Health: scheduling for specialty clinics. Assort’s agents answer patient calls for more than 140 clinics and have handled more than 5 million patient interactions. They book, reschedule, and answer office questions, working directly in the clinic’s EHR. Bookings rose 7%, operating costs fell 63%, and patient wait time dropped 89%. The hard part was getting dates, addresses, insurance plans, and drug names right every time, over phone audio, across thousands of edge cases.
SuperDial: the calls no patient sees. A lot of a clinic’s phone time is staff on hold with insurers. SuperDial’s agents make those calls: they navigate payer phone trees, wait on hold, and verify benefits or chase claims with the insurer’s own representative. Billing teams get four times as much done.
Arini: the front desk for dental practices. Arini answers calls around the clock for dental offices, books appointments, and answers insurance and billing questions in English and Spanish. Its co-founder put the stakes plainly: a dropped call can be worth upwards of $5,000 to a practice. Every bit of delay raises the chance a patient hangs up, which is why Arini switched to a provider that answered more than twice as fast.
There’s more on drawing the clinical line, HIPAA, and pilot metrics in conversational AI in healthcare.
Housing
EliseAI: one in six US apartments. About 1 in 6 US apartments use EliseAI. Its agents book tours, take maintenance requests, handle renewals, and help residents through eligibility and recertification, across more than 5 million voice calls a month. Multilingual voices let it serve affordable-housing residents in their own language, after hours.
EliseAI is the clearest example here of how to test conversational AI. Every new voice passes four gates: a public benchmark, a blind A/B test with human raters, a 1% production rollout measured on real outcomes, and a manual audit of sampled calls. When they tried regional voices (a Midwestern voice in Illinois, a Southern voice in Texas), tour conversions rose 5.3%, caller engagement 7.8%, and callers spoke 15.8% more words. More words matters more than it sounds: a caller who says more gives the agent more to work with.
Lending and financial services
Fundamento: rescuing stalled loan applications. Fundamento runs voice agents for Indian lenders and insurers, including Hero FinCorp, IIFL, and ICICI Prudential, across more than 20 million outbound calls a month. For Hero FinCorp, the agent calls borrowers who go idle for more than five minutes during a digital loan application. Often they’ve stalled because the form is in English and they aren’t native speakers. The agent works out the problem in the borrower’s language and guides them back. 42% of those dropped-off applicants came back and completed their application.
Banking brings its own rules: verification before account data, card numbers kept out of the model, call-frequency limits for collections. Conversational AI in banking goes through them.
Sales and outbound calls
2X Solutions: compliance as a latency problem. In the US, the FCC’s telemarketing rules (47 CFR 64.1200) treat an outbound call as abandoned if the person isn’t connected within two seconds of finishing their greeting, and cap abandoned calls at 3% of a campaign over 30 days. 2X’s previous voice provider took 500–900ms to start speaking, which ate a big piece of that two-second window once the rest of the pipeline was added. After switching, its abandoned call rate fell below 1.25%, across about 3 million calls a month.
Thoughtly: a voice for the CRM. Thoughtly plugs into Salesforce and HubSpot to call, qualify, and route leads for more than 1,000 teams, including Coldwell Banker and Farmers Insurance. Live calls run at under 200ms end to end, and teams can clone a branded voice from 20–30 seconds of audio for sales demos. Outbound is where the first two seconds decide everything, which is why both of these examples are really latency stories.
If outbound is your use case, read AI cold calling before you dial anyone. The consent rules for AI voices are stricter than most teams expect.
Many languages, one agent
Bolna: voice agents across India. Bolna’s platform handles more than 500,000 calls a day for companies deploying agents in Hindi, Tamil, Telugu, and other Indian languages. Callers switch between Hindi and English mid-sentence, and read out order IDs and alphanumerics that have to come out right on a phone line. Better language quality opened new regions for Bolna’s customers, and cut the engineering time spent fixing latency and quality by about 40%. India has 22 official languages, and an agent that sounds natural in only one of them reaches a fraction of the callers.
Beyond the phone line
Not every conversational AI deployment answers a phone.
Tavus: video avatars you can talk to. Tavus builds AI replicas that see, hear, and respond in a live video call. Getting end-to-end response under one second, so the avatar replies while the conversation still feels live, depends on speech generation that starts in under 100ms.
Ego: game characters that talk back. Ego lets players create 3D game worlds in natural language. In Thrall, a mod for the survival game Valheim, players give an AI companion voice commands (gather wood, join a fight) and it answers out loud, celebrating a good hunt or acknowledging the order.
Two smaller examples: Rox gives sales reps a voice mode for asking about their accounts while they’re walking through an airport, and Quora’s Poe lets millions of users hear AI responses in 100+ voices and 14 languages.
What the working examples have in common
Read the twelve together and the same five traits show up:
- One job, with a system of record behind it. Scheduling, card support, application follow-up. The agent can look up the right answer instead of improvising one.
- A latency budget someone measured. 2X had two seconds by law, and Arini counted dropped calls in dollars. In both, speed was a business number, not a spec sheet line.
- A clear way to a person. Lorikeet’s agents worked in a crisis because callers could still reach someone when the agent couldn’t fix the problem. The warm transfer guide covers how to hand off without making the caller repeat themselves.
- The caller’s language, not just English. Fundamento, Bolna, EliseAI, and Lorikeet each won callers because the agent spoke the language they were most comfortable in.
- Outcome metrics, tested in production. Bookings, completed applications, tour conversions. EliseAI’s 1% rollout is the model: prove it on real calls before turning it on for everyone.
What’s missing from the list is just as telling. None of these is a general-purpose assistant that tries to do everything. The ones that work start narrow and grow one call type at a time.
Build your own
There are two ways to build one on Cartesia. Managed Agents runs speech recognition, the language model, speech generation, and the phone connection for you. You write the instructions, connect your systems as tools, and get a phone number. Or use the API: Sonic for speech, Ink for streaming transcription and turn detection, and your own orchestration in between. That’s the path platforms like Retell and Vapi take. The AI voice agent guide explains how the pieces fit together, and the customers page has every story in full.
To try one before you build anything, create an agent in the Cartesia playground and call it from your phone.