The Voice AI agents landscape

Answers the practice or health-system phone line. Scheduling, rescheduling, intake, referrals, refill requests and triage routing, integrated into the electronic health record so the appointment actually gets written.

What to look for. Ask what share of calls are resolved without a human, and demand to see the appointment written into the EHR during the demo rather than described. Note how similar the descriptions below read. That is the category telling you something.

Agents that call outward on the provider's behalf: payers, pharmacies and other practices. Benefit verification, prior authorisation, claim status and denial follow-up. Hold-music drudgery converted into software.

What to look for. Measure this on completed determinations per week, not on call volume. Ask specifically how the agent copes when a payer changes its phone tree overnight.

Listens to the clinician and patient talking and produces a structured, billable note. It does not take a turn, ask a question or act. Filing these alongside conversational agents, as both reference maps do, is a category error.

What to look for. Clinical evidence and write access to the record are the moat, not transcription quality. Ask for published outcome studies and the named health systems running it in production.

Non-diagnostic clinical conversation at scale: post-discharge follow-up, chronic care check-ins, pre-operative screening and medication adherence. The agent initiates the call, which changes the safety and consent picture entirely.

What to look for. Ask what the agent is forbidden from saying, and how that boundary is enforced in the model rather than in the prompt.

First notice of loss, claims status, policy servicing, renewals and enrolment. Heavily scripted and heavily regulated, which makes it unusually well suited to voice because the caller already expects a structured conversation.

What to look for. Check how disclosures and recorded-consent language are enforced per state or jurisdiction. That is where these deployments fail audit, not where they fail technically.

Debt collection, payment arrangement, loan servicing, KYC follow-up and card disputes. Collections is the highest-volume voice use case in the sector and the one with the clearest measurable return.

What to look for. Compliance is the moat here. Ask how the agent recognises a dispute, a cease-communication request or a bankruptcy mention, and what it does in the same second.

General-purpose customer support across voice and chat, sold on resolution rather than deflection. The best-funded corner of the market: three companies here took roughly 91 percent of first-half 2026 category funding.

What to look for. Insist on resolution rate rather than containment rate, and get the definition of resolution in writing before the pilot starts.

The platforms that already run the contact centre and have added an AI agent to it. Their agent is often weaker in isolation, but the routing, workforce management, quality monitoring and reporting are already integrated and already bought.

What to look for. The question is never whether the challenger agent is better. It is whether the incremental resolution rate justifies a second vendor, a second integration and a split reporting surface.

Voice agents that run the first-round interview or screening call: availability, right to work, salary expectation, basic competency, then a scored write-up into the applicant tracking system.

What to look for. The applicant tracking system owns the candidate record, so ask what happens when it ships this natively. Also ask how bias testing is run and how often.

Answers the phone for trades, clinics, salons, law firms and any small business where a missed call is a lost job. Screens, quotes, books into the calendar and texts a summary. Sold on a converted missed call, not a headcount saving.

What to look for. Price it against twenty of your actual calls including overage, then try to break the agent yourself before signing.

Two different businesses filed under one label. Autonomous outbound dials and speaks on its own. Assisted dialing keeps a human on the line and uses AI for parallel dialing, notes and coaching. Enablement tools run roleplay against a simulated buyer.

What to look for. The two have completely different legal exposure. Only assisted dialing is currently safe at scale in the United States.

Drive-thru order taking, phone orders and reservations. Operationally the most mature category on this map: AI order takers process orders faster than humans and run every hour the restaurant is open.

What to look for. Order accuracy and upsell attach rate are the only two numbers that matter. Deepgram's acquisition of OfOne in January 2026 signals where this consolidates.

Service department scheduling, parts enquiries, sales follow-up and missed-call recovery, wired into the dealer management system where the appointment and the repair order actually live.

What to look for. Tekion shipped Service Advisor AI and expanded its agent suite at NADA 2026. Ask any third-party vendor what happens to them when the dealer management system ships this natively.

Leasing enquiries, tour scheduling, maintenance intake, renewals and rent collection for multifamily and property management, plus agent-side lead qualification in residential brokerage.

What to look for. Entrata already ships voice inside the property management system. Standalone vendors here need a reason the incumbent cannot copy within two quarters.

Carrier check calls, load negotiation, dock appointment scheduling and driver support. Enormous repetitive call volume and, unusually, no dominant system of record that has shipped voice yet.

What to look for. A thin category, which here is a genuine opportunity rather than a warning. Ask about integration with the transport management system you already run.

Non-emergency municipal lines, 311 services, benefits enquiries and 911 call-taker assistance. The agent triages, translates and documents while a dispatcher stays in control.

What to look for. Procurement cycles and certification are the moat, not the model. Expect a long sales cycle and a very short list of qualified vendors.

Intake and reception for law firms, and reservations, concierge and guest requests for hotels and restaurant groups. Both hinge on capturing a qualified enquiry outside business hours.

What to look for. In legal, ask how the agent avoids anything that could be read as advice. In hospitality, ask about integration with the property or reservation system.

Indian-market voice agents for BFSI, telecom, healthcare and commerce. The binding constraints here are different: language mix and mid-sentence code-switching, accent depth, carrier-level integration, DLT registration, and DPDP consent and residency.

What to look for. Ignore the language count on the slide. Insist on a live call in your actual language, with natural code-switching, on a real mobile network, with background noise.

The platforms that own the write path: the calendar, the record, the capacity, the customer history. A conversation that does not write here has produced nothing. These platforms are now shipping their own voice agents, with live availability that a third-party integration can only approximate.

What to look for. This is the single largest threat to the application layer above. Eight of the eighteen application categories on this map are marked GETS ABSORBED because of it.

The unglamorous core. Turn-taking, barge-in and interruption handling, conversation state, function calling, retries, warm transfer, telephony binding and observability hooks. This layer decides whether the call feels like a conversation or a phone tree.

What to look for. This is the layer buyers replace most often, so price the exit before you sign. LiveKit and Pipecat are missing from both reference maps because open source is not a fundable line item, yet LiveKit runs ChatGPT voice mode.

Audio in, audio out, with no human-readable text in the middle. Lower latency, and it preserves tone, hesitation and emotion that a transcript destroys.

What to look for. Not a peer of the three rows below it. A replacement for them. Choosing this eliminates the transcript, the vendor swap and most of your observability in one decision. Choose it deliberately, for reasons you can state.

Decides what to say and which tool to call. In a voice agent it has roughly 200 to 700 milliseconds of budget between recognition and synthesis, which is a completely different constraint from chat.

What to look for. Chat benchmarks are irrelevant if time to first token is slow. Benchmark the model on your own tool-calling schema, under load, at your concurrency.

Turns the model's text into audio the caller will tolerate for three minutes. Time to first audio and prosody under interruption matter far more than raw naturalness.

What to look for. The gap between a 40 ms and a 150 ms voice is real but small next to the gap between good and bad turn detection. Do not let a voice demo drive your architecture.

Converts caller audio to text in real time and, increasingly, decides when the caller has finished speaking. That second job, end-of-turn detection, matters as much as accuracy.

What to look for. Word error rate on clean English has plateaued at 2 to 3 percent across the top tier, and vendors themselves warn about benchmark overfitting. Test on fifty of your own recordings, not on a leaderboard.

Noise suppression, echo cancellation, voice activity detection and end-of-turn prediction. The capability that most determines whether a call feels like a conversation rather than a fight over the microphone.

What to look for. Has no box on either reference map, and one of them files Krisp as an application. It is already being absorbed into the orchestrators, which now ship turn detection natively.

Carries the audio. Numbers, SIP trunks, WebRTC sessions, concurrency and call records. It does not make the agent smart. It decides whether the agent can be heard, at what cost, and whether it is legal to dial.

What to look for. Has no box on either reference map. This is where per-minute economics, concurrency ceilings, carrier registration and data residency live. In India it is the first constraint a deployment meets, well before model quality.

Simulates thousands of synthetic callers before launch, then monitors and replays real traffic after it. The most under-bought layer in the stack and the one that separates a demo from a deployment.

What to look for. If a team cannot answer how it knows a prompt change did not break five hundred other call types, it does not have an evaluation strategy. Budget for this as a line item.

Consent capture before dialing, in-call AI disclosure, natural-language revocation, separate recording consent, retention limits, audit export and carrier registration. These are product requirements that must sit inside the call flow, not policy documents.

What to look for. No specialist vendor category has formed here, which means enterprises are carrying the exposure themselves. TCPA damages run 500 to 1,500 US dollars per call with no cap, EU AI Act Article 50 disclosure has applied since August 2026, and TRAI can disconnect unregistered Indian numbers within hours.