Skip to main content
usedby

What happens on a PolyAI call, from greeting to human handoff

Follow one inbound call through PolyAI: phone system, caller checks, a booking change in your systems, the handoff to an agent, and the data involved.

18companies identified by usedby

11with a case study, testimonial or press source

7on the vendor’s customer list

Public evidence snapshot: Nov 7, 2026

Our read

Plan a PolyAI deployment as four connections rather than one product: your telephony, the system that holds the booking or account, a source of answers, and the contact-center platform that receives transfers. The documentation lists connectors for each. Two points the sources leave open, the rules that trigger a transfer and who approves live changes, are what a pilot should settle.

12

An inbound-call agent, built in Agent Studio

PolyAI is a platform for building and running voice agents that answer inbound calls for enterprises. Teams configure the agent in Agent Studio, a no-code builder, or in code through a developer kit, and usage is priced per minute. For a buyer, the useful question is less what the agent says than what it is wired to. A single call touches four things: the phone network that delivers it, the system holding the booking or account, a source of answers, and the place where a human takes over.

The examples below come from case studies published by PolyAI: Atos, SafeRide Health, Carter's, Fogo de Chão, Audibel and Hawksmoor. Each describes one deployment, with figures from the vendor and its customers. None proves what a different company will get.

12345678

Greeting: connecting your number to the agent

The integration documentation gives several ways in. You can forward an existing number to a new PolyAI number, or connect any SIP-capable platform through a SIP trunk. Named telephony options include Five9, NICE CXone through its Signal API, Twilio, Amazon Connect, Genesys through Cloud BYOC, Zendesk Talk, Dialpad and Zoom. The choice matters beyond the greeting, because it decides where transfers can land later. Settle it first.

Customers show the range. Carter's lists Salesforce and Five9, while SafeRide Health lists NICE, Twilio and AWS. Atos replaced a touch-tone menu that routed calls to 440 agents with a voice assistant for FAQs and routing. Carter's describes an agent that opens with a plain invitation to say what the caller needs, instead of a menu tree.

1345

Understanding the caller, and what you may feed the agent

Understanding means intent plus identification. Audibel's agent collects a name, phone number and zip code, works out whether the caller wants an appointment or product information, then passes the call on. SafeRide Health's agent authenticates members before ride changes. According to its case study, the first, intent-based version struggled with natural speech and was rebuilt on an agentic design. PolyAI's security page says it authenticates through conversation rather than voice biometrics.

One restriction sits right beside this step. The data processing agreement says customers should not provide restricted data, including card data, health information and national ID numbers, unless their agreement says otherwise. Audibel's story calls its agent HIPAA-compliant and the product overview lists payments, yet the sources do not explain how those fit with that clause. Before you design an authentication step, ask which identifiers your contract lets the agent collect.

247910

Action: changing a booking in your systems

The documentation says the agent can look up, create and modify reservations during calls. Hospitality connectors include OpenTable, Tripleseat and HotSOS, Salesforce covers customer records and cases, and Gladly or other knowledge bases supply answers. Hawksmoor connected the agent to OpenTable so callers see the same table inventory a reservationist sees. Fogo de Chão also integrated OpenTable and, by its own account, launched a pilot within two weeks.

Where no connector exists, you build one. You define HTTP endpoints, set a base URL and authentication per environment, and call them from Python functions. MCP servers are a second documented route. SafeRide Health took the custom path: its agent was built on the company's own APIs to book, cancel and modify rides. So a call to change a reservation needs telephony, an authentication step, a read of the reservation, write permission on the booking system, and a confirmation by voice or SMS.

Two limits apply. Hawksmoor describes a caller-profile lookup by phone number, through OpenTable's guestbook API, as work in progress, so do not assume the agent recalls preferences. Audibel lists appointment scheduling as a next phase, so its example shows intake and routing, not booking. Editorial recommendation: in a pilot, script one modification that hits a sold-out slot and one where the write to the booking system fails, then listen to what the caller is told.

14678

Handoff: who receives the call, and with what

Transfers and live-agent handoffs sit in the telephony group of the documentation, while chat and SMS handoff reaches Salesforce, Zendesk, NICE CXone, Amazon Connect, Genesys, Webex or Five9. Audibel describes a warm handoff that carries the intake details, so the human starts with context. Carter's reports more than 3,000 live-agent transfers with no queue in its chat figures. SafeRide Health reports its speak-to-human rate falling from 55% to 2-3%, while keeping transfer for callers who want or need a person. Atos reports a 30% drop in agent call volumes, with complex queries routed to skilled agents.

Each figure has its own metric and population, so they cannot be compared with one another. The sources here do not state the rules that trigger a transfer, such as repeated misunderstanding, a request for a person or a sensitive topic, nor what happens when no agent is free. Ask for those rules in writing, and test them with a caller who asks for a person right at the greeting.

13457

Supervision, data terms and going live

The homepage lists guardrails, PII redaction, human-in-the-loop controls and data residency in the US, UK, EU or APAC. ISO 27001 appears on the security page, and a data processing agreement is published. Under it, the customer owns its service data, but PolyAI keeps a licence to use communications data such as recordings and transcripts to create analytics and improve its services. The homepage separately says customer data is never used for training without opt-in. The sources do not reconcile the two, so ask which applies to you. The agreement also sets 48-hour incident notice, 15 days' notice of sub-processor changes and a data copy on request within 90 days after the contract ends.

Day to day, Hawksmoor's team edits menu items and closures in Agent Studio, and its analytics showed callers dropping out when card details were requested, which led to removing that step for parties of two to four. SafeRide Health says flow changes that took weeks now take seconds. Atos included security checks in an eight-week build. Who approves a change before it goes live is not described in these sources, so make it a pilot question.

2348910

Method and limits

2026-11-07

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.

Named companies, tool by tool

Public evidence snapshot: Nov 7, 2026

Logos of the companies our public sources name, strongest proof first.

Companies identified by usedby

ToolCase studies, testimonials, pressVendor listsTotal
PolyAI11718
Distinct companies per tool and proof level, as of Nov 7, 2026. Source: usedby, public sources.
See the documented examples

This snapshot describes our coverage, not market share or product quality. Missing evidence does not mean non-use. Dates record our checks, not when adoption started.

Sources

  1. Introductiondocs.poly.ai
  2. PolyAIpoly.ai
  3. Atos · PolyAIpoly.ai
  4. SafeRide Health · PolyAIpoly.ai
  5. Carter's · PolyAIpoly.ai
  6. Fogo de Chão · PolyAIpoly.ai
  7. Audibel · PolyAIpoly.ai
  8. Hawksmoor · PolyAIpoly.ai
  9. Securitypoly.ai
  10. Dpapoly.ai