PolyAI

PolyAI builds voice AI agents that handle phone-based customer service for large enterprises: call routing, authentication, billing, bookings, and troubleshooting, deployed across industries including utilities, banking, hospitality, retail, and healthcare. The core product is Raven, a proprietary dialog model the company says was trained on more than a billion enterprise conversations, with turn-taking and agent behavior built into training rather than layered on with prompts. The platform also lets teams plug in external LLMs (GPT-5, Claude, Gemini) for reasoning while PolyAI's own stack handles the voice-specific parts (turn-taking, latency, telephony).

In July 2026, PolyAI released Dialog-RSN-1, which ingests caller audio directly instead of first transcribing speech to text — a different architecture from the transcribe-then-generate pipeline most voice AI vendors use. PolyAI says this lets the model pick up non-speech cues and has driven its lowest measured latencies, typically under 300ms, and that it's already running on live production calls. This was independently covered (not just republished from a press release) by SiliconANGLE, MarkTechPost, CMSWire, and Tech Times, all describing the same audio-native architecture. The latency figure and "already handling live production calls" claim are PolyAI's own; the independent coverage confirms the architecture and launch, not the numbers.

The live pricing page (poly.ai/pricing) states only that "ongoing use of the voice agent is priced on a per-minute basis" — no rate, no tiers, and every call to action leads to "request a demo." Separately, PolyAI announced on May 18, 2026 (via its own press release and covered by CMSWire, SD Times, and others) that it was opening the platform to self-serve signup at studio.poly.ai with two months free, installable from the terminal via an open-source CLI. As of this review (August 2026), that free-trial offer does not appear in the text of the current pricing page — it may have been a launch-window promotion rather than a standing offer. A buyer should confirm directly at studio.poly.ai whether the free period is still live rather than assuming it from the May announcement.

Founded in 2017 as a University of Cambridge spinout, headquartered in London, founded/led by CEO Nikola Mrksic (previously worked on Apple's Siri). In December 2025, PolyAI raised an $86M Series D co-led by existing investors Georgian, Hedosophia, and Khosla Ventures, with NVentures (Nvidia's VC arm), Citi Ventures, and Zendesk Ventures also participating, at a $750M valuation (up from $500M in May 2024) and over $200M raised to date. Per Forbes, audited revenue for the 12 months to January 2025 was $15M (up from $8.9M the year before), with an expected ~$40M ARR for 2025 and reported net losses of $26.6M for the 12 months to January 2024. This puts PolyAI meaningfully behind Sierra ($10B valuation, ~$100M ARR) in scale, ahead of Decagon ($1.5B valuation) on revenue efficiency per Forbes' framing, and it's raised at what Forbes characterized as a comparatively conservative revenue multiple next to Bay Area peers.

PolyAI names large customers including PG&E, UniCredit, FedEx, Marriott, Foot Locker, Caesars Entertainment/Golden Nugget, Metro Bank, and Fogo de Chão. Its PG&E case study (published by PolyAI, poly.ai/customers/pge) claims 67% call containment against an initial 3.5% target, 35,000 labor hours saved, a 22% CSAT increase on outage calls, and roughly 4 million calls a year with 41% resolved by the AI agent. These are PolyAI's own figures, not independently audited, though the scale and specificity (a named utility, a stated original target the deployment beat) is more concrete than most vendor case studies.

PolyAI shows up on both G2 and Gartner Peer Insights, with search-engine aggregation reporting roughly 12 reviews on G2 (all reportedly 5/5) and 4.7/5 across 23 reviews on Gartner. Both G2 and Gartner blocked direct automated access during this review, so these figures could not be independently confirmed against the primary page and should be treated as approximate. Recurring praise across sources: natural-sounding voice quality and responsive implementation support. Recurring criticism: weaker built-in analytics for understanding what callers actually asked, and occasional latency/lag.

PolyAI's security page (poly.ai/security) textually confirms ISO 27001 certification. It also displays several compliance badge images without accessible labels or names in the underlying page text, so specific claims like SOC 2 or PCI DSS could not be independently verified from that page alone. Separately, PolyAI's healthcare page states its agents are "HIPAA-compliant" — this is a vendor claim; no Business Associate Agreement terms or third-party attestation were found publicly. There is no dedicated trust center subdomain (trust.poly.ai does not resolve; poly.ai/trust 404s) and no public status page (status.poly.ai does not resolve).

PolyAI's Agent Development Kit (github.com/polyai/adk) is Apache 2.0 licensed, has 82 GitHub stars, and is under active daily development — releases were published on August 13, 14, 17, and 18, 2026 alone, with the latest version 0.44.3. It's a local, git-like CLI for pulling, editing, and pushing Agent Studio projects (install via uv tool install polyai-adk). This is developer tooling around the platform, not the voice AI product itself, which remains closed-source and hosted.