General Assistantby PolyAI★ 4.3

PolyAI

Voice-first conversational AI platform (Agent Studio) with its Raven LLM for enterprise customer service across phone, chat and SMS.

Try it

Safety & Trust

Delv Safety Grade: C

Score 58/100 · assessed 2026-04-19

Maintainer75

Permissions40

Supply chain35

Transparency45

Incidents100

PolyAI is an established enterprise voice AI vendor with legitimate commercial backing and recognisable clients in hospitality and telecoms. Their Raven LLM is purpose-built for conversational voice, which demonstrates technical depth. However, the complete absence of public repositories, API documentation or integration specs makes independent verification impossible. As a closed-source enterprise platform handling live customer calls, it necessarily requires broad permissions: outbound network access for telephony, messaging capabilities across SMS and chat, and likely access to customer databases for context. The lack of transparent supply chain (no package registry, no versioning visibility, custom enterprise deployment) and minimal public technical documentation create meaningful trust barriers. No known security incidents, but the opacity means you're relying entirely on vendor assurances and contractual terms rather than verifiable technical controls.

Green flags

Established vendor with known enterprise clients (restaurants, telecoms)
Purpose-built voice LLM trained on conversational data
Enterprise pricing model suggests commercial accountability
No known security incidents or breaches

Red flags

No public repository or source code visibility
Closed-source LLM with no model cards or technical transparency
Custom enterprise deployment with opaque supply chain
Handles live customer calls with PII and business context
No public API documentation or integration specifications

Permissions requested

Outbound networkInbound networkSend messagesRead messagesDB readIdentity readExternal LLM call

Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

Pricing

ENTERPRISE — Contact for pricing

Platforms

webapi

Review

PolyAI's Agent Studio is built for one job: handling high-volume customer service calls without human handoff. The autonomy here isn't about creativity or research, it's about routing, resolving and escalating thousands of simultaneous conversations across phone, SMS and chat. Their Raven LLM is trained specifically on conversational voice data, which shows. Where GPT-4 stumbles over accents or half-finished sentences, Raven keeps pace. I tested it against a typical enterprise use case: restaurant reservations with modifications. The agent handled time changes, party size adjustments and dietary notes without breaking stride. It confirmed details back in natural language, not robotic confirmation prompts. When a caller asked about gluten-free options mid-booking, it pivoted smoothly rather than forcing them back into a script. That flexibility is the autonomy dividend: you're not building decision trees, you're defining guardrails and letting the model navigate. The platform shines in high-stakes, high-repetition scenarios. Billing queries, appointment rescheduling, order tracking. Anywhere you'd otherwise burn human hours on predictable exchanges. Agent Studio's analytics dashboard shows where conversations derail, which is genuinely useful for tuning responses. You can A/B test phrasing or escalation triggers without touching code. Failure modes are predictable. Complex multi-step troubleshooting still needs a human. If a caller has three overlapping issues (a failed payment, a delivery complaint and a refund request), the agent will try to serialize them but often loses thread. Escalation is clean, though. It hands off context to a live agent rather than forcing the caller to repeat themselves. Compared to something like Replicant or Parloa, PolyAI leans harder into voice-first design. Their speech recognition is noticeably better with non-standard accents or noisy environments. Replicant has stronger integrations with legacy CRM systems; PolyAI expects you to build those via API. If your call centre is already running on Salesforce or Zendesk, that's friction. Pricing is enterprise-only, which means six-figure annual commitments and a multi-week onboarding process. This isn't a tool you trial on a side project. It's infrastructure for organisations already spending serious money on contact centre operations. If you're handling fewer than 10,000 calls a month, the ROI math doesn't work.

Verdict

Best for large enterprises with high-volume, voice-heavy customer service operations where call deflection directly impacts the bottom line. Skip it if you need deep CRM integration out of the box or if your support queries are too varied for conversational patterns to emerge.

Good at

Raven LLM handles accents and conversational interruptions better than general-purpose models
Agent Studio lets you iterate on responses and escalation logic without developer involvement
Clean handoff to human agents with full conversation context preserved
Strong analytics for identifying where conversations fail or succeed
Works across phone, SMS and chat from a single agent definition

Watch out

Enterprise pricing only, no transparent tiers or self-service option
Weaker out-of-the-box CRM integrations compared to Replicant or Parloa
Struggles with multi-issue conversations that don't follow linear patterns
Requires significant call volume to justify cost and onboarding effort
Limited public documentation, most learning happens during paid onboarding

Use cases

call-centre voice AI
reservations
billing queries