PolyAI

What it is

PolyAI is an enterprise voice AI platform that answers the phone call itself — booking a table, authenticating a caller, taking a payment, rescheduling an appointment — rather than steering the caller into chat. Founded in London in 2017 by three Cambridge dialogue-systems researchers, Nikola Mrkšić (CEO), Tsung-Hsien Wen, and Pei-Hao Su, it runs on its own speech stack rather than stitching third-party ASR and TTS together: the house dialog model, Raven, is trained on more than a billion enterprise conversations, with GPT-5, Claude, and Gemini available as alternates for the reasoning layer.

It raised an $86M Series D on 15 December 2025, co-led by Georgian, Hedosophia, and Khosla Ventures, with NVIDIA’s NVentures, Citi Ventures, Zendesk Ventures, Point72 Ventures, Sands Capital, Squarepoint Ventures, and the British Business Bank participating. SiliconANGLE reported the round at a $750M valuation, up from $500M in May 2024, taking total funding past $200M. The company reported over 2,000 live deployments across 100+ enterprises at the time of the round, and by May 2026 put its reach at 75 languages across 25 countries. Named customers include Marriott, Caesars Entertainment, PG&E, UniCredit, Foot Locker, FedEx, Fogo de Chão, Howard Brown Health, and Quicken.

Why it shows up in CX and support stacks

• It integrates with the systems the call is actually about. Every AI voice vendor connects to Salesforce and a CCaaS platform. PolyAI’s documented catalog goes past that into the operational systems the caller is phoning about: Epic, athenaOne, ModMed, Cerner (Oracle Health) and Raintree in healthcare; OpenTable, Tripleseat, HotSOS and Cendyn in hospitality. An agent that can read a chart or move a reservation is doing different work from one that can only read a CRM note, and this is the shortest list in the category that covers both.
• The managed-service model is no longer the only way in. Through 2025 the standing critique was that you did not build PolyAI agents — PolyAI’s team did. That changed twice in 2026. The Agent Development Kit, launched 22 April, is a CLI and Python package with Git-backed local development and a poly push deploy step, so agents live in your repo and go through your CI. On 18 May the company opened the platform to self-serve signup with the first two months free, alongside Poly Agent Builder, a natural-language authoring surface.
• It owns the audio path, and shipped a model that shows why. Dialog-RSN-1, released 30 July 2026, is audio-native: turn-taking, speech recognition, function calling, and response generation run in one model instead of a pipeline of four. PolyAI reports sub-300ms responses on A100 GPUs, a 37% latency cut at an insurance customer, and an 11% relative containment gain at a national restaurant group.
• Compliance is priced in, not bolted on. SOC 2, HIPAA, GDPR, and PCI DSS come as standard per the vendor, with PCI Pal and Stripe available as managed payment paths — the reason the healthcare and utilities logos are on the list at all.
• Agents can call MCP servers as tools. The integrations docs cover MCP for custom tooling, so an internal service you already expose over MCP is reachable from the voice agent without a bespoke connector. Note the direction: this lets the agent use your tools, not you query PolyAI’s analytics from Claude.

Pricing reality

Nothing published. The pricing page states one fact — usage is billed per minute — and routes every other question to a demo. There is no published rate, minimum, or overage figure, and no free tier beyond the two-month self-serve trial opened in May 2026.

Third-party buyer analyses put annual contracts starting around $150K and rising with volume, concurrency, language count, and integration depth; treat that as an estimate, not a quote. The same analyses note that telephony, CRM integration, and compliance work typically arrive on a separate invoice, so the platform number is not the first-year number. PolyAI’s own commissioned Forrester Total Economic Impact study claims a 391% ROI and an average $10.3M in savings — vendor-funded research, useful for a board deck and not for your model.

What matters operationally is that a per-minute meter tracks average handle time. An intent the agent resolves in 80 seconds costs a third less than the same intent at 120 seconds on identical call volume. Model it as minutes × containment rate, and get the per-minute rate, the annual minimum, the overage rate, and shortfall rollover in writing before you compare it against an outcome-based meter.

Best for

Enterprise CX and contact-center leaders in phone-dominant, regulated industries — healthcare, utilities, financial services, hospitality, and multi-site restaurant and retail groups — whose callers are asking about a record held in Epic, athenaOne, OpenTable, or HotSOS rather than in the CRM. The ROI band is best above roughly 500,000 annual voice minutes, where a few points of containment repay a six-figure commitment, and it is the strongest pick in the category when the integration list decides the project.

Skip it if chat carries most of your contact volume, if your first-year budget cannot absorb a six-figure commitment, if you need a published price you can approve without a procurement cycle, or if none of your systems of record appear in the integration catalog — in that last case you are buying custom API work at enterprise voice-AI rates.

Versus the alternatives

Start with what you already pay for. If you run NICE CXone, Genesys, or Five9, their native AI is the incumbent and adds no vendor to onboard; take it when its containment on your top intents clears your bar, which is usually the answer when your call reasons are routing and FAQ rather than transactions. Among AI-native voice platforms, Parloa is the closest like-for-like and the better-capitalized one — a $350M Series D in January 2026 at a $3B valuation, $50M+ ARR, and DORA compliance that PolyAI does not advertise. Pick Parloa if you are an EU financial-services buyer or your contact center is outsourced to Teleperformance, Concentrix, or Foundever, which already deploy it; pick PolyAI when the deciding factor is a healthcare or hospitality system of record.

The fastest-growing entrant is Sierra, at roughly $200M ARR by May 2026 — the pick when the agent must transact across both chat and voice and you will accept outcome-based billing and a larger vendor to get it. Decagon wins when chat deflection is the job, Ada when you want channel breadth and an invoice decoupled from agent performance, and Cresta when you are keeping human agents and want them coached rather than replaced.

If none fit, run the self-serve trial against your three highest-volume intents and keep humans on everything else. A voice agent bought before those intents are mapped produces a shorter path to the queue, not a resolved call.

Watch-outs

• The headline latency number belongs to a model you may not get. Dialog-RSN-1 is in limited availability and English-only; existing customers can enable it, new customers request early access. The sub-300ms figure and the 37% latency cut both come from it, while a multi-language deployment runs on the older pipeline. Guard: name the serving model, its language coverage, and its availability date in the order form, and benchmark the pilot against the model that will actually carry your traffic — not the one in the launch post.
• Per-minute billing charges you for latency and for a verbose agent. The meter that protects you from arguing about what counts as a “resolution” hands you the opposite exposure: slow responses and long prompts both show up on the invoice. Guard: instrument average handle time per intent from week one against your human baseline, fix the per-minute rate for the contract term, and write in a review trigger if AHT crosses a named threshold rather than assuming tuning will happen.
• The self-serve trial is not the enterprise product, and the integration work is where the money goes. Two months free with an email address exercises Agent Builder, not an Epic connector, a Genesys SIP path, or a PCI-scoped payment flow. Guard: scope and price the integration and compliance work as a separate line before signing the platform contract, and get written confirmation of which connectors are productized versus custom build — the third-party analyses agree that this is the invoice that surprises buyers.
• Filed revenue sits a long way below the valuation. UK accounts show $15M revenue for the year to 31 January 2025, up from $8.9M, with losses widening to $26.6M — against a $750M valuation set that December. Deloitte ranked it the UK’s fastest-growing AI business in 2025 on 248% three-year growth, so the trajectory is real, but a company priced for growth pushes multi-year terms and volume floors. Guard: cap the first term at one year, request the current SOC 2 Type II report and a signed BAA before contracting where PHI is in scope, and keep your agent definitions in your own Git repository via the ADK — that repo is your only cheap exit, and it is worth confirming in the agreement that you retain it.