# PolyAI review (2026): What it is and why enterprises choose Bland
By Samuel Oram-Jones
PolyAI is a managed voice AI provider whose own team designs, builds, and runs each customer's phone agent, and it quotes a price instead of publishing one. This review covers how PolyAI works, what it costs, what its contract says on uptime, billing, and changes, and why enterprises often choose [Bland vs PolyAI](/compare/bland-vs-polyai), as of November 2026.
## What is PolyAI and how does it work?
PolyAI is a managed service for AI phone agents and chat agents, sold to large consumer-facing companies such as hotels, restaurants, insurers, and telecoms. PolyAI's team scopes a customer's build in a statement of work, then builds the agent, tunes it, and runs it afterwards.
A call on PolyAI can run on its own models or on models from other companies. PolyAI has a proprietary language model called Raven and its own Owl speech recognition, and a customer can also choose GPT, Claude, or Amazon Nova as a language model.
PolyAI uses seven speech recognition providers, including Deepgram, Google, and Amazon, and eleven voice providers, including ElevenLabs, Cartesia, and MiniMax. The companies behind a call therefore depend on a customer's language, voice, and model settings, which is the opposite of a full stack design like Bland's.
## How much does PolyAI cost?
PolyAI publishes no rate for a managed agent. It says ongoing use "is priced on a per-minute basis", with a "Talk to our team" form beside it, so a buyer learns the price on a sales call.
The one per-minute figure PolyAI prints belongs to Dialog-RSN-1, its model API: 2¢ per connected minute, with "Telephony and orchestration Not included".
Voice output is "Free until 19 Nov" with no price after that date, text is billed per token, and concurrent calls start at 2 and rise to 10 included once a card is added. That is a language model a developer builds around, and a team can't put a phone number on it.
Billing rules differ between PolyAI's two products. PolyAI says "Billed on connected time, with no minimums and no rounding up to the minute." Its managed-service agreement says the "total handle time is rounded up to the next 30-second increment", so a 61-second call bills as 90 seconds.
### How is Bland priced?
On self-serve plans, Bland Start is $0.14 per minute with no platform fee, and Bland Build is $0.12 per minute with a $299 monthly platform fee, which is lower than Start above about 14,950 minutes a month.
Either rate covers a language model, transcription, and speech with no token charges, and phone is billed separately at carrier cost. Enterprise pricing is custom, set with Bland's sales team.
Enterprise is the plan regulated teams buy. It adds native iMessage agents alongside SMS and web chat, warm transfers, guardrails, a BAA, SSO, data residency, dedicated orchestration with dedicated GPUs run by Bland, and forward-deployed engineers who work beside a customer's team, with concurrency contracted to its volume and custom pricing.
## How easy is PolyAI to use?
A customer edits its agent in Agent Studio, and developers get an Agents API plus the Dialog-RSN-1 model API. PolyAI says a build takes "about six weeks", while its contract treats a build as ready for testing on "the ninetieth (90th) day" even when PolyAI hasn't sent notice, then deems it accepted 14 days later.
Changes after launch run through the same contract. A change that, "in PolyAI's determination", needs extra fees or moves the timeline requires a signed Change Order, so a team that wants to edit its own agent quickly has to ask who can ship a change before signing. PolyAI also doesn't offer iMessage agents, so its agents reach customers by voice, web chat, and SMS.
### How reliable is PolyAI?
PolyAI advertises "99.994% uptime, trailing twelve months" and a model response under 300ms, voice to voice, while Bland publishes sub-200ms and offers a 99.9% uptime SLA on every plan. The advertised figure is a measured history, and the contract figure is a promise.
That contract promises "a minimum uptime of 99.5% measured on a monthly basis", and 99% when a use case runs over the public phone network. In a 30-day month those figures allow 3.6 hours of downtime, or 7.2 hours on the phone network.
Support follows the same pattern. PolyAI says a per-minute price includes "24/7 support", then its contract sets fix times for urgent tickets only. A ticket is urgent when "at least 25% of all calls are failing", lower-priority tickets carry "No guaranteed fix time", and phone support goes only to "a designated number of your contacts".
Language coverage adds a dependency. PolyAI's Raven handles 24 languages, and some others, Danish for example, "require a third-party LLM", so the model behind a call changes with the language. The company behind that model can change or retire it, and PolyAI's speech and voice availability is "Provider-dependent".
## Which sounds more human, PolyAI or Bland?
Bland Speech v3 ranked 1st of 13 AI voices at 1365 on Design Arena's Audio Realism Bench on 4 August 2026, and no PolyAI voice appeared in those results.
In Design Arena's blind test, vetted native speakers hear two voices read one script and choose the one that sounds more like a real person. The highest outside voice PolyAI uses in those results, Cartesia's Sonic 3.5, ranked 6th at 1096. See [Design Arena's published results](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/).
## Is PolyAI secure enough for enterprise calls?
PolyAI reports SOC 2 Type II, ISO 27001, and GDPR compliance, but its wording on medical and card calls is softer than its marketing. PolyAI says its healthcare offer is "HIPAA-compliant and BAA-backed by default", then says elsewhere "Where relevant, our systems are designed to meet HIPAA". A BAA is the contract HIPAA requires before a vendor handles patient information.
Its data processing agreement has the customer agree, "save as otherwise agreed in the Agreement", not to provide protected health information or cardholder data. Its acceptable use policy bars sensitive data "except as expressly agreed in writing by PolyAI", including payment card information and protected health information. A regulated call needs written sign-off before it reaches PolyAI.
Card payments follow the same pattern. PolyAI lists PCI DSS under "Certified" on its home page, then says "Where relevant, PolyAI is committed to complying with the PCI-DSS". Bland self-attests to PCI DSS v4.0.1.
Each outside speech and voice provider on a call is also one more subprocessor, a company that handles caller data for a vendor, so a security review grows with each one a setup uses.
## Is PolyAI worth it for enterprises?
For most enterprise calls, PolyAI isn't the best choice as of November 2026, because it quotes a price on a sales call, routes changes through contract terms, and sends calls to outside speech and voice companies. PolyAI suits a large consumer brand that wants a vendor team to build and run an agent and is comfortable signing a quote-only contract.
### PolyAI pros and cons
Pros
- A vendor team designs, builds, and runs the agent for a customer
- Raven, a proprietary language model, handles 24 languages
- SOC 2 Type II and ISO 27001 are both reported
Cons
- No managed-agent rate is published, so a price comes from sales
- A Change Order covers any edit PolyAI judges to need extra fees or time
- 99.5% uptime in the contract against "99.994% uptime" in its advertising
- Calls round up to the next 30 seconds under the managed-service agreement
- Speech, voice, and non-Raven language models come from outside companies
- No iMessage agents
Pros
- One AI company on every call, so a security team reviews one vendor
- One published rate for all of its AI, with no token charges
- Dedicated orchestration and dedicated GPUs run by Bland, on Enterprise
- Its own SIP and media servers, with no outside telephony provider in the call path
- Call grading on the audio as well as the transcript
Cons
- Overkill for a small prototype, because Bland is built for production calls at volume
What these rows mean
- **AI on each call:** This row shows whether the language model, speech, and transcription behind each call belong to the platform or to other companies: Full stack means all three are the platform's own, Outsourced means its language model and transcription come from other companies, and Mixed means it runs some of its own models alongside outside ones. Bland builds and runs its own language, speech, and transcription models, so a call never passes through an outside AI provider.
- **Audio Realism Bench score, 4 August 2026:** Audio Realism Bench is Design Arena's blind listening test, where listeners pick whichever of two voices sounds more like a real person, and a higher score means listeners picked that voice more often. In its published results of 4 August 2026, Bland Speech v3 ranked first among AI voices. Each rival number is the highest-scoring voice model in those results available on that platform.
- **iMessage:** iMessage lets an agent text customers on a channel they already use. Bland offers it natively on Enterprise.
- **SOC 2 Type II:** SOC 2 Type II is an independent audit of how well a company's security controls work over time.
- **ISO 27001:** ISO 27001 is an international standard for managing information security.
- **AI rate per minute:** Headline rates cover different things, so a buyer should check what each one includes before comparing them. Bland's self-serve rate covers all of its AI: language model, transcription, and speech. Phone is billed separately at carrier cost, and Enterprise pricing is custom, set with Bland's sales team. PolyAI quotes its managed agent per minute and publishes no rate.