
If you have spent any time building voice AI for a UK business in the last twelve months, you have stared at these three tabs: Retell AI, Vapi and Synthflow. All three promise the same thing — a working voice agent in under a day, with speech-to-text, an LLM, text-to-speech and telephony bundled together. The truth is they drift further apart the deeper you look.
I have shipped production voice agents on all three in 2025 and 2026 for UK clients — dental practices, property firms, a logistics company, a couple of e-commerce shops. This guide is the honest trade-off map, with the stuff you only learn by actually running calls through them at 11pm on a Friday when something breaks.
The short version
Best all-round for agency-grade builds. Cleanest call-flow builder, deepest analytics, excellent UK voice options when paired with ElevenLabs.
Pick when: you want polish and you bill a client for the work.
Developer-first. Best SDK, clearest logs, most predictable webhooks. Slightly less hand-holding in the UI.
Pick when: you are the engineer and you want control.
Friendliest UI. Non-developers can ship. CRM integrations out of the box. Marginally cheaper at volume.
Pick when: an ops person will own day-to-day changes.
How these platforms are actually built
Before comparing the three, it helps to understand what a voice AI call actually involves. Each one is a four-layer pipeline. The platform picks a default for each layer and lets you swap most of them.
| Layer | What it does | Common choices |
|---|---|---|
| Telephony | Connects a phone number to the pipeline | Twilio Voice, Telnyx, Vonage |
| Speech-to-text (STT) | Converts caller audio to written text in real time | Deepgram Nova, AssemblyAI, Whisper |
| Large language model (LLM) | Generates the reply based on conversation + knowledge | GPT-4o, Claude Sonnet, GPT-4o-mini, Groq Llama |
| Text-to-speech (TTS) | Converts the reply text back to natural voice | ElevenLabs, Azure TTS, Play.ht, Cartesia |
| Orchestration | Holds the conversation context, turn-taking, tool calls | Retell, Vapi, Synthflow |
The orchestration layer — that top-level coordinator that decides when the AI speaks, when to listen, when to call a tool, when to hand off — is the job Retell, Vapi and Synthflow all do. Where they differ is how much they force your choices in the other four layers and how much they let you tune.
Feature-by-feature comparison
| Feature | Retell AI | Vapi | Synthflow |
|---|---|---|---|
| Visual call-flow builder | |||
| Bring-your-own LLM | |||
| Bring-your-own STT | |||
| Bring-your-own TTS | |||
| ElevenLabs integration | |||
| Twilio number porting | |||
| Native webhook events | |||
| Node.js SDK quality | Good | Excellent | OK |
| Python SDK | |||
| Call recording & transcript | |||
| Sentiment / intent tagging | beta | ||
| CRM integrations (native) | HubSpot, Salesforce, GoHighLevel | Via webhook only | 40+ native connectors |
| Calendar booking (native) | |||
| Tool / function calling | |||
| Call hand-off to human | |||
| EU data residency option | |||
| UK data residency | |||
| SOC 2 Type II | |||
| HIPAA available | |||
| Free tier | 60 min/mo | 10 min/mo | 50 min/mo |
| Starter price (UK, Apr 2026) | $39 + per-min | $0 + per-min | $29 + per-min |
= fully supported · = with workaround · beta = beta · = not available. Prices exclude VAT, converted from USD.
Real pricing — the bit the marketing pages hide
All three quote a cheap-looking base price and then charge you per-minute for what actually matters. The sensible way to budget is in per-minute blended cost: telephony + STT + LLM tokens + TTS combined. Plug your minutes into the slider below.
Monthly cost at your real call volume
Per-minute pricing covers speech-to-text, LLM tokens, text-to-speech and telephony. Drag to see where you land.
Retell AI — the agency favourite
Retell sits in the middle of the developer-friendliness spectrum and, in our experience, produces the best end-client experience. The visual call-flow builder is properly designed: branching logic, tool calls and hand-offs all live in the same canvas and the UI keeps up as flows get complex. It is the one I reach for when I am building something for a client rather than for myself.
Where it shines
- Native ElevenLabs integration. A British voice that sounds like a human sits six clicks from a fresh account.
- Post-call analytics are the best of the three — sentiment per call, topic clusters, completion rates, hand-off reasons.
- Branded telephony numbers in minutes. Twilio and Telnyx both work without a separate account.
- The docs are current. I cannot overstate how rare that is for a platform this young.
Where it bites
- The webhook payloads are occasionally inconsistent — a call that ends via hand-off fires a different shape from one that ends on its own. Minor, but it catches you out once.
- Built-in CRM connectors are limited. Anything outside HubSpot, Salesforce and GoHighLevel means writing a webhook handler.
- Latency on first-word response creeps up with long system prompts. Keeping the prompt under ~1,500 tokens helps.
Vapi — the engineer's pick
Vapi is what you pick when you trust yourself more than a vendor's UI. The SDK is clean, the dashboard is honest about failures (it shows actual errors rather than "something went wrong"), and the cost breakdown per call is itemised down to the cent. When I am debugging why a voice agent is failing, I am doing it from Vapi's logs not its dashboard.
Where it shines
- Best-in-class Node.js and Python SDKs. Any engineer who has touched Stripe will feel at home.
- Webhook events are stable and versioned. You can count on the shape not changing without notice.
- Tool calling works identically to OpenAI's function-calling. No proprietary DSL to learn.
- Great for agentic flows — one agent handing off to another agent mid-call is a first-class concept here.
Where it bites
- The visual builder lags behind Retell. Complex flows are more comfortable in code than in the UI.
- Built-in analytics are sparse. You are expected to ship transcripts off to your own warehouse.
- Onboarding assumes technical literacy. A non-engineer will feel lost within twenty minutes.
Synthflow — the non-developer's friend
Synthflow is the only one of the three that your marketing manager can realistically maintain without involving a developer. The template gallery is practical (recruiter screening, restaurant booking, pest-control triage, dental confirmations) and forty-plus CRM connectors work without a single webhook. For solo operators and small teams without engineering, it is the obvious starting point.
Where it shines
- Genuinely usable by non-developers. That is rare in this space.
- Largest native integration library. Booking, CRM and messaging platforms mostly work click-and-save.
- Slight cost advantage at higher volume once you hit the committed-minute tiers.
- Template marketplace is a time-saver. Most starting flows are 80% of the way there.
Where it bites
- Custom logic beyond the templates gets awkward quickly. You eventually find yourself writing JavaScript inside the UI, which has worse developer ergonomics than writing it in your own editor against Vapi's SDK.
- STT quality on regional UK accents is noticeably weaker than Retell or Vapi with Deepgram Nova.
- Webhook reliability had two noticeable wobbles in late 2025. Better now but worth monitoring.
Which one to pick — by scenario
| If you are building… | Pick | Why |
|---|---|---|
| An agency-delivered receptionist for a dental or legal client | Retell | Polish, analytics, end-client confidence |
| A bespoke AI that hands off between two or three specialised agents | Vapi | Agentic flows are first-class; SDK supports it cleanly |
| A small e-commerce booking-and-returns bot managed by a marketing team | Synthflow | Non-technical ownership, Shopify connector built in |
| An internal tool that calls patients to confirm appointments | Retell | Outbound call support + GDPR-friendly EU hosting |
| A high-volume AI front desk for a call-centre-scale operation | Vapi | Most predictable at load, cleanest per-call cost accounting |
| A prototype to show a client next week | Synthflow | Fastest from blank to live demo, fewest footguns |
| Anything regulated where UK data residency is non-negotiable | None — use LiveKit self-hosted | All three route through US or wider EU infrastructure |
Switching between them
Good news: the pieces you actually spend time on — system prompt, tool schemas, knowledge base, voice choice — move between platforms with minimal rework. I have migrated two clients from Retell to Vapi and one from Synthflow to Retell. In each case the conversation quality was effectively identical after a day's tuning. The platform is the cheapest part to change; the hard work is in your prompt, your tools and your analytics.
UK-specific gotchas
- Ofcom call recording disclosure — if the AI is recording (and all three do by default), a clear in-call notice is required under UK regulations and GDPR. Every platform offers a starter script; check you are using it.
- Barge-in behaviour — UK callers are quicker to interrupt than US callers. Set interruption sensitivity higher than the default or you will feel slow.
- Post-code handling — ask for the post-code digit-by-digit and confirm readback. Standard numerical STT butchers alphanumeric UK post-codes otherwise.
- 0345 / 0800 / 03xx prefixes — test inbound on each, not just on geo numbers. Call routing through Twilio occasionally mis-handles non-geographic UK numbers.
- VAT — the platforms list prices in USD ex-VAT. Budget +20% if your business cannot reclaim it, and remember the FX.
Frequently asked questions
The honest verdict
If you are an agency or technical founder starting a new UK voice-AI build in April 2026, default to Retell for client work and Vapi for anything where you want full engineering control. Pick Synthflow when a non-developer has to own it day to day. None of them are wrong choices for the mainstream use cases — the differences only start to matter past the first month of production traffic.
Want us to build and host it instead?
We run voice AI for UK businesses on whichever of these three (or LiveKit, for regulated work) fits the brief. Fixed-fee build, parallel-run migration, a named human on WhatsApp when something goes wrong at 6pm on a Sunday.
Related reading
AI receptionist UK pricing comparison · AI receptionist UK complete guide · Voice AI complete guide · How to choose an AI receptionist


![How to Choose an AI Receptionist [2026]](/_next/image?url=%2Fimages%2Fblog%2Fchoose-ai-receptionist-comparison.jpg&w=3840&q=75)