Sub-processors
We are Third Frontier Ventures Ltd. trading as Arcla (“we”, “us”, “our”). This page lists every third party that processes personal data on our behalf, what each one does, and where it processes that data. Our main privacy policy is at arcla.com/privacy-policy.
How to read this page
When your organisation uses Arcla, your organisation is the data controller for the calls it records. Arcla is the processor, acting on your instructions. The companies below are sub-processors: they process that data on our behalf. We require data processing terms with each of them that limit them to processing your data on our instructions and prohibit them from using it for their own purposes. For most providers this is their published standard Data Processing Agreement; for a minority that paperwork is still being completed.
There is no legal difference between a company that hosts our infrastructure and one we call over an API. Both process personal data on our behalf, and both are listed. A vendor is a sub-processor whether or not its staff ever look at the data: storing it is processing.
Two different things are listed separately, because they often disagree.
- Where the company is is the country the company itself is based in. It tells you which legal system the company answers to.
- Where it processes is the region your data is actually stored or handled in. This is the one that matters for data location.
The two often disagree. Sentry is a United States company that processes our data in the EU. Qdrant is a German company that processes ours in the United States. Neither is what the company name would suggest.
Where personal data is transferred outside the UK or EEA, we rely on UK International Data Transfer Agreements or EU Standard Contractual Clauses. Completing that documentation across every provider on this page is work in progress.
Data residency
We do not currently offer region locking. You cannot contractually require that your organisation’s data stays in a particular region.
In practice our processing is spread across regions rather than concentrated in one. Our primary database and cache are in the United Kingdom. Call recording, speech-to-text and AI inference are in the United States. Analytics and error monitoring are in the EU. Hosting runs on a global edge network with no region restriction configured. The tables below give the region for each service.
Formal data residency is on our roadmap. If it is a procurement requirement for you, contact us at security@arcla.com to discuss timing.
Infrastructure and hosting
| Sub-processor | Where the company is | Purpose | Data it handles | Where it processes |
|---|---|---|---|---|
| Cloudflare | United States | Application hosting, edge compute, object storage, call database, AI gateway, knowledge retrieval (AI Search), language models and embeddings (Workers AI), DDoS protection, cookieless aggregate web analytics on the marketing website | Call transcripts, knowledge base content, call metadata, website content, aggregate page-view and page-performance measurements | Global edge network. No region restriction is configured on our databases or object storage. Our answer pipeline runs in AWS us-east-2 (United States) |
| Vercel | United States | Dashboard application hosting | All dashboard request traffic, account data, approximate location derived from IP | Global edge network |
| Supabase | United States | Primary database: call records, participants, transcript snapshots | Call metadata, transcript extracts, account and profile data | United Kingdom (London) |
| Upstash | United States | Redis cache for rate limiting and short-lived OAuth state | Transient per-user state, rate-limit counters | United Kingdom (London) |
| Microsoft (Azure) | United States | Off-site disaster-recovery backups of our database and stored files | Encrypted copies of call records, transcripts, knowledge base files and account data. We encrypt them before upload and hold the only key, so Microsoft cannot read them | United Kingdom (London) |
Call capture and speech-to-text
| Sub-processor | Where the company is | Purpose | Data it handles | Where it processes |
|---|---|---|---|---|
| Recall.ai | United States | Desktop recording capture and upload; calendar and meeting integration | Call audio and video recordings, transcripts, participant details, calendar events, and the access tokens for any calendar you connect | United States (N. Virginia) for production recordings. An EU instance (Frankfurt) exists and is in validation, but is not yet used in production |
| AssemblyAI | United States | Speech-to-text for live call transcription | Call audio, resulting transcripts | United States |
| Deepgram | United States | Speech-to-text for live call transcription | Call audio, resulting transcripts | United States |
| Cartesia | United States | Speech-to-text for live call transcription | Call audio, resulting transcripts | United States |
We use more than one speech-to-text provider so that transcription continues if one is unavailable. Which one handles a given call depends on how the call was captured and on availability at the time.
AI model providers
Each of these receives extracts of call transcripts and knowledge base content in order to generate answers, detect questions, and analyse calls. We use several providers and route between them for performance and availability, so a given request may go to any of them. Requests to these providers pass through Cloudflare’s AI gateway, with two exceptions noted in the table: call title generation calls OpenAI directly, and vocabulary extraction calls Anthropic directly.
We do not train any models of our own on your data, and none of these providers is engaged to do so. We are working through explicit zero-retention and no-training terms with each of them, and that work is not yet finished. If it matters to your procurement, contact us at security@arcla.com for the current position on each provider.
| Sub-processor | Where the company is | Purpose | Data it handles | Where it processes |
|---|---|---|---|---|
| Anthropic | United States | AI inference: onboarding and manifest generation, script assistance, website classification, FAQ and vocabulary extraction. Through the gateway, except vocabulary extraction, which is called directly | Script and knowledge base content, organisation profile content | United States |
| OpenAI | United States | AI inference: in-call question detection and knowledge chat, post-call analysis and summaries, call title generation. Through the gateway, except call title generation, which is called directly | Transcript extracts, script and knowledge base content | United States |
| United States | AI inference: fallback for in-call assistant answers | Transcript extracts, knowledge base content | United States | |
| Cerebras | United States | AI inference: primary provider for in-call assistant answers | Transcript extracts, knowledge base content | United States |
| Together AI | United States | AI inference: fallback for in-call assistant answers | Transcript extracts, knowledge base content | United States |
| OpenRouter | United States | Text embeddings only, as a fallback when our primary embedding path is unavailable. We do not use it to route language models | Search queries derived from calls, knowledge base content | United States |
| Perplexity | United States | Web search, when the in-call assistant needs information beyond your knowledge base | A search query generated from the question asked on the call | United States |
Knowledge base and search
| Sub-processor | Where the company is | Purpose | Data it handles | Where it processes |
|---|---|---|---|---|
| Qdrant | Germany | Knowledge base storage and search | Knowledge base content and its embeddings, search queries | United States |
| Supermemory | United States | Knowledge base storage and search | Knowledge base content, your website content, search queries | United States |
| Pinecone | United States | Legacy knowledge base index from an earlier version of the product. It is no longer written to and is not used by the live retrieval path; retiring it and removing the residual data is in progress | Knowledge base content and its embeddings held from earlier use | United States |
| Firecrawl | United States | Reading your organisation’s website to build your knowledge base | Your website URLs and page content | United States |
Identity, operations, and support
| Sub-processor | Where the company is | Purpose | Data it handles | Where it processes |
|---|---|---|---|---|
| Clerk | United States | Authentication, user and organisation management | Name, email address, profile image, session data | United States |
| Cloudflare Turnstile | United States | Bot protection on sign-up, provided through Clerk | IP address and browser signals at sign-up | Global edge network |
| Inngest | United States | Background job orchestration | Call, organisation and user identifiers; waitlist email addresses | United States |
| Composio | United States | Connecting third-party knowledge sources you choose to link | Access credentials for the tools you connect, and content retrieved from them | United States |
| Attio | United Kingdom | Customer relationship management for waitlist and onboarding | Name, email address, and the answers you give when joining the waitlist | European Union (Ireland) |
| Cal.com | United States | Meeting booking on our marketing website | Name, email address, and chosen meeting time | United States |
Analytics, monitoring, and logging
| Sub-processor | Where the company is | Purpose | Data it handles | Where it processes |
|---|---|---|---|---|
| PostHog | United States | Product analytics and feature flags. Requests are routed through our own domain at p.arcla.com |
Dashboard usage events. With your consent, your name, email address, and account dates | European Union |
| Sentry | United States | Error monitoring and crash reporting, including session replays of the dashboard | Error reports with call, user and organisation identifiers; recordings of dashboard sessions, which may show transcript and knowledge base content | European Union |
| Axiom | United States | Application log management | Call, user and organisation identifiers; call metadata; redacted log previews | United States |
Advertising measurement on the marketing website
These three run only on arcla.com, never on the dashboard, and only after a visitor allows the analytics category (Google Analytics) or the marketing category (Google Ads, Meta, LinkedIn) in our cookie banner. They never receive customer call data. Each also uses what it receives for its own advertising platform under its own privacy policy, so for that use it is an independent controller rather than a sub-processor; we list them here so the full picture is in one place.
| Sub-processor | Where the company is | Purpose | Data it handles | Where it processes |
|---|---|---|---|---|
| United States | Website analytics (Google Analytics) and advert conversion measurement and retargeting (Google Ads), each once enabled | Pages visited on arcla.com, demo requests, advert click identifiers, cookie identifiers, IP address and browser signals | United States | |
| Meta Platforms | United States | Advert conversion measurement and retargeting for Facebook and Instagram | Pages visited on arcla.com, demo requests, advert click identifiers, cookie identifiers, IP address and browser signals; a one-way hash of the email address for a demo request reported from our servers | United States |
| United States | Advert conversion measurement and retargeting | Pages visited on arcla.com, demo requests, advert click identifiers, cookie identifiers, IP address and browser signals; a one-way hash of the email address for a demo request reported from our servers | United States |
Calendar providers
If you connect a calendar, we and Recall.ai call your provider directly to read your meeting schedule. These are your own providers rather than vendors we chose, but we list them for completeness.
| Provider | Purpose | Data it handles |
|---|---|---|
| Calendar and meeting access | Your calendar identity and events | |
| Microsoft | Calendar and meeting access | Your calendar identity and events |
What is not on this list
We use a number of services that never receive customer personal data: source control, continuous integration, code review and security scanning tools, error-free build and release tooling, and internal project management. These are not sub-processors and are not listed.
Changes to this list
We review this list at least annually and update it whenever a sub-processor is added or removed. Material changes are announced on this page with a revised date above.
To ask about our data processing terms, or to raise an objection to a sub-processor, contact us at support@arcla.com.