Skip to content
Sub-processors

Sub-processors

The third parties involved in running Tephlo, what each one is used for, which data reaches it, and where it operates. This list is part of the Data Processing Agreement.

Last updated 24 October 2026Revision Draft 1.2 — October 2026 refreshEffective 24 October 2026

In plain language

Running an AI assistant on WhatsApp and Telegram means other companies are involved: one generates the answers, two carry the messages, one hosts the database, one sends email, one stores files. Each is listed below with what it does, what it receives, where it operates and in what role, as configured on the live deployment on 9 October 2026.

  • Several are off by default — voice transcription, spoken replies, image reading, web search and error monitoring only apply if they have been switched on. On the live deployment the first three are available to a workspace that turns them on, web search applies only where a workspace allows it, and no error-monitoring project is configured.
  • No analytics, advertising or tracking company appears on this list, because none is used.
  • These providers operate internationally. We cannot promise data stays in one country, and this page says where each one is based rather than implying more precision than exists.

Core provider roles

These describe supported and documented configurations, not proof that each provider is active for every workspace. Endpoint settings and the operator’s provider agreements determine the actual recipients.

In use — all workspaces
ProviderPurposeData categoriesLocation
OpenRouterGenerating the assistant’s replies. Every AI answer goes through this provider, which routes the request on to the selected model provider.The customer’s message and recent conversation, the passages retrieved from the workspace’s knowledge base, and the workspace’s instructions to the assistant.Whatever a customer writes in a message can therefore reach this provider.United States, and the country of whichever model provider it routes to.
OpenAITurning knowledge base text and customer questions into search vectors, so an answer can be found in the right document.Knowledge base text uploaded by the workspace, and the text of customer questions used to search it.On the live deployment these requests travel through OpenRouter (the row above) to OpenAI’s embedding model.United States.
Meta Platforms (WhatsApp Business Cloud API)Receiving and delivering WhatsApp messages. Required for the WhatsApp channel.Customer phone numbers and the full content of messages in both directions.Global infrastructure operated by Meta.See the note on channels below — with your own WhatsApp number, Meta is your provider, not our sub-processor.
TelegramReceiving and delivering Telegram messages. Required for the Telegram channel.Telegram user identifiers and the full content of messages in both directions.Global infrastructure operated by Telegram.
CloudflareObject storage for files customers send and for brand assets; hosting for the public marketing, documentation and legal sites.Uploaded files and their extracted text; brand images. The public sites involve only ordinary connection data.Storage region set when the bucket was created: Cloudflare R2, automatic placement. Site hosting is distributed.
Heroku (Salesforce)Application hosting, the managed PostgreSQL database, and the managed key-value store used for queues and short-lived state.All application data at rest, including conversations, knowledge bases and accounts.Region chosen at provisioning: Europe (Ireland).The platform can also be self-hosted, in which case this row is replaced by the operator’s own infrastructure.
MailtrapDelivering operational email: signup verification, team invitations, password resets and escalation alerts.Staff names and email addresses, and the content of those emails.Escalation alerts carry no customer message content — they say a conversation needs a person and link to the console.Operated by Railsware Products Studio, LLC (United States), with infrastructure in the European Union and the United States. A workspace deployment may instead use its own SMTP provider.

Optional — off unless switched on

These are disabled by default. Each requires the platform to be configured for it and, where it affects conversations, the workspace to opt in. If they are off for your workspace, no data reaches them.

Optional — enabled per deployment or per workspace
ProviderPurposeData categoriesLocation
OpenAI (speech to text)Transcribing voice notes so the assistant can answer them.The audio of a voice note, and the resulting transcript.United States. The request travels through OpenRouter to OpenAI’s Whisper model.
Paystack (workspace plan billing)Hosted checkout, payment verification and supported renewals for workspace plans, where billing is enabled. This does not process payment for a customer’s saved cart.Payer email, amount, currency, payment references and transaction metadata. Payment details are entered directly with the provider; the platform can retain a reusable authorization reference and limited payment-method display details returned by it.Nigeria: Paystack Payments Limited, Lagos. Paystack is an independent controller for the payment itself, which the payer completes on Paystack’s own page; we receive the reference, status, amount, fee and the display details above.
OpenRouter, routing to Deepgram (spoken replies)Turning an approved assistant reply into audio, where spoken replies and a compatible speech endpoint are enabled.The approved reply text, selected voice and language-related settings, and the resulting audio. This can contain customer information already present in the reply.United States. The request goes to OpenRouter, which routes it to Deepgram’s speech model; both operate in the United States.
OpenRouter, routing to OpenAI (image reading)Interpreting customer images and, when separately enabled, reading scanned PDF pages without selectable text.Customer images or rendered pages of a customer document.United States. The request goes to OpenRouter, which routes it to an OpenAI vision model.
TavilyWeb search, for question types a workspace has allowed the assistant to look up.A search query derived from the customer’s question.United States.
SentryError monitoring — collecting crash reports so faults get fixed.Error type, stack trace and technical context. Message content, identifiers, tokens and addresses are redacted before logging.No error-monitoring project is configured on the live deployment, so nothing reaches Sentry today. If one is configured, this row will say where before it starts.
Telegram GatewayDelivering password-reset codes to a staff member’s phone, where they have chosen that method.A staff phone number and a short one-time code.Global infrastructure operated by Telegram.

A note on messaging channels

Who contracts with WhatsApp and Telegram matters. Most workspaces connect their own WhatsApp Business account or their own Telegram bot. In that case the provider processes under your relationship with them and their terms apply to you directly — they are not our sub-processor, and we cannot influence what they do with the message. Where a workspace uses a number the platform shares between businesses, the relationship is ours and the row above applies. Either way, the provider necessarily sees every message sent over its own channel.

What is deliberately absent from this list

  • No analytics provider. The marketing, documentation and legal sites load no analytics script — their content security policy blocks third-party scripts entirely.
  • No advertising or tracking network, no advertising pixels, and no data brokers.
  • No third-party font service. Public application scripts and fonts are served by the site. Published logos and icons may load from explicitly configured API or asset-hosting origins.
  • No cart-payment processing. The built-in potential-order flow records a customer request for business review. The optional plan billing provider above is used to pay for the workspace itself.
  • No AI model training on your data. Conversations are sent to a provider to generate a reply; they are not contributed to model training by us.

Changes and objections

Before a new sub-processor begins processing customer data, this page is updated and workspace administrators are notified by email, with 30 days’ notice. If you object on reasonable data protection grounds, section 7 of the Data Processing Agreement sets out what happens next.

Dated, not timeless. Provider locations and configuration change. The rows above describe the live deployment as configured on 9 October 2026; a new provider, or a change of region for one listed here, is published on this page with the notice the DPA requires before it takes effect.

Contact

General questions about this document go to hello@tephlo.com. Questions about personal data, including requests from individuals, should go to the data protection contact at hello@tephlo.com. A request about your own data — access, erasure, correction, objection or a complaint — can also be sent through the privacy request form, which puts it in front of the team as a tracked request rather than a message in a mailbox.

Postal address: TEPHLO AI VANGUARD SYSTEMS LTD, 9A Kenneth Lekara Luka Street, Lokogoma, FCT, Nigeria.