Data Processing Agreement
Last updated: 4 October 2026
What changed on 4 October 2026
- Encryption at rest, isolated code execution and access logging are no longer claimed; we describe what we can show.
- Sub-processors and model providers are listed from the same source as the Privacy Policy, including China for our default chat model.
- Local models and model training are described in the same words as in the Privacy Policy.
- We no longer say a deleted account is deactivated at once and purged whole: we list what the 30-day sweep deletes, what it does not yet remove, and that sign-in still works until it runs.
- We disclose the keeper wallet behind cloud-run trading agents.
- Model providers are named; the other sub-processors are still listed by category.
- We no longer claim role and ownership checks on every record at the database layer.
- We no longer say a deleted account can be restored during the grace period.
- Psyche: we no longer say withdrawing a recorded consent stops the processing; the one consent we record is the sync consent.
- We no longer promise short-lived tokens: the sign-in token lasts up to 90 days by default.
- We no longer say each sub-processor is bound by terms no weaker than these: Article 28(4) flow-down is not yet in place.
- We no longer claim a written duty of confidence we cannot show you; ask us for the terms.
- The data export is described by what it holds: your account record, chat conversations, file list and security log.
- Audit logging is described for the endpoint that records erasure requests, not for every deletion.
- We list Alibaba Cloud's DashScope (China, where we have configured it), Deepgram and Perplexity among the model providers.
This agreement applies where you use Poly to process personal data for which you are the controller — for example, a business account whose staff put customer information into a chat, a workspace, or an agent. It forms part of the Terms of Service.
It is written to Article 28 of the UK and EU GDPR. Where Article 28(3) requires a contract to stipulate something, there is a section below saying what we do about it, in the same words we would use to describe the code.
Where this agreement does not apply
For your own personal use of Poly, you are not a controller and we are not your processor — we are the controller of that data and the Privacy Policy governs it. This document is for the case where you are answerable to someone else for data you put into Poly.
Roles and subject matter
- Controller: you. Processor: Polysystems, Inc.
- Subject matter: providing the Poly platform — chat, agents, workspaces, documents, research, mail and the other surfaces you use.
- Duration: for as long as your account is open, plus the deletion period described below.
- Categories of data subject: whoever you choose to write about. We do not control that and cannot enumerate it for you.
- Categories of personal data: likewise — whatever you submit. Prompts and uploads are free text and free files.
Special category data. Psyche data (assessments, therapy, journal, check-ins, games and readings) is health-adjacent and we treat it as Article 9 special category data. What reaches our servers, and the one consent we record, is set out here: Psyche — assessments, therapy sessions and messages, journal and check-ins, games and readings — is stored on your device. Copies reach our servers only after you turn on sync, with two exceptions: readings are produced by our servers and the conversation that produced each one is kept in your account's chat history; and a therapy message answered by a cloud model because a local model you chose was unavailable is kept there too. When you use an AI feature (a therapy message, a reading, a journal action), the text you send passes through our servers to our model provider. Turning on sync is the consent we record: it is saved in your browser and, when you are signed in, against your account. Using a Psyche feature without sync is not recorded as a consent. Turning sync off stops syncing, and deleting removes the copies on our servers. Do not put other people's special category data into Poly unless you have your own lawful basis for doing so.
(a) We process only on your documented instructions
Your instructions are: this agreement, the Terms, the settings on your account, and the requests you make through the product. We do not process your content for any other purpose.
We do not use your content to train AI models for us or anyone else. The one model trained on your content is a custom style model (LoRA) you ask for in Studio: it is trained, through Replicate, on the images you choose, for your use. Model providers process your requests under their own terms, which decide whether and how long they keep them; our list of model providers links each provider's terms.
If we ever believe an instruction of yours requires us to break the law, we will tell you rather than quietly comply or quietly refuse.
(b) Confidentiality
Access to customer content is limited to people at Poly who need it for support, security or debugging. For the confidentiality terms they are bound by, ask legal@poly.inc.
(c) Security
Measures currently in place:
- TLS in transit. For our hosting providers' encryption at rest, ask legal@poly.inc.
- Authentication by signed tokens: the API cookie lasts one hour, and the sign-in token kept in your browser lasts up to 90 days by default (signing out removes it from the browser); workspace API keys stored only as a SHA-256 hash, so a database dump does not yield a usable credential.
- Queries are scoped by user or workspace on the server, not filtered in the client.
- Keys for wallets you connect and for the Poly Credit wallet are never transmitted to us; a Credit seed is encrypted in your browser under a passphrase we have never seen. The exception is the keeper wallet behind cloud-run trading agents: If you run a trading agent in the cloud, it trades from a keeper wallet that Poly creates and holds for you: its private key is generated and stored, encrypted, on our servers. Trades from a keeper wallet run only while we have switched keeper execution on for the service.
- Rate limiting on our API routes.
- Audit logging of security-relevant events, including consent changes, data exports and erasure requests made through our data-request endpoint.
We do not currently hold ISO 27001 or SOC 2 certification. If a certificate is a procurement requirement for you, we do not meet it today and would rather you knew now.
(d) Sub-processors
You give general authorisation for us to engage sub-processors. We remain responsible for what they do with your data. They process it under their own published terms (each model provider's is linked below). We have not yet put in place, with each of them, terms that pass on the obligations in this agreement as Article 28(4) requires; that is a gap, next to the transfer gap below.
| Category | Who | What they receive |
|---|---|---|
| AI model providers | The model providers listed below | The prompt, context and files for the request you made, and the response. |
| Media generation | Image, video, music and speech generation providers | Your generation prompt and any source media. |
| Payments | Stripe | Billing details and subscription events. |
| Hosting and storage | Cloud compute, database and object storage providers | The data we store to run the service, and data in transit to and from it. |
| Operational infrastructure | Rate limiting, caching and email delivery | Your user id and request metadata; for email, your address and the message. |
| Analytics | Vercel Web Analytics | The page-view data listed in our Cookie Policy. Page addresses can contain ids, such as the id of a chat or document. |
| Data sources | Web search, market data, blockchain RPC and research APIs | The query or address you asked about. |
| Messaging | Telegram, where you connect it | The messages your agents send you. |
The model providers we send your requests to, and where each says it processes them:
| Provider | What for | Where | Its terms |
|---|---|---|---|
| DeepSeek | Our default chat model | the People's Republic of China, per its privacy policy | Privacy policy |
| OpenRouter | Describing images you attach, and other models you choose | the United States and other countries, per its privacy policy | Privacy policy |
| Anthropic | Chat models you choose | the provider's own region — see its terms | Privacy policy |
| OpenAI | Chat models you choose, and speech transcription | the provider's own region — see its terms | Privacy policy |
| Z.AI | Reading text in images and documents, and speech | Singapore, per its privacy policy | Privacy policy |
| ElevenLabs | Speech | the provider's own region — see its terms | Privacy policy |
| Replicate | Training a custom style model (LoRA) in Studio | the provider's own region — see its terms | Privacy policy |
| Alibaba Cloud (DashScope) | Chat, research and agent models, where we have configured it | the People's Republic of China (its Beijing region), per its Model Studio documentation | Privacy policy |
| Deepgram | Transcribing call and meeting recordings uploaded in Workspace, where we have configured it | the United States, per its privacy policy | Privacy policy |
| Perplexity | Answering web-research queries in Research | the provider's own region — see its terms | Privacy policy |
Model providers are named above; the other sub-processors are listed by category, and that is a real limitation. Article 28 contemplates your being told which sub-processors we use so that you can object to one, and a categorical list does not let you do that. If you need the names in the other categories and a change-notification commitment in writing, ask legal@poly.inc and we will provide them for your account.
Local models. When you choose a local model (Ollama or OMM), Poly's server sends your message to the local model runtime configured for the service. If that runtime is unavailable, your message is sent to a cloud model provider through our servers instead; in Chat, a notice tells you so.
(e) Helping you answer data subjects
Product features do part of this without needing us in the loop: from Settings you can export your account record, chat conversations, the list of files you stored and your security log, delete individual items, and delete the account. Where a request needs something the product cannot do, write to privacy@poly.inc and we will help within the statutory period.
If a data subject contacts us directly about data you control, we will not answer on your behalf — we will point them at you and tell you it happened.
(f) Helping with breaches, impact assessments and consultation
We will notify you without undue delay after becoming aware of a personal data breach affecting your data, with what we know at the time rather than waiting for a complete picture. You are the one who has to notify a regulator within 72 hours, so a late notice from us is a failure of ours.
We will give you the information you reasonably need for a data protection impact assessment or a prior consultation with a supervisory authority.
(g) Deletion and return
From Settings you can export your account record and chat conversations, and delete the account from Settings > Account. For any other data, write to privacy@poly.inc.
Deleting an account from Settings signs it out everywhere at once. After a 30-day grace period an automated sweep permanently deletes its conversations, research, flows, memories, knowledge graph, synced files and Psyche data. The sweep does not yet remove the account record (email, username, password hash and profile photo), connected-account tokens, workspace content such as office documents, agents and call and meeting recordings, keeper wallets and their keys, login history, two-factor secrets, consent records, or usage and billing records; until it does, write to privacy@poly.inc to have them erased. Until the sweep runs the account can still be signed in to, and signing in does not cancel the deletion. Backups are kept on a rolling cycle so we can recover from failure; deleted data leaves the backups as that cycle replaces them.
Two exceptions, and both are limits on what we are able to do rather than choices:
- Billing and tax records are kept for as long as tax law requires — typically seven years.
- Anything written to a public blockchain is permanent and outside anyone's deletion right, including ours. Nothing we do removes it.
(h) Information and audits
We will make available the information you need to verify that we are meeting these obligations, and respond to security questionnaires.
Audit here means answering, not admitting. We do not currently offer on-site inspection or third-party audit access to our infrastructure, and we hold no third-party audit report to offer in its place. Article 28(3)(h) contemplates both; we meet the first half. If your own regulator requires more, tell us before you rely on this document.
International transfers
Poly operates from the United States and our sub-processors are located in the United States and elsewhere, including China for our default chat model and for Alibaba Cloud's DashScope where we have configured it, so data from the EEA or the UK is transferred out of it.
We do not have counter-signed Standard Contractual Clauses in place with you, and we have not completed a transfer impact assessment. This is the largest gap in this document and it is stated here rather than buried: if you are exporting EEA or UK personal data and need a documented Chapter V transfer mechanism, we cannot yet give you one. Write to legal@poly.inc before you rely on Poly for that processing.
Agents you publish or sell
If you publish an agent, other people can run it. If you sell one through Poly Credit, buyers' prompts reach your agent and you become a controller or processor of that data in your own right — this agreement does not cover you in that direction. The obligations that would apply to you are set out in the Poly Credit seller agreement.
Liability, order of precedence and changes
The limitations of liability in the Terms apply to this agreement. Where this agreement and the Terms conflict on the processing of personal data, this agreement wins.
We will give notice before changing this document in a way that reduces your protection. Questions: dpo@poly.inc.
Not legally reviewed
This was written from what the software does, by the people who built it, and has not been reviewed by a lawyer. It is offered because an accurate description of our processing is more useful to you than a template describing somebody else's — but have your own counsel read it before relying on it, particularly the transfers section.