← Blog

Together AI Compliance Requires a Route-Level Approval Record

Parminder Singh
Parminder Singh··6 min read
Summarize with AI

Together AI compliance for serverless inference depends on the customer account, approved API route, training choice, Zero Data Retention setting, and evidence attached to each request. Together documents the provider-side terms. The customer must prove which identities and data classes were permitted to use the route.

AI Security Solutionsai-complianceai-governancecomplianceauditpolicy-enforcementidentity-and-authorization
Together AI Compliance Requires a Route-Level Approval Record

Together AI compliance starts with the serverless inference request. An authenticated caller submits content through an API route, and Together returns an output. The customer's selected privacy settings govern the provider's handling of that exchange. The approval record should name the account and route, then connect them to the permitted data classes and training choice.

Together's two public records cover provider duties. Customers need separate evidence showing that a specific user or agent was authorized to send a specific request. I would reject any review that stops at a screenshot of the provider logo on an approved-vendor page. That image says nothing about the route carrying production data.

TL;DR

  • Approve Together's serverless inference API as a named route with an account owner and evidence location. Record the permitted identities and data classes.
  • Together says customer data stays out of model training unless the customer explicitly opts in and consents.
  • Zero Data Retention takes effect prospectively through Privacy & Security settings, so preserve activation evidence and test requests made after activation.
  • Keep customer authorization records separate from Together's provider security statement and monitoring rights. Keep them separate from reliability and confidentiality terms.

Together AI compliance begins with the inference route

Together's Privacy Policy, updated December 17, 2025, covers programmatic APIs and web interfaces. This approval should stay narrower: the serverless inference API route selected by the customer. Fine-tuning, GPU clusters, the web playground, and Together's internal operations require separate assessments because they introduce different activities and evidence owners.

Create one route record for each approved production path. Name the Together account and customer owner. Record the API endpoint and permitted calling application. Add the user or agent population and allowed data classes. Add the training choice and ZDR state. Link the record to the applicable contract and provider policies. Then identify where request evidence will live.

The AI vendor risk assessment template can hold the provider review. The route record proves which customer-controlled path received approval and which restrictions apply when traffic enters that path.

Training requires an explicit customer choice

Together states in its Privacy Policy that it will not use customer data to train models without explicit opt-in and consent. Preserve that statement with the policy date, then pair it with a dated export of the customer's actual account selection.

The approval should identify who can change the choice and who reviews it. Attach the current setting and review date to the account identifier. If the customer later opts in, treat that as a new data-use decision with a named approver, affected route, effective date, and permitted data scope. Keep the prior state so an auditor can reconstruct which rule applied to an earlier request.

Use the AI data classification guide to express route restrictions in operational terms. A customer may permit public documentation. It may block secrets or regulated records. The provider's training statement never decides which customer material belongs in the request. That decision remains with the customer.

ZDR activation has a clear time boundary

Together says customers can enable Zero Data Retention prospectively through Privacy & Security settings. Under ZDR, submitted content and outputs are removed after processing. They are excluded from retention and training, as well as product improvement and secondary purposes beyond providing the service.

Once ZDR removes material, Together says it cannot retrieve or correct it. It also cannot export or delete it. The customer therefore needs its own request evidence before relying on the provider for later reconstruction. Store controlled metadata or references according to the customer's policy. Avoid assuming Together can recover the submitted content after processing.

Capture the account page immediately after activation with the account identifier and setting. Include the timestamp and reviewer ticket in the same frame. Then preserve the first approved request made after that moment. I prefer this paired record because it ties a configuration decision to the traffic that followed it instead of treating a settings page as timeless proof.

The Anthropic Zero Data Retention acceptance guide describes the same evidence principle for a different provider. Keep the provider terms separate while reusing the route-level acceptance method.

Account and API responsibility stays with the customer

Together's Terms of Service, updated May 19, 2026, says Together supplies APIs and access tokens while customers provision and manage their user accounts. The customer approval should translate that allocation into named operating duties. Assign an owner for the Together account and token issuance. Assign responsibility for user access review and the application route.

A valid token proves access to the provider endpoint. Customer authorization evidence must answer a narrower question: was this caller permitted to use this route for this data class under the policy active at that moment? Record the application-supplied identity or agent identifier and approved route. Add the data classification and policy version. Keep the outcome and timestamp. Keep the provider token out of that evidence record.

The Terms also say Together may monitor use for agreement compliance. It may also monitor for operational quality and improvements. Treat that statement as a provider-side activity covered by the contract and policy review. The HTTP-layer policy enforcement guide shows where the customer can make and record its own route decision before forwarding a request.

Contract claims need their own evidence file

Together states that it maintains a security program for customer content. Preserve the current Terms and applicable order form with the provider assurance file. Assign customer route authorization to the operating record, with the permitted caller and data class attached to the request decision.

Reliability commitments require an order form under the standard Terms. Record the signed order form when the customer depends on a reliability promise. The standard Terms also say no confidentiality obligations apply unless the parties agree to them separately in writing. A security review should therefore locate the separate written confidentiality agreement instead of inferring one from the public Terms.

These contract points belong beside the procurement decision. Customer route authorization belongs in the operating record. Keeping those files separate prevents a signed document from being presented as proof of a production request, while preserving the contract evidence that governs provider commitments.

The boundary stops at routed inference traffic

This control scope covers authenticated serverless inference requests deliberately routed between customer users or agents and Together's LLM endpoint. Fine-tuning and GPU clusters remain outside it. The Together web playground also needs its own approval because a browser session follows a separate customer path. Provider monitoring and security operations retain Together as the evidence owner.

The customer owns identity and route approval. It also owns data classification and request-policy evidence. Downstream use of an output remains with the customer application. Direct traffic that bypasses the approved route falls outside the route record and requires separate detection and access controls.

Draw the boundary on the architecture page. Put the approved HTTP route in a solid box. Mark the playground and fine-tuning outside that box. Put cluster services and provider operations outside it as well, each with its own owner. A reviewer can then see which evidence this approval supplies and where another control must answer.

DeepInspect

DeepInspect sits inline on authenticated HTTP traffic deliberately routed between customer users or agents and LLM endpoints. For an approved Together serverless inference route, it can evaluate application-supplied identity and destination. It can also evaluate data classification and the current customer policy before forwarding the request.

Each decision produces customer-owned evidence with the caller and route. The record also contains the policy version, outcome, and timestamp. That record supports the customer's authorization decision while Together's Privacy Policy and Terms remain the source for provider-side handling. DeepInspect does not control Together's ZDR setting, training choice, security program, monitoring, order form, confidentiality agreement, or internal operations.

Book a technical deep dive at deepinspect.ai.