Skip to content

Security and governanceChecklist

A user asks to delete their data: what must an AI gateway erase?

An erasure request does not stop at your database. A prompt can sit in gateway logs, a response cache, a provider's retention window, a vector store and a trace backend. This checklist maps each location to its owner, its deletion route and the evidence you can keep.

By INFRO Editorial

Published

8 min read

A grid of dark glass cards with small bar charts, a few of them lit amber, on a dark ground

In short

A gateway must erase whatever it stored for that person: optional payload logs and their files, cached responses, derived datasets and exports. It cannot erase provider-side copies, so those need a separate request under your provider agreement. Prove each step with a dated record.

Key takeaways

  • Inventory every store a prompt can reach before the first request arrives; an inventory built during an erasure deadline will miss something.
  • A gateway can erase only what it stored: payload logs and their files, cached responses, datasets and exports. Provider-side copies need their own request under your provider agreement.
  • Retention windows that expire on their own are a deletion route, but a window you cannot shorten per person means you need a targeted delete for the early request.
  • Tag each request with an end-user identifier at call time, or you cannot find that person's records afterwards.
  • Keep a dated erasure record per location; a statement that logging was off by default is not proof for a period when it was on.

Start with the map, not the request

When a user asks you to delete their data, the first question is not "what is in our database". It is "where could a prompt or response from this person have been written down since they started using the product". An AI gateway adds several places to that answer, and each has a different owner.

This checklist walks those places in the order a request travels. The EUR-Lex text of the GDPR returned nothing when I fetched it, so I do not quote article numbers here. Have your privacy counsel confirm the legal basis, exemptions and deadline for your case. What follows is the engineering side: what exists, who can delete it, and what evidence you can keep.

The scope of your own role matters too. INFRO's EU compliance page says an organization may act as a controller or as an authorized processor, and that provider agreements must match your actual processing arrangements.1 Decide which you are before you answer the user, because it changes who must act on the request.

The map: where a prompt can persist

LocationWritten whenWho deletesEvidence to keep
Gateway payload logOnly if content logging is onYou, in gateway settings or by vendor request; otherwise expirySetting history for the period, deletion confirmation, retention window
Files beside a log record (images, audio, video)With the logged requestDeleted with the recordSame as the record
Response cacheOpt-in, on a cacheable answerYou, or expiry at the time-to-liveCache settings, TTL, purge confirmation
Gateway metadata (usage, cost, ids)Every requestVendor policy; may be kept for accountingWhat fields it holds, the stated retention basis
Datasets, evals, scores built from logsWhen someone saves themYouDataset list before and after
Exports, SIEM and OTel destinationsWhen pushed or pulledYou, in your own systemsDestination inventory, deletion tickets
Provider retentionAt the provider, per your agreementThe provider, via your contractProvider ticket and written reply
Vector storeWhen your app embeds and indexesYouIndex ids removed, re-query test
Your app's own logs and chat historyIn your codeYouTicket per system
Places a prompt or response can persist, who deletes it, and how to show it.

The rest of this article takes these in turn. Everything marked "you" is a task for your team, because the gateway does not know which of your downstream systems copied the data.

Gateway payload logs

Check whether content was ever captured. INFRO's content logging is off by default and turns on only when an owner or admin enables request logging; a per-request logging: false disables capture even when the organization has it on.2 So the first fact to establish is the history of that setting: was it on, for which period, and which halves were recorded (prompts, responses or both).2

  • Pull the setting history for the user's whole active period, not just today's value.
  • Record which halves were captured, since a prompts-only organization has no stored answers.
  • Identify the retention window in force. INFRO's plans cap optional payload retention at 24 hours, 14, 30 or 90 days depending on plan, and the stored window is the smaller of the organization's choice and the plan maximum.2
  • Remember attached files. Images, audio and video a recorded request carried are kept beside the record with the same retention and deleted with it.2

Two behaviors matter for the proof step. Content is never returned after its retention deadline even if deletion has not run yet, and shortening the retention setting applies to content already captured.3 That second point is a useful lever: if a user's request is the only thing you need to purge, shortening the window is a blunt tool that affects everyone, so ask the vendor for a targeted deletion and keep the reply.

Do not mistake the request-history window for a purge. INFRO states that the queryable-history window describes access through request APIs and exports, and is not a promise to erase all underlying metering records at that date; usage aggregates and records required for accounting, security or law are kept separately.2 Say so in your reply to the user rather than claiming everything is gone.

The response cache

A response cache is a deliberate second copy of an answer. INFRO's is opt-in and exact-match: an identical request inside the time-to-live is answered from the cache and never reaches your provider.4 The entry is keyed on the organization, project and key environment, the model or route, the guardrail policy, the messages and the generation parameters.4

That key tells you how to search it. A cached answer written for one person's prompt is only returned for an identical prompt, but the stored text still exists until it expires. The plan sets the longest time-to-live: 1 hour on Developer, 24 hours on Startup, 7 days on Team and 30 days on Business and Enterprise.4

  • Find out whether the cache was on for any project the user touched.
  • Note the TTL, which is the longest an entry can outlive the request.
  • Ask what the purge function removes. The cache page is described as covering purge and privacy, so read that section and test it against a known entry before you rely on it.4
  • If the answer contains the user's personal data and the cache serves other users, consider excluding such routes from caching altogether.

For the wider risk of serving one person's answer to another, see the semantic caching article; an exact-match cache is narrower but not free of it.

Provider retention is a separate request

The gateway cannot reach into a provider's systems. INFRO's own docs say its logging settings do not change provider-side retention.1 Your prompts were processed under your account with each provider, including every fallback target a failed request was retried on, so the list of providers to ask is the list of connections the route could use, not only the primary one.1

What you can ask depends on the contract, not the consumer privacy page. Anthropic's privacy policy states that it does not apply to content processed on behalf of customers of its business offerings, whose handling is governed by customer agreements.5 OpenAI's enterprise privacy page says a DPA can be executed for the API, and its 30-day deletion statement is written for ChatGPT workspace products rather than the API.6 I did not confirm API retention periods on either page. Read the current API data-control documentation and your DPA for the figure, and do not infer it from the workspace wording.

Provider prompt caching is a related copy. INFRO passes it through, and says caching happens at the provider and that cached entries live with the upstream that served the request.7 Ask the provider how long cached prefixes live and whether they are covered by their deletion process.

  • List every connection a route may have used for this user's traffic.
  • Send each provider a deletion request referencing your organization and the time window.
  • Save the provider's written reply with the date; that is your proof, since you cannot inspect their storage.

For which providers offer zero-retention terms and how to verify them, see the zero data retention article.

Derived copies that people forget

Several features create copies that outlive the log. INFRO datasets are built from logged requests and kept until you delete them, and evaluator verdicts and scores attach to the requests they judged. Those are your derived stores; deleting the log record does not remove a dataset row someone saved from it.

  • Search datasets and saved evaluation sets for the user's content.
  • Check console history. INFRO Studio keeps generations for a window the plan sets and deletes them after it; the Playground's history stays in the browser, so it is out of the gateway's reach.
  • Inventory exports. A CSV from the request log, a SIEM stream, or spans pushed to your own OpenTelemetry collector are now in systems you own and must clean yourself.
  • Check vector stores. INFRO's embeddings endpoint does not store vectors, and its inputs follow your payload-logging setting, so the vectors live in your index.2 Delete by document id and re-query to confirm.

Your own application logs belong here as well: the chat history table, error trackers and analytics events often hold the same text under a different retention rule.

For tenant separation in indexes, see the multi-tenant RAG article.

Make the next request findable

Deleting is only possible if you can find the person's records. Put an end-user identifier on every request at call time. INFRO can group cost by an end-user id you attach, which doubles as a lookup key afterwards. Use a stable pseudonymous id rather than an email address, so the key itself is not another copy of the personal data you are about to erase.

Reduce what is written in the first place. Gateway guardrails can block, redact or flag patterns such as email addresses, phone numbers and card numbers before the provider sees them, though detection can miss things. Redaction narrows the problem; it does not remove the need for the inventory above.

Prove it, and keep the proof

An erasure answer is a set of dated statements, each tied to a system. A workable record per request holds:

  1. The identifier used to search, and the date range searched.
  2. For each location in the map, the action taken, the person who took it and the timestamp.
  3. For vendor-held copies, the ticket reference and the vendor's written confirmation.
  4. What was retained, with the stated reason (for example accounting records), as INFRO's docs describe for usage aggregates.2
  5. A re-check a few days later, since asynchronous deletion and cache expiry can lag the request.

The audit log records member, key and policy changes with an actor and a timestamp, which helps show who changed a logging or cache setting and when; it is not a record of data deletions, so keep your own. Finally, keep the erasure record free of the deleted content itself.

Pre-incident checklist

  • Data map of the nine locations above, reviewed each quarter.
  • Payload logging off unless a written purpose exists, with the window set as short as that purpose allows.
  • Response cache off for routes that carry personal data, or TTL set to the minimum.
  • Provider contacts and DPA references on file for every connection, including fallbacks.
  • End-user id attached to every request.
  • A template reply that states what is deleted, what is kept and why.

Questions readers ask

Can the gateway vendor run the deletion for me?
INFRO's EU compliance page says to contact support for assistance with access or deletion requests, and that the platform does not replace your organization's data-protection obligations.1 Treat vendor help as one step in your own record, not as the answer to the request.
Does turning logging off erase what was already captured?
Not by itself in every case. INFRO's trace docs say shortening the retention setting applies to content already captured, and that content is never returned after its deadline even if deletion has not run yet.3 Confirm the actual deletion with the vendor, and record the date.

From the team that builds it

Where this shows up in INFRO

INFRO is a control plane in front of the provider accounts you already hold. The capabilities below are the parts of it this article touches, each with the documentation that defines its behaviour.

  • Request traces and payload logs

    Available with a plan

    Per-request timing detail, with prompt and completion logging off by default and a retention window you choose.

    Read the documentation for Request traces and payload logs
  • Response cache

    Available with a plan

    Opt-in, exact-match cache for chat completions. An identical request inside the time-to-live you set is answered from the cache and never reaches your provider, so your provider does not bill it. Every plan; the longest time-to-live depends on the plan. No semantic matching.

    Read the documentation for Response cache
  • Embeddings

    Available with a plan

    POST /v1/embeddings in OpenAI's shape on your own OpenAI, Azure OpenAI, Mistral or OpenAI-compatible connection: routing that fails over only to the same model, budgets checked before the call, and the estimated provider cost of each request. Inputs follow your payload-logging setting; vectors are never stored.

  • Datasets

    Available with a plan

    Inputs and expected outputs built from logged requests in one click, uploaded as JSONL or CSV, or sent from code, kept until you delete them. On every plan.

  • Usage, request and FOCUS exports

    Available with a plan

    GET /v1/usage and GET /v1/requests on any gateway key, the request log and usage series as CSV, plus a FOCUS v1.4 Cost and Usage dataset as JSON or CSV, so AI spend lands in your FinOps platform beside the rest of your cloud cost.

    Read the documentation for Usage, request and FOCUS exports
  • OpenTelemetry export

    Available with a plan

    Spans pushed to your own collector from the request path, with the endpoint and headers set in the console. Included on every plan.

    Read the documentation for OpenTelemetry export
  • Spend attribution

    Available with a plan

    Group cost by model, project, key, or an end-user id you attach to the request.

    Read the documentation for Spend attribution
  • Guardrails

    Available with a plan

    Pattern-based checks on chat completion prompts and responses: block, redact or flag email addresses, phone numbers, card numbers, IBANs, US Social Security numbers, IP addresses, API keys and your own terms, plus optional classifier checks on your own connection to a decision model's provider. Detection can miss things. Included on every plan.

    Read the documentation for Guardrails

INFRO accounts are self-serve. Provider usage is billed to you by your providers; INFRO charges a software subscription and takes 0% of model spend. See plans.

Sources

Numbered in order of first citation. Each was opened and checked against the claim on the date shown.

  1. 1.EU compliance (opens in a new tab) — INFRO Docs. Accessed Nov 21, 2026.
  2. 2.Privacy & data (opens in a new tab) — INFRO Docs. Accessed Nov 21, 2026.
  3. 3.Request traces (opens in a new tab) — INFRO Docs. Accessed Nov 21, 2026.
  4. 4.Response cache (opens in a new tab) — INFRO Docs. Accessed Nov 21, 2026.
  5. 5.Privacy Policy (opens in a new tab) — Anthropic. Accessed Nov 21, 2026.
  6. 6.Enterprise privacy at OpenAI (opens in a new tab) — OpenAI. Accessed Nov 21, 2026.
  7. 7.Prompt caching (opens in a new tab) — INFRO Docs. Accessed Nov 21, 2026.

About this article

Read on 2026-11-21: INFRO docs pages on privacy, request traces, response cache, prompt caching and EU compliance; OpenAI's enterprise privacy page; Anthropic's privacy policy. The EUR-Lex GDPR page returned no text, so no GDPR article is quoted or cited. Provider API retention periods were not confirmed and are left to your contract.

Last checked against its sources on . The next review is due by . No separate named reviewer is recorded for this article.

How articles are written, sourced and corrected is set out in the editorial standards.