# AI / LLM Data Handling (/docs/trust/ai-llm)



Layers uses LLMs and image/video models for:

* Content generation (copy, image, video, scripting).
* The ads-optimization agent.
* App Machina's SDK-instrumentation coding agent.
* UGC researcher + creative agents.

## Providers [#providers]

| Provider                  | Used for                                               |
| ------------------------- | ------------------------------------------------------ |
| Google Vertex AI / Gemini | Primary LLM for content generation, ads agent, scoring |
| Anthropic (Claude)        | Selected agent tasks                                   |
| OpenAI                    | Selected agent tasks                                   |
| Replicate                 | Image / video model hosting                            |

Per-request only — we don't enroll in any training-data sharing
program, and agent calls are made with the providers' zero-retention
inference APIs where those are offered.

## What gets sent [#what-gets-sent]

| Context            | What's in the prompt                                             |
| ------------------ | ---------------------------------------------------------------- |
| Content generation | Brand brief, voice, banned words, product info, reference links. |
| Ads agent          | Campaign config, current metrics, pool of eligible creatives.    |
| App Machina        | Repo files relevant to the task, user prompt, compile errors.    |
| UGC researcher     | Creator allowlist, post content from SIFT, niche filters.        |

## What does NOT get sent [#what-does-not-get-sent]

* End-user PII from SDK events.
* Vault-stored tokens / credentials.
* Anything not relevant to the specific task.

Each agent's system prompt scopes what it reads.

## Transparency [#transparency]

AI-generated creative includes metadata tagging it as AI-generated.
User-facing output is labeled where legally required (e.g., California
SB 942, EU AI Act transparency obligations).

## Human oversight [#human-oversight]

Every automated decision has:

* Guardrails (min-age, min-spend, never-empty-adset — enforced by the
  optimizer).
* Manual override (you can always intervene).
* An audit trail in the partner audit log where the action is
  partner-visible.

See [Safety & guardrails](/docs/concepts/safety-guardrails).