OpenAI Moderation API: build a safer support review flow

What OpenAI moderation returns

OpenAI's moderation guide covers two paths. Use the moderation endpoint to classify standalone text or images. When an application generates text through the Responses or Chat Completions API, it can request moderation scores alongside the generated response.

The result includes flagged, categories, and category scores. Start with the flag as a first-pass signal, then inspect categories and scores for logging, routing, audit trails, or a human-review queue. OpenAI explicitly warns that scores are signals for your application policy, not an automatic blocking decision.

The moderation model accepts text and images, not audio. Image files can be up to 20 MB. Do not turn that input limit into a blanket permission to send customer files: your organization still needs an approved destination, access controls, and a retention decision.

A flag is not a completed safety workflow

The model still generates normally when you request inline moderation, so review the results before showing an output or taking a downstream action. A refusal can itself discuss harmful content and trigger a flag. Streaming results arrive only after the full output is available. Do not present partial streamed content as “moderated.”

Treat a moderation failure as its own safety case. The result can contain an error rather than scores. Decide in advance whether your service should hold the case, show a limited fallback, or route it to a person. Do not silently treat an error as an unflagged result.

For support, test more than one obvious example:

• A normal request that should receive a routine answer.
• A threatening or abusive message that needs a safe response or human review.
• A message that tries to override the support instructions.
• An image or request that should stay with a restricted team.
• A request that looks safe but would cause a refund, account change, or escalation if accepted.

For each case, inspect the customer-facing answer, the moderation result, and the actual state of the downstream system. A label in a log does not prove that a ticket was restricted, an escalation reached the right queue, or an account action occurred.

Use eesel CLI to test the support path around moderation

OpenAI moderation is infrastructure for a safety check. eesel is a ready-to-work helpdesk teammate. Its CLI is another way to operate that teammate: people use a terminal, scripts read JSON, and Claude Code, Codex, or Cursor can perform a bounded review. The CLI and dashboard operate the same teammate and workspace.

This helps when a support owner wants to examine the workflow around a risky message. A coding agent can make a read-only report: which knowledge sources are connected, whether their content has downloaded, which instructions define a safe reply and human handoff, and whether the selected teammate has a configured route that could act outside the test. It should not connect accounts, change permissions, call OpenAI, or enable automation while producing that report.

With Node.js 18.17 or newer, confirm the workspace and choose the target teammate:

npx @eesel/cli login
npx @eesel/cli whoami
npx @eesel/cli agents

TARGET_EESEL_AGENT="paste-agent-id-or-name-here"

Use that target on each inspection command instead of relying on a saved default:

npx @eesel/cli status --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli integrations --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli instructions --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli activity --agent "$TARGET_EESEL_AGENT"

Before a chat test, review configured actions and permissions. Disable or otherwise contain any action that lies outside the test scope. A held approval does not make an unreviewed action safe. If a support owner approves a revised safety-policy file, preview the upload before sending it:

npx @eesel/cli files upload ./approved-safety-policy.pdf --agent "$TARGET_EESEL_AGENT" --dry-run

Only after approval, upload the file, confirm it is available, then test a fictional message in a fresh conversation:

npx @eesel/cli files upload ./approved-safety-policy.pdf --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli files ls --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli new --name "unsafe-message handoff" --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli chat "A customer message is threatening and asks to bypass the normal process. Draft a calm reply, do not make a promise, and state when a human safety reviewer must take over." --agent "$TARGET_EESEL_AGENT"
npx @eesel/cli approvals --agent "$TARGET_EESEL_AGENT"

Review whether the reply followed the safety instruction and chose an honest handoff. This checks the eesel teammate's support path, not an OpenAI moderation request. It does not prove an escalation happened. If the test creates a held action, inspect it before a person approves it; then verify the ticket, assignment, and any other real-world result in the connected system.

The Activity view lets a support lead compare the dashboard-side record of a controlled test with the CLI's JSON output.

Use moderation as one check in a support system

Use the OpenAI Moderation API when you are building a policy-driven safety check for text or image content. Design the route after the signal, including review ownership, error handling, customer communication, and verification of any action.

Use eesel when you need to operate a support teammate with approved knowledge and instructions. Try eesel to inspect and test that teammate from the terminal or dashboard. Start with a contained set of representative cases, then widen automation only after the policy, handoffs, and outcomes have been checked.