Clef-omni for Customer Voice-Note Routing: API Limits, Costs, and Indonesian Evaluation

When is Clef-omni useful for support?

Clef-omni is worth evaluating when a routing decision needs both a customer’s voice and attachments. For example, a team may need to decide whether a voice-note complaint and parcel photo warrant a claims specialist’s review. Begin with internal suggestions; decisions that affect customers require separate checks.

Cloudflare introduced Clef-omni on October 9, 2026 for text, image, audio, and video input. This decision model scores application-defined options rather than composing a free-form reply. Sources for this article were checked on October 11, 2026. All examples below are synthetic and have not been run against the API. They establish no implementation results, Indonesian-language accuracy, or proven savings.

For support categories and queues, see the Decisions API triage guide. This article focuses on the additional work introduced by voice and attachments: file validation, endpoint selection, conflicting evidence, and media-processing costs.

Choose the endpoint before designing the workflow

The same model may accept different modalities through different providers. At this check, OpenRouter supports only text and images for Clef Omni, with audio and video still coming soon. Changing the model name in an OpenRouter integration is not enough to process voice notes.

Use the direct Workers AI route for the design discussed here. Cloudflare’s official example uses @cf/cloudflare/clef-omni and a JSON body containing model: "clef-omni", state, questions, and embedded media in images, audio, or videos. It names WAV/MP3 audio and MP4/WebM video; the example uses base64 data URLs.

Do not assume that files from a customer channel already fit. Check the container, codec, MIME type, duration, and decoding result. If conversion is necessary, account for its cost and quality changes; renaming a file extension is not conversion. Recheck the endpoint documentation before implementation.

Media limits to validate

These hosted Workers AI limits come from the Clef-omni documentation:

Remote URLs are rejected. Context: 64,000 tokens. Media + questions exceeding context: request fails; text state: truncated to fit. Limits and context reference.

Design recommendation: check size and duration before base64 encoding, then also limit the outgoing request body. Validate both combined and individual limits; passing one does not mean passing the other. Unreadable files, incompatible codecs, and missing required attachments should go to manual handling rather than being silently skipped.

Keep the minimum context essential to the decision, including the relationship between each message and attachment. Do not rely on automatic truncation to select the right evidence. Flag incomplete evidence when important context cannot fit.

Synthetic example: a voice note and parcel photo

Imagine a 35-second voice note saying: “The box is dented. I haven’t opened the item yet; please check it first.” The attachment shows the outside of the parcel. This is a training scenario, not a customer recording or a model output.

A reasonable task is to suggest whether a member of staff should review a damage claim. Do not turn “dented box” into a conclusion that the contents are damaged. Agree labels and evidence boundaries with the team:

• Review the claim: there is a concern for a person to inspect, without deciding its cause or compensation eligibility.
• Request context through staff: the order, attachment, or complaint is unclear.
• Review conflicting evidence: speech and the photo disagree, or the attachment’s origin cannot be established.

These labels are proposed application policy, not a vendor-provided schema. Separate observations from operational decisions. A high probability does not establish that a photo is genuine, an order number matches, or a customer qualifies for a refund. Classification must not itself authorize messages, claim-status changes, or money transfers.

An n8n design to test

This is an untested HTTP architecture, not a native Clef-omni integration or an import-ready workflow. The n8n HTTP Request node supports REST calls and JSON bodies; file retrieval and media conversion still need to be designed.

1. Receive the ticket identity and retrieve attachments only from authorized storage or channels. Restrict download sources; do not let URLs inside a message freely determine fetch destinations.
2. Minimize personal data, validate type/size/duration, and construct a compatible media payload. Keep access tokens and base64 files out of logs and workflow exports.
3. Call Workers AI through HTTP Request. Manage credentials separately from the body, set a timeout, and handle failures explicitly.
4. Validate response status, answers keyed by the requested question IDs, allowed choices, and numeric values. Empty responses, unfamiliar shapes, and results that fail the review policy go to staff.
5. If using Switch in Rules mode, set Fallback Output → Extra Output and connect it to human review. Handle request errors before this routing stage.
6. Record suggestions internally using a deduplication key for the ticket and attachment version. Bound retries to prevent duplicate notes or unlimited repeated costs.

During initial evaluation, staff continue handling tickets normally. Retain only necessary evaluation data, restrict media access, and define when copies are deleted. Start with synthetic data. For real records, ensure processing and transmission to the provider are authorized under organizational policy. Voices and photos can contain other people or sensitive information that the task does not need.

Calculate whole-workflow cost

Workers AI pricing lists US$0.15 per million input tokens for Clef-omni. The model documentation says media becomes input tokens and output tokens are not charged. This is not a fixed price per voice note.

For an arithmetic illustration, if a request records 2,000 input tokens, inference costs US$0.0003; 10,000 identical-usage requests cost US$3. The 2,000-token figure is a calculation assumption, not a measurement of the 35-second example. This excludes workflow hosting, storage, conversion, retries, taxes, and reviewer time.

Compare two paths on the same tickets: direct media processing with Clef-omni versus transcription followed by text classification, with equivalent image handling where required. Calculate cost per ticket reaching the correct person, including misrouting corrections. Removing a model stage does not prove lower total cost or faster service.

A text alternative: Microsoft-Decision-1

Microsoft announced Microsoft-Decision-1 on October 9, 2026 in public preview in Foundry. Built on Qwen3.5-9B, it focuses on text decisions with probabilities over predefined choices. Vercel also announced its availability in AI Gateway.

Evaluate this option on the existing-text or already-transcribed path. Do not treat it as a direct substitute for voice and photo input. Compare the same label definitions, transcription quality, image handling, and cost across every stage. A new model does not remove the need to verify evidence.

Evaluate Indonesian before changing queues

Build a labeled test set representative of the team’s work. Separate examples used to tune instructions from the final test set. Have human reviewers assess the complete evidence and record disagreements between reviewers themselves.

• Language: formal and conversational Indonesian, abbreviations, mixed English, and accents relevant to the customer base.
• Audio: quiet speech, street noise, long pauses, multiple speakers, and cut-off recordings.
• Attachments: blurry or unrelated photos, multiple parcel views, and speech that contradicts a photo’s first impression.
• Negative cases: a damaged box with uninspected contents; a normal parcel with a status query; messages that do not support a decision.
• Misleading instructions: speech or text inside an image asking the system to ignore its rules. Treat customer media as evidence, never as authority to change the workflow.
• System failures: oversized files, decoding failures, timeouts, incomplete responses, and repeated delivery of the same ticket.

Measure misrouting by language group and media type, human-review share, missed important cases, time to the correct staff member, and total cost. Agree acceptance criteria with the process owner before seeing results. No universal probability threshold is automatically safe for every accent, product, or category.

An initial evaluation succeeds when it helps staff without hiding uncertainty. Moving difficult cases into an unread review queue does not solve the problem. Assign a review owner and preserve a way back to manual handling.

Prepare a learning session with Fahmi

Discuss AI automation training with Fahmi Fahreza using one support process, safe practice media, and clear success criteria. Agree learning needs and tool access before scheduling practical work.

See corporate AI training for team-session context, or AI agents with n8n for workflow foundations.