AI avatar sales agent: live SDR or personalised clips?

An AI avatar sales agent comes in two shapes. A live one holds a video conversation with a prospect; Tavus's documentation lists a Sales Development Rep PAL as one of four example use cases, next to an interviewer, customer support and medical intake (read 2026-10-03). A rendered one is a short personalised clip sent to each lead. Sume does the second: one script per lead in, one avatar video of 4-60 seconds out. It does not take live calls.

Pick live when the lead needs to ask questions in the moment, and rendered when the job is to get a clear, reviewed first message in front of many people.

When does a live avatar SDR make sense?

When the next sentence depends on the prospect. Tavus's PAL pipeline supports knowledge-base documents (document_ids), memory tagged per participant and a custom_greeting, according to its create-conversation reference (read 2026-10-03). That fits qualification, pricing questions and objection handling.

The cost model follows the conversation. Tavus lists plans from a free tier of 25 minutes with one concurrent stream, to Starter at $59 a month with 100 minutes and up to 3 concurrent streams, and Growth at $397 with 1,250 minutes and up to 10 (Tavus pricing, read 2026-10-03). The limit that matters for outbound is concurrency: each live prospect occupies a stream while they talk.

When do personalised clips beat a live agent?

For the cold first touch. A clip is watched on the prospect's schedule, needs no concurrency, and can be reviewed before it goes out. With Sume you can render a preview first and approve the first frames before paying for the full render (Avatar video previews).

Personalisation is whatever you put in the script. Sume does not look anyone up; your code fills the name, company or product line into the text and submits. Give each submit its own Idempotency-Key, derived from the lead id, so a retry does not create a second clip.

import os, requests

API = "https://api.sume.com/v1/avatar-1.0/talking-video"
HEADERS = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
leads = [
    {"id": "001", "first": "Dana", "product": "the onboarding kit"},
    {"id": "002", "first": "Sam", "product": "the analytics add-on"},
]
for lead in leads:
    script = (f"Hi {lead['first']}, I recorded this short note about "
              f"{lead['product']}. Reply if you want a walkthrough.")
    r = requests.post(
        API,
        headers={**HEADERS, "Idempotency-Key": f"outreach-{lead['id']}"},
        json={"avatar_handle": "sume_clawra", "script": script,
              "aspect_ratio": "9:16", "quality": "standard",
              "mode": "webhook",
              "webhook_url": "https://example.com/hooks/sume"},
        timeout=30,
    )
    r.raise_for_status()
    print(lead["id"], r.json()["data"]["status_url"])

How do the two compare for outbound?

The deciding fields are who has to be present and what you review.

Should I preview every clip?

Preview the template, not every lead. When the only difference between clips is a name or a product line, create one avatar-video preview with a sample lead, check preview_image_url, and then submit the batch. The preview stills are tier-independent, and generate-video can override quality for the final render only, so approving a draft framing does not lock you into a tier (Avatar video previews).

Spot-check a few rendered outputs from the batch afterwards. Names are the usual failure: a script that reads a handle or a company in an odd way is cheaper to fix in the template than in a hundred sent emails.

What about the follow-up call?

A clip works best as the opener to a human or live-agent conversation, not a replacement for it. The clip states one useful thing, and the next step is a reply, a booking link or a live session. Tavus's conversation API can start a PAL from a link (conversation_url) or send it into a meeting with meeting_url (create-conversation reference, read 2026-10-03), so a rendered clip followed by a live link is a workable sequence.

Keep records of what was said. With a rendered clip, the exact script is the record; store it with the lead id and the Idempotency-Key. That is a small but real advantage when someone later asks what a prospect was told.

What should the clip say?

Keep it under the length of the script window and make the ask one step: reply, book, or watch the product. If the lead replies with a question, that is the moment to hand over to a person or to a live agent. The webhook posts a job.completed event with the artifact URL; the verifier on your side should refuse to run with an empty secret (Webhooks).

• Use a clip for the first touch and a live agent for the reply.
• Derive the Idempotency-Key from the lead id.
• Do not claim the avatar is a person; say it is an AI-generated presenter.
• Read the preview before sending a batch.