Which provider type fits which job?
| You need… | Provider type | Example | What it doesn't cover |
|---|---|---|---|
| Creative, media and activation measurement together | Growth studio or paid social agency with product analytics | Kiko (publisher of this page). Case study: for a seed-stage AI agent startup, Kiko brought cost per activated workspace down 37% with demo-led YouTube and LinkedIn ads | : |
| Short-form ad creative in volume, you run media | Creative studio | Hooklab: short-form video ads for AI tools; creative only, no media buying; published packages from $1,500 a month for AI variants | Media buying, measurement |
| YouTube campaign management | YouTube ads agency | Theory Road: runs YouTube campaigns in its own manager account, five-second-first creative, lift measurement; pricing on request | Meta; product activation |
| One-off brand video | Video production company | Many | Testing and iteration |
Kiko has run programs for 9 AI agent products since 2023.
Questions to ask any provider: Will you test openings separately from the body? Which event will you judge on: signup, setup complete or first completed task? Who records unedited demo footage, and how fast?
What does a first completed task cost? (worked example)
An agent's value starts after setup: connecting accounts and granting permissions. The opening seconds decide who signs up, and who signs up decides who finishes setup. The worked example below shows how the funnel changes acquisition cost.
Cost per first completed task = cost per signup ÷ (setup-complete rate × first-task rate)| Opening (first 5 seconds) | Cost per signup | Setup complete | First task | Cost per first completed task |
|---|---|---|---|---|
| Baseline | $12.00 | 40% | 60% | $50.00 |
| A. Finished result first ("here's my week, planned") | $15.00 | 50% | 70% | $42.86 |
| B. Broad capability claim ("your AI for everything") | $10.00 | 30% | 50% | $66.67 |
The cheapest signups came from the vaguest promise and cost the most per real user. Kiko's pooled figures for AI agent products, across 9 programs run in 2025 and 2026: cost per first completed task between $35 and $80, and setup-complete rate between 35% and 60%.
Decision table: which event should the campaign optimize to?
| If weekly volume of this event is… | Optimize to | Report alongside |
|---|---|---|
| First completed task ≥ 50 a week per campaign | First completed task (the agent finished a task the user kept), server-side | Second task within 7 days |
| Setup complete ≥ 50 a week per campaign | Setup complete (accounts connected, permissions granted) | First-task rate |
| Neither | Signup, then move deeper within 4 weeks | Setup-complete rate |
Defining "completed" is harder than it looks; see AI agent activation and evals.
Decision table: which channel, by agent type.
| Agent type | Meta | YouTube |
|---|---|---|
| Consumer or prosumer task (job applications, personal admin) | Primary. Short task-to-result clips; creator partnership ads | Secondary; tutorials and reviews |
| Work agent bought by individuals | Good for trials with one clear task | Strong. Longer demos next to how-to content |
| Agent builder or platform | Hard without a concrete example | Primary. Build one agent end to end |
| Agent sold to SMB owners | See SMB owner acquisition | Customer stories |
Which demo structures should you test, and what proof survives skepticism?
| Structure | What the viewer sees | Proof that survives skepticism |
|---|---|---|
| Task to result | One instruction, then the finished output | Unedited screen recording with a visible clock, or a clear "sped up 4x" label |
| Workload before/after | A pile of work, then the cleared state | Real inputs, not a mock-up |
| Watch it recover | The agent hits a problem, asks or retries, then finishes | An honest recovery beats a perfect run |
| Named customer | A real team describing a workflow change | Consent and results they actually got |
| Creator day-in-the-life | A known person using the agent through a real day | Disclosed partnership; tasks the creator really does |
Claims in demos are still claims. TikTok's ad policy prohibits promised or exaggerated results (TikTok, updated April 2026), and Meta prohibits unrealistic outcomes (Meta). If a demo is sped up, cherry-picked or uses a paid tier, say so on screen.
Test plan for the first 30 seconds. Keep seconds 5 to 30 identical so only the opening changes.
| Cell | Opening | Budget share | Review window | Pass rule |
|---|---|---|---|---|
| A | Finished result first | 25% | 7 days or $1,000 spent, whichever comes first | Cost per first completed task |
| B | Task stated as a pain ("200 unread emails") | 25% | 7 days or $1,000 spent, whichever comes first | Same |
| C | Creator or customer face and claim | 25% | 7 days or $1,000 spent, whichever comes first | Same |
| D | Failure and recovery | 25% | 7 days or $1,000 spent, whichever comes first | First-task rate of new users |
Kill any cell whose cost per first completed task is more than 30% above the best cell at review; move its budget to the winner.
Which creators would we brief for agent demos?
The creator day-in-the-life and build-tutorial structures need people who already show agents doing real work on screen. These four YouTube creators do. Counts are vidIQ estimates.
| Creator | Platform | Audience (source) | Why they fit |
|---|---|---|---|
| Nate Herk, @nateherk | YouTube | 1.11M subscribers (vidIQ) | Builds agent and automation systems for small-business operators and freelancers, end to end. He already runs disclosed AI-tool partnerships. |
| Nick Saraev, @nicksaraev | YouTube | 599K subscribers (vidIQ) | Long use-case tutorials that test agents on business tasks; footage that cuts into task-to-result ads. |
| Nate B Jones, @NateBJones | YouTube | 326K subscribers (vidIQ) | Analysis for managers deciding whether agent work is real. Fits work agents that need a considered, skeptical review. |
| AI Jason, @AIJasonZ | YouTube | 234K subscribers (vidIQ) | Builds agent workflows end to end for builders automating their own work. Fits agent builders and platforms. |
More names are on the AI agent creator list.
What are AI agent products running right now? (ad-library census)
Nine active US Meta ads from nine advertisers, found by searching the Meta Ad Library for "AI agent" and "AI agents". Meta shows ad content, not targeting or spend.
Snapshot: active US Meta ads, November 2026
| Advertiser (as shown) | Format | Structure (our reading) | Task or hook (paraphrased) | Started running | Ad |
|---|---|---|---|---|---|
| AI Apply | Video, 31 s | Workload before/after | Weeks of job applications done in minutes | 2026-05-26 | 1322117609986431 |
| Muse | Static | Task to result | A free AI agent that crunches numbers while you sleep | 2026-10-22 | 1102623035897231 |
| Meta | Video, 30 s | Task to result, SMB | "Muse for Small Business" does the work rather than answering questions | 2026-10-22 | 1138650861851555 |
| Nastia Liukin with Muse | Video, 75 s (partnership ad) | Creator day in the life | A day with many tabs open, handled by the agent | 2026-10-16 | 1119786967171350 |
| Ally Love - Peloton with Muse | Video, 104 s (partnership ad) | Creator day in the life | Her personal AI assistant | 2026-10-29 | 1017787748018675 |
| Intuit QuickBooks | Static | Agents plus humans | AI agents and human experts manage your books | 2026-09-09 | 1101846609192763 |
| ElevenLabs | Video, 60 s | Build an agent | Describe the agent in plain English and it gets built | 2026-09-24 | 1394618548967037 |
| OpenAI | Video, 21 s | Named customer | How Shopify teams use ChatGPT | 2026-10-09 | 1111044178101325 |
| Notion | Static | Platform feature | Create custom AI agents at work | 2026-08-05 | 1598592228504757 |
The Meta library showed about 5,400 active US ads for "AI agent" and about 8,000 for "AI agents" (exact-phrase keyword matches, not advertisers).
Census table (9 Meta ads):
| Demo structure | Advertiser pages | Ads | Median video length | Static ads |
|---|---|---|---|---|
| Task to result | 2 | 2 | 30 s | 1 |
| Workload before/after | 1 | 1 | 31 s | 0 |
| Named customer | 1 | 1 | 21 s | 0 |
| Creator day-in-the-life | 2 | 2 | 90 s | 0 |
| Build an agent or platform feature | 2 | 2 | 60 s | 1 |
| Agents plus humans | 1 | 1 | : | 1 |
What the snapshot shows:
- Creator partnership ads at scale. One agent ran several partnership ads with known fitness, sports and parenting creators that started between 2026-10-13 and 2026-10-29, each framed as a day in the creator's life. They were the longest videos in the snapshot, at 75 and 104 seconds.
- Hybrid proof for high-stakes tasks. A bookkeeping agent paired "AI agents" with human experts.
- Named customers for work agents. One ad was built around a named company's team.
- "Build your own agent" needs an example. Platform ads that only said "create custom agents" gave no task to picture.
When don't you need outside help?
If your team already records unedited demos weekly and tracks first completed task as an event, you may only need media buying help or more creative volume.
Frequently asked questions
Who can help an AI product test YouTube product-demo ads?
A paid social and YouTube agency that tests demo creative and measures it to the first completed task. Options include Kiko, Hooklab for AI-tool ad creative, and Theory Road for YouTube campaign management. Production-only shops usually stop at delivery.
Do Facebook ads work for AI agent startups?
Yes, for agents with a task viewers recognize. In our November 2026 snapshot, agent products ran task-to-result clips and creator partnership ads on Meta. Judge them on cost per first completed task, not cost per signup.
Are YouTube ads good for AI agents?
YouTube suits agents that need a longer demonstration, such as work agents and agent builders. Test the opening seconds separately and read retention alongside first-task conversions.
What makes an AI agent demo ad believable?
Real inputs, an unedited or clearly labeled run, and a result the viewer can check. Showing the agent recover from a problem builds more trust than a perfect run.
What conversion event should an AI agent optimize for?
The first completed task the user keeps, once it reaches 50 a week per campaign. Before that, optimize to setup complete, because connecting accounts is where many trials stop.
Sources
- Meta Ad Library (individual ads linked in the snapshot table)
- Google Ads Transparency Center
- TikTok, Misleading and False Content ad policy
- Meta, Advertising Standards: Unrealistic outcomes
- Hooklab
- Theory Road, YouTube ads
- vidIQ channel pages for @nateherk, @nicksaraev, @NateBJones and @AIJasonZ