Your AI agent closes a ticket overnight. Nobody watches it happen. So what goes in the video? The usual advice says skip the screen and go abstract: glowing orbs, networks, a metaphor. But 42 of 47 real AI agent product videos still show a product screen. Not one shows a robot. Your buyer still wants to see where the agent works.
An AI agent video sells software that does work on its own. A no-screen video, the kind the title calls interface-free, shows a product screen for under a tenth of its length. A screen-heavy video shows screens for half its length or more. Mixed videos sit in between. Anything shown in place of the screen is a stand-in: moving text, a diagram, an orb or live-action shots. ADVIDS Research studied 47 agent videos from 37 companies, including Anthropic, SAP and Workday.
Six findings stand out. Most videos show where the agent works (Work Screen). Dashboards show up more than chat windows (Receipt View). Half the screen-heavy videos drop the voiceover, mostly at two companies (Music Bed). Moving text is the top stand-in (Type Layer). Words open the video, and the screen follows (Twelve Seconds). With no screen, videos draw the system (Drawn System).
ADVIDS Research · 47 AI agent videos · 37 companies · coded frame by frame · published 14 September 2026 · 17 min read
Top takeaways
Work Screen
89% of the AI agent demo videos show a product interface screen at some point.
Receipt View
70% of the videos show a dashboard, and 51% show a chat window.
Music Bed
8 of the 16 screen-heavy videos run on music with no spoken voiceover, and 5 are Anthropic's.
Type Layer
57% use moving text on screen, more than any other stand-in.
Twelve Seconds
71% of videos with both open on words or a diagram before the first product screen.
Drawn System
6 of the 7 no-screen videos use 2D animation, and 5 have a narrator.
Questions this report answers
How do AI agent videos show a product with no interface?
They draw it. That's the Drawn System. 7 of the 47 AI agent feature videos have no screen. 6 of those 7 use 2D animation, and 5 have a narrator. They lean on moving text and diagrams of what the agent connects. Companies include Atlassian, Celigo, Genpact, GitLab, etc.
Do agentic AI videos still show the product UI?
Yes. That's the Work Screen. 89% show a product screen at some point (42 of 47). It may be a dashboard, a chat window, a terminal or a screen recording demo. The agent works on its own. But the video still shows where it works, so your buyer can check.
Should an AI agent video use a chat interface or a dashboard?
Lead with the dashboard, unless the agent is the conversation. That's the Receipt View. 70% of the videos show a dashboard of finished work, and 51% show a chat window. A dashboard shows what the agent did. A chat shows what it was asked. Your buyer checks the first.
What visuals replace the interface in autonomous AI videos?
Moving text, most of all. That's the Type Layer. 57% of the videos use animated text on screen, typically by 0:04. Diagrams come next. Glowing orbs and live-action shots are rare, and no video shows a robot. So write your on-screen lines with the script.
Do AI agent product videos need a voiceover?
Most do. That's the Music Bed. 60% of the videos have a voiceover, and the rest run on background music only. Screen-heavy videos split: 8 of the 16 drop the voice. Most of those come from Anthropic and OpenAI. If strangers can follow your screen on mute, music can carry it.
When should the product screen appear in an AI agent video?
After a few seconds of words. That's Twelve Seconds. In 71% of videos with both, moving text or a diagram comes before the first product screen (24 of 34). The typical first screen lands at 0:12. So open on a line or a diagram, and have the product ready by then.
How was this analysis done?
ADVIDS Research screened 188 product videos against a rule set before the work began. 47 qualified, from 37 companies. Each was coded frame by frame. The six findings, Work Screen, Receipt View, Music Bed, Type Layer, Twelve Seconds and Drawn System, are simple counts. The full dataset comes with the report.
How to read this report
This report looks at 47 AI agent videos from 37 companies, posted to YouTube from January to August 2026. Each figure counts videos, like the 42 of 47 that show product screens in use. It's a small set with no view data, so read it as a guide and test what persuades your buyer.
Finding 01 · Work Screen
1. Most videos still show where the agent works.
You might think an agent with no one at the controls needs no screen in its video. Not so. 89% of AI agent videos still show a product screen (42 of 47). Very few skip it. If your agent works where a buyer could later look, show that screen, even when nobody clicks.
Autopilot flies the plane. The pilot still watches the instruments. Your product screen is that instrument panel.
And the screen isn't a quick cameo. In the typical video, a product interface screen stays up for 43% of the time.
Example: IPRally's agent video runs on music alone. At 0:14 the agent's Steps panel appears, working through a patent claim one stage at a time. Nobody types, but you watch it work.
Videos that show a screen: companies in this group include Anthropic, SAP, OpenAI, Workday, etc.
Exception: an agent that works inside tools your buyer already owns. Swimlane's Symphony of AI Agents video shows no product screen at all. Moving text and live-action shots of a security analyst show the analyst's role instead.
Source: Advids analysis of 47 AI agent videos
A no-screen video is the exception, not the rule.
Finding 02 · Receipt View
2. Dashboards show up more often than chat windows.
When an agent finishes work, most videos show the record of it. 70% show a dashboard (33 of 47), and 51% show a chat window. A dashboard shows what the agent did. A chat shows what it was asked. If someone has to check your agent's work, lead with the dashboard.
You don't check a waiter by rereading your order. You check the bill. Your dashboard is the bill.
Chat often shows up first. But the dashboard stays on screen longer. A few screen-heavy videos lead with a terminal or a SaaS screencast instead.
Example: Savant Labs' finance video opens on title cards. At 0:32 the dashboard lands and stays for more than half the video. A flow canvas lays out each step the agent took, ready to check.
Videos with a dashboard: companies in this group include Alloy, Amplience, Asana, Savant Labs, etc.
Exception: an agent that is the conversation. Zoom's Virtual Agent tour shows the chat window more than the dashboard. When the agent is the support chat, the chat window is the product.
Source: Advids analysis of 47 AI agent videos
Chat isn't the default picture of an agent. The record of its work is.
Finding 03 · Music Bed
3. Half the screen-heavy videos drop the voiceover.
Screen-heavy videos split on sound. 8 of the 16 run on music alone, with no voiceover. But that's mostly two companies. Every screen-heavy video from Anthropic and OpenAI is music only. Outside that pair, just 1 of 9 drops the voice. So if your screen labels each step in plain words, music can carry it.
Watch a cooking video on mute. If every jar has a label, you can follow along. Your screen labels are the jars.
The typical screen-heavy interface demo video speaks just 8 words. No-screen and mixed videos mostly keep a narrator.
Example: Anthropic's Claude Code agent video runs 0:39 with zero words. The terminal is on screen from 0:00. Sessions sit in groups marked Needs input and Working, so the labels do the talking.
Music only, screen-heavy: companies in this group include Anthropic, OpenAI and Make.
Exception: a viewer who has to follow a build. Blink's Agent Builder video keeps a short voiceover over its screen recording. As a prompt turns into an app, the voice names each step.
Source: Advids analysis of 47 AI agent videos
The silence comes from two companies. Other companies mostly keep a narrator.
Finding 04 · Type Layer
4. Moving text is the top stand-in for the screen.
When an agent's work happens out of sight, most videos write it out as moving text. 57% use moving text on screen (27 of 47). It's also the first stand-in in 61% of the videos that use one. If your agent works out of sight, write those lines with the script.
At the airport, you never see your plane leave the gate. You read one word on the board: Boarding. Your moving text is that board.
Diagrams come second, in 43% of videos. Before any diagram, orb or robot, videos put the agent's work into words on screen. Most no-screen videos use animated typography too.
Example: Dext's AI Assist video opens on moving text at 0:00: “It learns how”, above a large orange “you”. The dashboard follows at 0:13. One line says what the agent does, then the screen proves it.
Moving text on screen: companies in this group include Atlassian, Board, Celigo, Contentstack, etc.
Exception: one picture that carries the claim. Genpact's accounts-payable spot uses no text and no product screen. Live-action shots of an orchestra whose players have invoice-paper heads carry it on music alone.
Source: Advids analysis of 47 AI agent videos
Moving text shows up in 27 videos. Flowcharts and network diagrams show up in 12 each.
Finding 05 · Twelve Seconds
5. Words open the video, and the screen follows.
Most videos with both a screen and a stand-in open on the stand-in. 71% show moving text or a diagram before the first product screen (24 of 34). The typical first screen lands at 0:12. If your video has both, open on words and show the screen by then.
Your GPS names the turn before you reach it. The words get your eyes ready. The screen is the turn.
Many openings go to a title or logo card: 47% start on one. Just 5 videos start on a product screen.
Example: Board's planning video opens on a title card. At 0:07 a glowing card reads Controller Agent. The first dashboard lands at 0:15, and a network diagram follows.
Words first: companies in this group include Dext, ElevenLabs, Genpact, GitLab, etc.
Exception: an agent whose output is a queue to approve. Campfire's Ember Agents video shows its accounting dashboard at 0:04, right after the title card. When the buyer's job is approving items, the queue opens the video.
Source: Advids analysis of 47 AI agent videos
The typical first product screen lands at 0:12.
Finding 06 · Drawn System
6. With no screen, videos draw the system the agent runs.
Some agents have no screen to show. 7 of the 47 videos are no-screen videos. Of those 7, 6 are built with 2D animation, and 5 have a narrator. If you have no screen, draw what your agent connects. Add a line of text for each step, and a narrator.
A subway map isn't the tunnels. But it's the only picture of them anyone can read. Your diagram is the map.
5 of the seven are motion graphics videos, and most open on a title card.
Example: Celigo's Agent Builder video shows a flowchart from 0:04. Apps like CRM and ERP sit on one side, linked through a Celigo Agent icon to people. Moving text takes over at 0:33.
No-screen videos: companies in this group include Celigo, GitLab, IBM, Swimlane, etc.
Exception: a product that is the coordination itself. Salesforce's Agentic Orchestration explainer uses no screen and no diagram. A photo collage of jazz players stands in, and the narrator casts the orchestrator as the bandleader.
Source: Advids analysis of 47 AI agent videos
5 of the 7 also have a narrator.
Benchmarks by video type
Check your own video against each column. The no-screen group holds just 7 videos.
Screens, stand-ins, words and length
| Measure | No screen (7) | Mixed (24) | Screen-heavy (16) | All (47) | What it means |
|---|---|---|---|---|---|
| Typical runtime | 1:14 | 1:17 | 1:03 | 1:14 | Half the videos run shorter than this |
| Typical screen time, percent of the video | 0 | 37 | 68 | 43 | How much of the video shows a product screen |
| Typical stand-in time, percent of the video | 71 | 35 | 8 | 25 | Moving text, diagrams, orbs, live-action shots or alerts |
| Typical first product screen | 0:44 | 0:14 | 0:04 | 0:12 | Time into the video, for videos that show a screen |
| Typical first stand-in | 0:04 | 0:06 | 0:08 | 0:05 | Time into the video, for videos that use one |
| Typical words spoken | 148 | 140 | 8 | 115 | Music-only videos count as zero |
What almost nobody does
| Pattern | Videos | Companies |
|---|---|---|
| Shows a robot | 0 of 47 | none |
| Shows a phone or tablet screen | 3 of 47 | Contentstack, Ofelia and Workday |
| Shows a glowing AI orb or shape | 4 of 47 | ElevenLabs, IBM, Nexthink and Nooks |
| Uses live-action shots | 4 of 47 | Genpact, SAP, StackAI and Swimlane |
| Puts a presenter on camera | 4 of 47 | Nooks, Ofelia, SAP and StackAI |
No video shows a robot, and just 4 show a glowing AI orb.
Six questions for your next brief
Answer these six on one page before anyone opens a storyboard.
- Finding 01 Can a buyer see where your agent works? Yes: budget for a clean pre-sales demo environment. No: go to Finding 06.
- Finding 02 Does your agent leave a record, or is it the conversation? A record: design that screen first. The conversation: lead with chat.
- Finding 03 Can strangers follow your screen on mute? Yes: music can carry it. No: keep the voiceover.
- Finding 04 What stands in for the work nobody sees? Write the on-screen lines with the script, not after the edit.
- Finding 05 What fills the seconds before the screen? Storyboard those seconds as text or a diagram, and land the screen by 0:12.
- Finding 06 No screen at all? Brief a 2D animation studio and a narrator, and draw what the agent connects.
Now go back to the agent that closes a ticket overnight, with nobody watching. 42 of 47 videos still showed where the agent worked. The other findings decide what surrounds that screen, or what replaces it.
This week, list the screens where a buyer could check your agent's work. Answer the six questions on one page. Then write the brief.
Zero-UI is how the agent works. A screen is how the buyer checks it.
Methodology
| Vertical | Videos |
|---|---|
| devtools | 9 |
| fintech | 7 |
| cx | 4 |
| productivity | 3 |
| ai | 3 |
| enterprise | 2 |
| other verticals | 19 |
| Coding scheme | Values | Definition |
|---|---|---|
| Strategy class | interface-free · blended · interface fallback (screen-led) | Computed from coded segments: interface-free when product-screen motifs cover under 0.10 of runtime and the coded surface lists no dashboard, chat or terminal; interface fallback when they cover 0.50 or more; blended otherwise. Fallback videos are sub-typed chat-led, dashboard-led or other screen-led by the screen motif with the most runtime. |
| Screen motif | dashboard_ui · chat_ui · terminal_ui · handheld_device_ui · screen_recording_generic | Coded per timed segment from frames, with the first timecode and the share of runtime each motif covers. |
| Interface-free motif | kinetic_type · diagram_flowchart · diagram_network · ai_orb_abstract · cinematic_broll · robot_hardware · notification_alert | Coded per timed segment from frames. cinematic_broll is live-action b-roll; notification_alert is a toast or alert card. |
| Surface | dashboard · chat · terminal · diagram · none | Coded from frames: the interface a video shows. A video can list several. |
| Cards | title_card · brand_logo_card | Coded per timed segment. The first and last segments give the opening and closing frame. |
| Build and format | 2d_animation · ui_mockup · screen_recording · kinetic_type · live_action · 3d_cgi · screencast · motion_graphics | Coded from frames with a confidence of high or medium. A video can carry several. |
| Sound and frame | voiceover · music_only · word count · brightness 0 to 255 | Voiceover or music only from audio analysis; words from captions; brightness computed over the nine sampled frames, not the interface alone. |
| Not coded | cast · altitude | Who appears and what the video argues. The library records neither, and this report does not code them from transcripts. |
Cite this report
Jai Ghosh, Advids. Decoding the Zero-UI Challenge: 47 Autonomous AI Videos Audited for Interface-Free Narrative. 14 September 2026. https://advids.co/blog/zero-ui-autonomous-ai-video-narrative
Author & editor bio
About Advids
Advids produces video for AI, SaaS and DeepTech companies. ADVIDS Research publishes coded studies of what those companies ship, with the dataset attached so every count can be checked.