48 AI agent demo videos from 40 companies, split Q1 2026 (24) against April to August 2026 (24), both periods under the 30-video floor, show three movements of their own: a dashboard is the interface surface in 9 later videos against 15, network diagrams rise from 4 to 8 and title cards fall from 20 to 13, while the rest of the enterprise AI launch set moved the other way on all three. Anthropic, Salesforce and Microsoft illustrate them. Pitching and the modern enterprise are the reader's test: the files hold no buyer response.
ADVIDS Research · 48 agentic demo videos coded · 40 companies · 432 sampled frames · 490 timed segments · published 14 September 2026 · 13 min read
Top takeaways
AI agent demo videos moved off the dashboard
A dashboard is the coded interface surface in 9 of 24 AI agent demo videos from April to August 2026, against 15 of 24 in Q1 2026, while chat panels hold at 15 and 13; later dashboard-surface videos include Anthropic, SAP, Foyr, etc.
Network diagrams and multi-agent language rose in agent demo videos
Network diagrams rise from 4 of 24 AI agent demo videos in Q1 2026 to 8 of 24 from April to August 2026, and multi-agent language from 4 to 8; companies drawing the diagram include Atlassian, Board, Salesforce, etc.
Title cards fell out of AI agent demo videos
Title cards appear in 13 of 24 AI agent demo videos from April to August 2026, down from 20 of 24 in Q1 2026, while kinetic-type segments hold at 14; later videos with no title card include ElevenLabs, Genpact, Microsoft, etc.
Questions this report answers
What are the trends in AI agent demo videos in 2026?
Across 48 AI agent demo videos from 40 companies uploaded between 12 January 2026 and 3 August 2026, 3 movements separate Q1 2026 from April to August 2026: dashboards left the screen, network diagrams and multi-agent language rose, and title cards fell. Each period holds 24 videos, so every count is directional.
Do AI agent demo videos still show a dashboard?
Fewer do: dashboards are the interface surface in 9 of 24 videos from April to August 2026, down from 15 of 24 in Q1 2026, and hold a quarter of the runtime in 6, down from 13. LangChain's LangSmith Engine video is one of those 6, with dashboards on screen for 29 per cent of the runtime.
Are AI agent videos pitching multi-agent systems?
More of them: 8 of 24 videos from April to August 2026 draw a network diagram, up from 4 of 24 in Q1 2026, and multi-agent or orchestration language moves from 4 to 8. Both rest on 8 later videos, exactly the materiality floor.
Do agent demo videos pitch autonomy or human approval?
Neither moved as a category: human-in-the-loop language moves from 5 to 8 of 24, a move of 7.7 points without SAP, and autonomy language from 5 to 8, 6.4 points without Genpact. On screen alone, approval language moves from 1 to 7.
How does this compare with the enterprise AI launch report?
46 of these 48 videos are also in that report, and its UI mockup, light-frame and voiceover findings hold here while its live-action and closing-logo shifts do not. Outside this set, as that group shrank from 28 to 21 videos, dashboard surfaces went from 11 to 13, network diagrams from 8 to 4 and title cards from 15 to 13.
How was this analysis done?
ADVIDS Research screened 188 visually coded videos against a rule written before intake; 93 matched an agent term and 48 qualified from 40 companies, dated by YouTube upload. Counts compare 24 videos from Q1 2026 with 24 from April to August 2026 across 432 sampled frames and 490 timed segments.
How to read this report
48 AI agent demo videos from 40 companies qualified from 188 screened against a rule written before intake. Every one is a product demo video, launch video or explainer whose subject is an AI agent sold to business buyers, uploaded to YouTube from 12 January 2026 to 3 August 2026.
A figure such as 8 of 24 counts videos, never frames. The window splits at 1 April 2026 into 24 videos in Q1 2026 and 24 from April to August 2026; both bases sit under the 30-video floor, so every count is directional.
A shift counts as a trend only when it moves at least 10 points, involves at least 8 videos and still clears 10 points without its largest April-to-August company and without Anthropic.
Trend 01 · Interface · Dashboard · Q1 2026 against April to August 2026
1. AI agent demo videos moved off the dashboard
A dashboard is the coded interface surface in 9 of 24 AI agent demo videos uploaded from April to August 2026, against 15 of 24 in Q1 2026. Dashboard segments appear in 15 of 24, down from 20, and hold a quarter of the runtime or more in 6, down from 13. Chat panels hold at 15 and 13. Of the 9 later videos with no dashboard segment, 2 show a terminal, 2 a chat panel and 2 a network diagram. In the enterprise AI launch videos outside this set, a dashboard surface went from 11 to 13 while that group shrank from 28 to 21 videos, so the fall belongs to the agent videos. The dashboard stopped being the default stage for an agent.
Companies with a dashboard as the surface include, in Q1 2026, Alloy, Amplience, Asana, etc.; and from April to August 2026, Anthropic, SAP, Foyr, etc. Without Anthropic and SAP in either period, the fall is still 35.8 points; without Anthropic alone, 30.2.
LangChain goes the other way. Its May 2026 LangSmith Engine video runs 1:12, with product panels coded as dashboards for 29 per cent of it: an error search, an evaluator set-up screen and a dataset view. Where the agent's job is improving other agents, the evaluation panels are the record of its work.
Source: ADVIDS Research coded set, 48 videos from 40 companies, uploaded 12 January 2026 to 3 August 2026; each period under the 30-video floor
- Q1 2026, n=24
- Apr to Aug 2026, n=24
Median dashboard time across all videos falls from 26 to 7 per cent of runtime.
Trend 02 · Interface · Diagrams and multi-agent language · Q1 2026 against April to August 2026
2. Network diagrams and multi-agent language rose in agent demo videos
Network diagrams appear in 8 of 24 AI agent demo videos uploaded from April to August 2026, against 4 of 24 in Q1 2026, and language about multiple agents or orchestration, spoken or on screen, moves from 4 to 8. Flowcharts fall from 7 to 4, so the diagram changed shape as well as count. In the enterprise AI launch videos outside this set, network diagrams went from 8 to 4. Both rises rest on 8 later videos, exactly the materiality floor, and the multi-agent row clears the concentration test by less than a point without Genpact (10.6 points), so this is the thinnest of the three chapters. A third of the later videos draw the agents as a system.
Companies drawing a network diagram include, in Q1 2026, GitLab, Perplexity, Zoom, etc.; and from April to August 2026, Atlassian, Board, Salesforce, etc. Without SAP in either period, the network-diagram rise is 11.9 points; without Anthropic, 22.6.
AppsFlyer goes the other way. Its April 2026 AI Assistant video runs 1:18 with one assistant in a chat panel beside the analytics dashboard, answering campaign questions, and no diagram of agents. Where the product is a single assistant next to the data it answers from, one panel carries the whole pitch.
Source: ADVIDS Research coded set, 48 videos from 40 companies, uploaded 12 January 2026 to 3 August 2026; each period under the 30-video floor
- Q1 2026, n=24
- Apr to Aug 2026, n=24
3 later videos carry both, so the drawn system and the named one mostly arrive in different videos.
Trend 03 · Structure · Title cards · Q1 2026 against April to August 2026
3. Title cards fell out of AI agent demo videos
Title cards appear in 13 of 24 AI agent demo videos uploaded from April to August 2026, against 20 of 24 in Q1 2026. Videos opening on a title or logo card fall from 13 to 10, and median time on title and logo cards falls from 17 to 13 per cent of runtime, while kinetic-type segments hold at 14 and 14. Kinetic type listed as a format falls from 9 to 3, the same fall the enterprise AI launch and UI animation reports measured. In the enterprise AI launch videos outside this set, title cards went from 15 to 13 while that group shrank from 28 to 21 videos, so the title-card fall belongs to the agent videos. The words stay on screen; the card that frames them goes.
Companies using a title card include, in Q1 2026, Anthropic, Campfire, Celigo, etc.; and from April to August 2026, Anthropic, Salesforce, AppsFlyer, etc. Later videos with none include ElevenLabs, Genpact, Microsoft, etc.
Anthropic goes the other way. All 4 of its later videos carry a title card; its April 2026 Claude Code desktop video runs 0:37, opening on a 'Code' card and closing on a 'New desktop app now available' card around a working session. Where one company ships several short launches in a quarter, the card is the series frame.
Source: ADVIDS Research coded set, 48 videos from 40 companies, uploaded 12 January 2026 to 3 August 2026; each period under the 30-video floor
- Q1 2026, n=24
- Apr to Aug 2026, n=24
Anthropic keeps a title card in all 4 of its later videos; without Anthropic the fall is 37.6 points, not 29.2.
Benchmarks
Descriptive figures for placing your own AI agent demo video against the set.
Length, interface and sound
| Measure | Q1 2026 (24) | Apr–Aug 2026 (24) | What it means |
|---|---|---|---|
| Runtime, median | 1:14 | 1:07 | Half the videos in the period run shorter than this |
| Runtime, middle half | 1:00 to 1:40 | 0:56 to 1:42 | The runtime span of the middle 50 per cent of videos |
| First interface on screen, median | 0:12 (n=23) | 0:14 (n=21) | When an interface motif first appears, among videos that show one |
| First chat panel, median | 0:11 (n=15) | 0:21 (n=13) | When a chat panel first appears, among videos that show one |
| Interface coverage, median | 47 | 43 | Per cent of runtime with an interface motif on screen |
| Chat panel coverage, median | 11 | 10 | Per cent of runtime with a chat panel on screen, all videos |
| Words spoken, median | 122 | 88 | Words in the captions; videos with no captions count as zero |
| Videos with a spoken track | 15 | 13 | Videos whose captions carry 15 or more words |
| Frame brightness, median | 137 | 161 | Mean luminance across the nine sampled frames, 0 to 255 |
What did not move
| Coded value | Q1 2026 (24) | Apr–Aug 2026 (24) | Points | Reading |
|---|---|---|---|---|
| Agent, agents or agentic in the video's own words | 20 | 21 | +4.2 | Held |
| Human-in-the-loop language, spoken or on screen | 5 | 8 | +12.5 | SAP's movement: 7.7 points without SAP |
| Human-in-the-loop language on screen | 1 | 7 | +25.0 | Emergence watch: one video under the count floor |
| Autonomy language, spoken or on screen | 5 | 8 | +12.5 | Genpact and SAP: 6.4 points without Genpact |
| Names an action the agent completes | 9 | 6 | -12.5 | Base changed: 4.8 points without Anthropic |
Named actions fall because Anthropic's 4 later videos carry no spoken track, so the base changed rather than the pitch.
What to do with this
- Trend 01 If 6 of 24 AI agent demo videos from April to August 2026 hold a dashboard for a quarter of the runtime, ask whether your demo video needs one before building the demo tenant.
- Trend 02 If 8 of 24 draw a network of agents, decide in the creative brief whether you are selling one assistant or a system, and write the diagram into the script if it is a system.
- Trend 03 If title cards fell to 13 of 24, write the on-screen lines over the product rather than on cards between sections.
Methodology
| Vertical | Videos |
|---|---|
| devtools | 7 |
| fintech | 6 |
| productivity | 4 |
| cx | 4 |
| martech | 3 |
| itops | 2 |
| ai | 2 |
| other verticals | 20 |
| Coding scheme | Values | Definition |
|---|---|---|
| Interface motif | dashboard_ui · chat_ui · terminal_ui · notification_alert · screen_recording_generic | Coded per timed segment, with the first timecode and the share of runtime each motif covers. |
| Interface surface | dashboard · chat · terminal · diagram · none | Coded from frames: the interface the video is organised around. A video can carry several. |
| Diagram | diagram_network · diagram_flowchart | Coded per timed segment. A network diagram draws nodes and links; a flowchart draws a sequence. |
| Structure | title_card · brand_logo_card · kinetic_type | Coded per timed segment. The first segment gives the opening frame. |
| Build | ui_mockup · screen_recording · 2d_animation · 3d_cgi · kinetic_type · live_action · stock | Coded from frames, with a confidence of high or medium. ui_mockup is an animated recreation of an interface; screen_recording is capture of the product. |
| Format | screencast · motion_graphics · kinetic_type · talking_head | Coded from frames: the genre the video reads as, plus a live-action genre value. A video can carry several. |
| Pitch flags | autonomy · human in the loop · actions · named third-party app · integrations · worker framing · quantified outcome · multi-agent · trust | Phrase lists fixed before any change was computed, matched over the transcript (videos with 15 or more words) and, separately, over on-screen text. Apps owned by the video's own company do not count as third-party. |
| Vocabulary | agent · copilot · assistant · bot | Matched over the video's own words: title, transcript and on-screen text. |
| Frame and sound | light · dark · mixed · brightness · voiceover · music_only | Computed: luminance over nine sampled frames, and audio analysis. Word counts come from auto or manual captions. |
| Not coded | altitude | What the video argues. The library records no altitude, and this report does not code it from transcripts. |
Cite this report
Jai Ghosh, Advids. We Audited 48 Agentic Demo Videos: 3 Trends in Pitching AI Bots to the Modern Enterprise. 14 September 2026. https://advids.co/blog/agentic-ai-demo-video-trends
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.
Image credits and licensing
Video frames on this page are reproduced for analysis and commentary. Copyright in each frame remains with the company credited in its caption; to license a frame, contact that rights holder. Advids grants no rights over them.
Charts and images credited ADVIDS Research, and the coded dataset behind them, are © 2026 Advids and licensed CC BY 4.0: reuse them with attribution to Advids and a link to this page.