--- title: "What we realized about B2B video production after analyzing 45,000 videos" description: "We analyzed 45,000 new videos. AI-agent-built videos went from zero to 19% in five months, and they get exported 2.5× as often as screen recordings." url: "https://www.clueso.io/blog/ai-agent-video-study" updated: "2026-10-22T09:02:00.322Z" --- # What we realized about B2B video production after analyzing 45,000 videos We analyzed 45,000 new videos. AI-agent-built videos went from zero to 19% in five months, and they get exported 2.5× as often as screen recordings. - Agent-built videos went from 7 in May to 3,314 in the first three weeks of October, which is 19% of every new video on Clueso. - Most of them start in Claude, Claude Code, or ChatGPT, not in a code editor. - People who make videos both ways export 52% of their agent-built videos and 10% of their screen recordings. - The median agent-built video is exported 15 minutes after it's created, after four rounds of revisions, and runs 63 seconds. ## Introduction We launched [Clueso Agents](https://www.clueso.io/agents) in early July. The premise was simple. Agents navigated your product and created videos on demand based on whatever input you gave them. They could also identify old videos that needed an update by identifying code changes in your product and determining which processes or screens had gone out of date. Since then, adoption for Clueso Agents has been phenomenal. In fact, in October, our users have made 3,314 videos so far, and are five times as likely to finish those videos compared to the ones they recorded themselves. But what's causing people to use agents for video production so much? AI makes mistakes, well a lot of them. But ever since our launch we've seen an almost meteoric rise in video production done primarily through AI agents. To understand what was going on, we went through three months of product data: about 45,000 new videos, with every Clueso employee and every workspace any of us belongs to removed from the numbers (we make a lot of videos internally!). This is what it shows about who makes videos with agents, what those videos look like, and why so many more of them get finished. ## AI agents built 19% of new Clueso videos in October ## AI agents built 19% of new Clueso videos in October The first agent-built videos showed up in our data in May 2026. That was when Clueso Agents was still in beta and was only offered to a select few users. By July, after the official release, that number rose to over 1,500 videos a month. ![A graph showing AI agent led video production numbers over the last three months](https://d37f9k611f70xr.cloudfront.net/agent-share-chart.svg) _Share of new videos built by AI agents, May to October 2026._ By October, agent-built videos made up 19% of every new video on Clueso, and October was only three weeks old when we pulled the data. It's also the first month people built more videos through an external LLM via [our MCP](https://www.clueso.io/mcp) than through the AI chat inside our own product: 3,314 against 3,253. We find that very encouraging. These aren't one-off experiments, either. The 545 people building with agents in October made about six videos each. ## It isn't (just) developers ## It isn't (just) developers One would normally assume that heavy MCP usage would indicate developers are at it again on their terminals, but non-developers have been rapidly adopting Claude Code and Codex as the starting point for all their projects. Over the last 90 days, people built videos with Clueso Agents from: - Claude, on the web or desktop app: 628 people - Claude Code: 518 people - ChatGPT: 459 people - Codex: 71 people Claude Code makes sense for teams documenting software, since that's where most release work already happens. But the two biggest groups are general-purpose assistants, used by people who would never call themselves developers. A product marketer who drafts launch copy in ChatGPT can ask the same assistant to turn it into a narrated walkthrough, and plenty of them do. So if your trainers and product marketers already spend their day in a chat assistant, that's where your video production is likely to start too. ## Agent-built videos are exported five times as often as screen recordings ## Agent-built videos are exported five times as often as screen recordings This is the number we kept re-running. Of the videos people started between mid-July and mid-October, 53% of the agent-built ones were exported. Compared to this figure, screen recorded videos sat at 21% and videos made with prompts inside Clueso were at 33%. Our initial assumption was that agent users are just keener, more technically proficient and motivated; especially on teams that ship a lot of videos. So we compared people with themselves. > Among the 710 people who made videos both ways, 52% of their agent-built videos were exported. Only 10% of their screen recordings were. Same people but five times the finish rate. We also widened the scope to understand how these videos were being used. People usually were publishing agent made videos to a help center, copying an embed code for landing pages, or sharing a link with other people. Agent-built videos still came out at 55%, and recordings at 24%. We feel some of this is also intent. Plenty of recordings are rehearsals, false starts, or a capture someone meant to come back to on Friday. (That Friday never came.) But when you ask an agent for a video, you usually want a finished video that can be shared externally. As more and more people are starting to use Clueso Agents, their confidence in AI being able to create a polished video end-to-end is increasing. A lot of this also can be attributed to how accurately Clueso Agents are able to navigate products, understand a product's branding, and stay close the the prompts people are providing when generating media from scratch. ## Most agent-built videos are revised and exported in one sitting, in about 15 minutes We also wanted to understand how efficient users felt when working with agent-built videos. The median agent-built video was exported 15 minutes after it was created, and 78% were out for external purposes within the hour. Screen recordings took a median of 43 minutes, and 57% made it out within the hour. Of course, people aren't just taking the first draft and running with it. Before its first export, the median agent-built video went through 31 tool calls (the agent adding a clip, writing narration, moving a callout) and four rounds of revisions: retime this, cut that, try a different voice. 98% got at least one. This shows people are bringing in their taste and vision when directing these videos, not rubber-stamping them. What has dramatically reduced though, is the back-and-forth. Of all the videos that we analyzed, **84%** of agent-built videos were finished in a single working session, inside one conversation, instead of across a week of handoffs between whoever recorded it, whoever edits it, and whoever signs it off. ## Agent-built videos are half as long as screen recordings ## Agent-built videos are half as long as screen recordings The median exported agent-built video ran 63 seconds, and two-thirds come in at 90 seconds or less. On the other hand, the median exported screen recording ran 114 seconds, and only 39% made it under 90. It makes sense once you think about how these two videos start. A recording lasts as long as the task took, including the part where you hunted for the settings menu. A video built from a description lasts as long as the description. They sound different, too. Two-thirds use an AI voiceover, usually in a specific voice someone asked for rather than the default. 43% add music, and about one in eight includes AI-generated visuals. Users who are sensitive to how many minutes agent-built videos are consuming are also seeing better results with this approach. Since these videos don't spend time hunting for buttons or screens, they already know what to present and when to present it to a viewer and offer better pacing overall. ## So, are screen recording based videos dead? (Not really) ## So, are screen recording based videos dead? (Not really) None of this means recording your screen is the wrong way to make a video. The recordings in our data aren't stalling because people are bad at making videos (We promise). They're stalling where videos have always stalled: after the capture, when someone has to turn a long recording into 90 seconds that make sense, add narration, fix the part that went wrong, and get it signed off. That gap between captured and finished is where most training and product video end up taking time. Agents are closing that gap by doing the turning-it-into-a-polished-video part much faster and in the same conversation where you said what you wanted. You still review it. You still change the voice and cut the clip that doesn't look right. You just don't have to open an editor, find the timeline, and do it all by hand. ## Should your team make videos with an AI agent? ## Should your team make videos with an AI agent? That's the most important question for teams struggling with producing videos at scale. We'd say if your team's video backlog is mostly recordings nobody finished, our data says it's worth changing the starting point (process) before you change the people. Every agent-built video in this study was made through the [Clueso MCP](https://www.clueso.io/mcp), which lets Claude, ChatGPT, Cursor, and other assistants build and edit Clueso videos directly: clips, narration in a voice you pick, callouts, captions, and the export. Connect the assistant you already use and ask it for the next video on your list. We'll be counting. ## Methodology ## Methodology We used aggregated product analytics from Clueso. We read no individual person's, company's, or video's data, and every figure is an aggregate. The smallest group we report is Cursor users, at 16 people. We excluded every Clueso employee, every workspace a Clueso employee belongs to, and all development and test environments. Adoption covers new videos by month from May to October 21, 2026. Outcomes cover videos created between about July 16 and October 14, 2026, so every video had at least a week to be exported. "Exported" means at least one finished video file was rendered. "Shipped" adds publishing to a help center, using an embed code to publish the video elsewhere, or sharing a link externally. Medians are used throughout, because a small number of very long or very active sessions would skew averages. ## FAQ ## FAQ ### How many people use AI agents to make videos? On Clueso, more than 1,100 people in over 1,100 workspaces built at least one video through an AI agent between mid-July and mid-October 2026. In the first three weeks of October, agent-built videos made up 19% of all new videos, up from under 1% in June. ### Are videos made with AI agents more likely to get finished? In our data, yes. 53% of agent-built videos were exported, against 21% of screen recordings. Comparing the same 710 people across both methods, 52% of their agent-built videos were exported, against 10% of their recordings. ### Which AI assistants do people use to make videos? Over 90 days, the most common were Claude on web and desktop (628 people), Claude Code (518), and ChatGPT (459), followed by Codex (71) and Cursor (16). ### How long are videos made with AI agents? The median exported agent-built video in our data runs 63 seconds, and two-thirds are 90 seconds or shorter. That's about half the median length of exported screen recordings, which run 114 seconds.