Why maintaining product videos now costs more than making them

Earlier this month we published a study of 45,000 new videos. Its headline finding was about speed: the median video built by an AI agent was exported 15 minutes after someone started it. We wanted to know what happens to a video after that first export, so we went back to our product data and asked how often anyone opens one again.

We took 31,510 videos our customers published between October 2025 and July 2026 and checked whether anyone came back later to export each one again. Once a video was a month old, only 3.4% were ever exported again, about one in 30. Past the three-month mark, it was 1.6%.

A video nobody re-exported isn't necessarily wrong or a bad thing. Some of these videos show a settings page that hasn't changed in a year, and some teams delete an old video and make a new one instead of editing it. But, most of those videos show software, and the software kept shipping new features and screens while nearly all of those videos stayed exactly as they were on the day they were published.

I think this is where the cost of video has moved. AI has cut the first version of a product video down to minutes. The fifth version, the one you need after four releases have moved the screens it shows, still costs about what it always did, because someone has to record it again. For any team with more than a few dozen videos, keeping the library accurate now costs more than building it, and the only way to afford that is to stop re-recording videos by hand and rebuild them from the product itself.

Making a product video stopped being the hard part

A screen-recorded product video used to take a person through four separate jobs: record the workflow, write a script, record the voiceover, and edit all of it into something watchable. Each job needed a different skill, which is why so many companies had one person, or one agency, who made every video.

Most of that work is automated now. AI voiceover reads a script in whatever voice you choose, a rough narration can be cleaned up into the script, and the written step-by-step guide comes out of the same recording. AI agents go even further and build the whole video from a prompt. In our study, the median agent-built video reached export 15 minutes after it was created, after four rounds of revisions.

Faster production means bigger libraries. Agent-built videos went from almost none in May to 19% of all new videos on Clueso in October, and people who make videos both ways export 52% of their agent-built videos against 10% of their screen recordings. Every one of those videos shows screens that the next release might change, and the 15 minutes covers only the first version of each.

Every release sends a screen-recorded video back to the start

You can't edit a screen recording the way you edit a document. The pixels are the content, so when a button gets renamed, the old name is baked into every frame that shows it, and the narration still reads it out.

A customer education lead told us about a release that changed their product's navigation menu: "I had to update every single video with the new navigation." Navigation changes are the worst case. Almost every workflow starts by opening a menu, so one release can touch nearly every video in the library.

The work per video adds up quickly. Someone sets up the demo account again, records the workflow, matches the narration to the new timing, adds the zooms and callouts back, and exports. If you've translated the video, every language version needs the same treatment, and the written guide made from the old recording is now wrong as well.

It's worth doing this sum for your own library. Count the videos that pass through the part of the product your last release changed, then multiply by the hours your last re-record took, review included. Say a navigation change touches 40 of your 200 videos and each re-record takes three hours: that's 120 hours, or three working weeks for one person, before anyone makes a new video. Supporting four languages multiplies the recording and narration work by four.

Faced with that bill, most teams re-record the handful of videos that someone complains about and leave the rest alone. That matches what we see in our data, where only about one video in 30 was exported again after its first month. We've written separately about why product videos take so long to make in the first place.

A stale video doesn't look broken

Nothing tells you a video has gone stale. It still plays in your help center and still collects views, so every dashboard you have says it's working. The only problem is that the button it tells customers to click now has a different name.

Help articles go stale too, but text is easier to check. Someone can search a help center for an old label and find every article that uses it. Nobody can search a video's frames for an old menu name, so a stale video usually surfaces when a customer follows the old steps and opens a ticket, or when someone on the team happens to watch it for another reason.

One customer education lead described launching a course and hearing a week later that "the whole course has to be redone," because the product had changed underneath it. That team found out within a week. In a library nobody is checking, customers keep watching an outdated video for as long as it takes someone to notice.

Every customer who watches it in that time learns the old way of doing something, and some of them file a ticket about it. Teams that turn repeat support questions into videos feel this most, because an out-of-date answer creates the tickets it was meant to prevent.

Agents fix maintenance by rebuilding videos from the product

A better editor doesn't solve the maintenance problem. A team that edits twice as fast still has to work out which videos a release broke, and still has to capture the new screens. Those two jobs take most of the time, and both happen outside the editor.

Agents change this, because an agent can use your product the way a person on your team would. It can run a recorded workflow again on the current build, capture the screens as they look today, regenerate the narration to match, and hand back a draft.

Vamsi Duvvuri, whose firm Dharana Capital led our Series A, puts it this way: agents are turning video "from static media files into living, context-aware assets that continuously adapt as business processes and products change."

We built Clueso Agents to do this. A browser agent logs in to a demo or staging account in your product, navigates the workflow, records it, and builds the video, which arrives as an editable draft for your team to review. Seismic's customer education team uses Clueso to create product tutorial videos "and update them as the product evolves," and their Director of Customer Learning Experience says it has let the team "move faster while maintaining product accuracy and instructional quality."

A person still reviews every update before it goes out, because an agent can misread a workflow and an accurate-looking video of the wrong steps does real harm. Reviewing a draft takes minutes, though, where re-recording took hours. A team that could only afford to fix the videos customers complained about can now check every video a release touches.

Count your stale videos before you plan new ones

You can get a rough measure of your own maintenance backlog this afternoon. Pull up your release notes for the last six months. Then take your ten most-watched videos and play each one against the product as it is today, checking every screen and label.

Write down how many of the ten still match. For each one that doesn't, note the release that broke it and how many weeks it has been live since then. Multiply the failures by the hours your last re-record took, and put that number in front of whoever plans next quarter's video work, next to the list of new videos they've asked for.

Frequently asked questions

How do teams handle re-recording videos when the product changes constantly?

Most teams ration it: in our data, only about one video in 30 is exported again after its first month. Teams re-record the videos customers complain about and leave the rest, because a re-record means setting up the demo account, recording the workflow again, matching the narration, and redoing every translation. The alternative is to have an agent re-run the workflow on the current product and produce an updated draft, so the team's time goes into reviewing drafts instead of recording.

Can you update a training video without re-recording it?

For small changes, yes. If a label or a line of narration changed, you can edit the script, regenerate the voiceover, and replace only the clip that shows the old screen. A redesigned workflow usually needs a fresh capture of that part. We cover both cases in how to update a training video without re-recording it.

How do you find out which videos are out of date?

Compare the videos against your release notes. Keeping a record of which screens appear in which video helps: when a release changes a screen, you can look up every video that shows it instead of rewatching the library. The same approach works for finding old branding after a rebrand.

Why does AI make video maintenance more important?

AI made new videos fast to produce, so libraries are growing faster than before. Every extra video shows screens a future release can change, and recording a video again by hand takes as long as it always did. A team that builds 200 videos instead of 20 has ten times as many videos to keep accurate.

How do you keep a large training video library up to date?

Check the library against every release. Look up which videos show the parts of the product that changed, rebuild those from the current product, and review them before the old versions come down. Our guide to managing a training video library covers how to organise one so this is possible.