Skip to the content

Case study

When the AI video tool changed its interface, the ads stopped

Presenter lines stopped being made while the screen still promised a finish time of a few hours, and tall clips came out wide. What a quiet change at an AI video service did, and how to check for it.

Key takeaways

  • A service that changes its screens usually tells nobody, so the only reliable check is what your automation actually produced.
  • An automation can read a missing button as a fact about the account, which turns a small screen change into work that never starts.
  • Keeping one account outside the automation gives you a safe place to see the service as it is today.
  • A finish time guessed from past speed is worse than a plain message saying the job is stuck at the service, with the reason.
  • Before giving up something the owner cares about, check whether a later step in the process already takes care of it.

The presenters in the ads stopped being made. Nothing failed and no error appeared: the lines simply waited, while the screen went on showing a finish time of a few hours. At the same time, the short background clips that should be tall for phones started coming out wide.

That is how a quiet breakage looks when your daily work runs on someone else's software. For one client, our engineer traced it to a change the AI video service had made to its screens, and had the automation working with both the old screens and the new ones within about eight hours of one day.

Work that stopped without failing

The client's video team makes its ads with an outside AI video service, driven automatically through many accounts on that service. When the presenters stopped, nothing in the process reported a fault. The screen kept promising a finish time of a few hours, so a team watching that screen had no reason to look for a problem. One ad went almost a day without a presenter clip while its screen still promised a few hours.

Other faults ran alongside it. Background clips asked for in the tall shape came out wide, and now and then a request reached the service with only the photo and none of the words. The lost words cost twice: the clip is wrong, and the attempt has still been spent from what the account is allowed.

When the AI video tool changed its interface

The service was moving accounts to a new interface a few at a time over about a week. The automation was built on the old one, and it read every missing button as a fact about the account rather than a change to the screen. A missing voice menu became an account with no voices. A missing picture-shape menu became no instruction at all. An empty text box became a request ready to send.

This is the ordinary risk of building on someone else's software, and it is why automation output that comes out wrong is harder to catch than automation stopped working after update. A job that fails shows up. A job that finishes and produces the wrong thing looks like work being done.

A third-party tool changed without warning, and the automation carried on confidently. Because screens arrive a few accounts at a time, colleagues on the same service can be looking at different things on the same day.

How our engineer found it

The trail started with the presenter lines. The automation had marked a growing share of accounts as having no voices, so it gave them no presenter work, yet the voices on those accounts were fine.

Opening one of those accounts by hand was the turning point. The service's new interface was in place, and in it the video maker is a skill inside a chat; the menu that used to list voices is gone. A missing menu had been read as an account without voices.

Before touching anything, we added a way to take a single account out of the automatic work, so a person could use it by hand without the automation interfering. That account then showed the other faults plainly. The new interface has no menu for the shape of the picture, so a clip asked for in words alone came out wide. And the box where the words are typed sometimes held nothing, so the request went with the photo alone. Before the fix, about nine in ten of the tall background clips made on new-interface accounts had come out wide.

What we changed

The automation now works with both interfaces. On the new one, the tall shape is said in the words of the request itself, a tall video, taller than wide, which gives the tall clip the old menu used to give.

The words are now read back out of the box before the request is sent. If they are not there, they are typed in again, then pasted in a second way; if they still do not stay, nothing is sent, and the log records what the box held. That way an attempt is not spent on a request carrying only a photo.

The wrong no-voices marks were lifted.

The finish time no longer misleads. An ad with no new presenter clip for 45 minutes is now shown as stalled at the video service, with the reason in plain words, and the ads waiting behind it say so too.

The voice decision

The hardest part was the voice, and the answer came from the owner. On the new interface a voice can only be set by describing it in words, and that gives the service's own voice for the photo rather than the one the team chose.

The owner's first instruction was that the voice must never change, so new-interface accounts were given no presenter work at all. Then the owner made a point we had not weighed: every ad made over ten days had its chosen voice applied afterwards in any case. With that in view, ads are again made on any account with the voice described in words, and only the work that must keep the service's own voice stays on the old interface.

What the business has now

The change has been live on the client's platform since the same evening. In tries by hand on the account we had set aside, presenter lines came out in well under a minute each in the voice described, and a background clip asked for as tall came out tall. The presenters stood still for about a day before the cause was found; the pause for hand use, the voice, the prompt box, the picture shape and the honest finish time were all built within about eight hours of one day.

A full week of ads made since has not been measured yet, so there is nothing to say about that. Tracing a quiet stoppage back to the tool that caused it, and making the automation admit when it is stuck, is the kind of work our engineers do for clients.

What you can check yourself

  • Look at a sample of what your automation actually produced this week, not only at whether each job finished.
  • Keep one account with the outside service that you use by hand, outside the automation, so you can see what that service's customers see today.
  • When a finish time on your screen has not moved for hours, ask what it is based on.
  • Read the change notes or the news of the services your work depends on at least once a week.

Questions people ask

How do I find out that an outside tool has changed before a deadline is missed?

Watch what comes out, not only whether the job finished. A service that changes its screens usually does not send a message, and the work either stops silently or comes out slightly wrong. A quick look at a sample of this week's output, including the small details like shape and wording, shows more than a queue of completed jobs.

Why did the automation mark good accounts as having no voices?

Because the menu it used to read the voices from was no longer there. The automation treated a missing button as a fact about the account rather than a change to the screen, so it stopped giving those accounts presenter work. The voices on those accounts were never the problem.

Why did tall video clips come out wide after the update?

The new interface had no menu for the shape of the picture. A clip asked for in words alone was made in the default wide shape, because nothing told the service otherwise. Saying the shape in the words of the request itself, a tall video, taller than wide, gave the tall clip the old menu used to give.

Should I keep one account outside my automation?

Yes, and it is worth doing before anything goes wrong. An account you can use by hand shows you what the service's customers see today, which is often different from what your automation assumes. It also gives you somewhere to try things without disturbing live work.

Why does my screen show a finish time when nothing is moving?

Often the time is worked out from how long the job took before, not from anything the service is reporting now. When the work has quietly stopped, that time keeps promising what will not arrive. Ask whoever looks after the automation what the time is based on, and whether it can say stalled instead.

From Tech Concierge

Want this handled for your business?

Your own engineer builds and runs the website, the app, the CRM and the AI assistant that answers your customers.

The first 7 days are free.

Read next

Case studyAI and automation

Why your AI assistant invents an upload button that does not exist

A client asked our own chat assistant how to send us two page designs. It sent them to an upload button and an email address that did not exist. We changed the portal, the team's view and the assistant's instructions, and added a check before replies leave.

5 min read

Case studyAI and automation

Why an AI tool stop button still running costs you

An AI tool's Stop button often stops only the visible part of the job, while queued and late work carries on being paid for. We traced why in a tool we built for a client, and what a real stop has to do.

4 min read

Case studyAI and automation

Why jobs finish out of order and older ones wait all day

Ads queued in the evening were still running the next day while newer ones finished. There was no queue: a free slot went to whichever job asked first. What we found, what we changed, and the questions to ask about any shared tool.

5 min read

We count visits to this blog with our own analytics, never Google's, and share what you read with nobody. May we remember this browser with a cookie? See the privacy policy.