Why the AI in Our Testimonial Tool Is Not Allowed to Write

Akteora collects video, voice and text testimonials and turns the best moments into short captioned clips. AI is a big part of how that works. But one rule was set on the first day: the AI is never allowed to write a word.

The temptation

When you build a testimonial tool today, the obvious AI features are:

• “Clean up” what the customer said.
• Write a punchy headline quote or a summary.
• Generate captions that sound a little better than the real words.

They are easy to build and easy to sell. They are also a bad idea. A testimonial is worth something only because a real person said it. The moment software improves their words, it is no longer their testimonial. It is marketing copy with their face on it.

The legal side

There is a hard reason too. In the US, the FTC’s rule on fake reviews and testimonials took effect on 21 October 2024. It bans fake testimonials, including AI-generated ones, and it can reach the tools that help produce them, not only the business that publishes them.

Our customers publish testimonials under their own brand. If our software quietly changed what someone said, they would be the ones taking the risk. For the detail, see are AI-generated testimonials legal and our plain English guide to the FTC rule.

What the AI actually does

The model has one narrow job when it helps build a reel: find the best moments.

1. The customer records. A video or voice note on their phone, in their own words.
2. Speech becomes a transcript. With a start and end time for every single word.
3. The model reads the transcript. And answers with start and end times only. Numbers, not text.
4. Every answer is checked. Against the real transcript and the real recording. Anything that does not line up is rejected.
5. The real video is cut. At those times. Captions come straight from the transcript, word for word.

The checks are strict. A moment has to start exactly where a word starts and end exactly where a word ends, sit inside the recording, and be a sensible length. If the model’s answer fails, it gets one more try with the reasons. If it fails again, or no model is available, a simple rule-based picker takes over instead.

The model can choose. It cannot create. Every word in the final clip was actually said by the customer, and the person building the reel can move any cut themselves.

What about typos and filler words?

Sometimes people do want a small cleanup, like removing “um” from a spoken testimonial or fixing punctuation. Akteora allows that, carefully:

• The original is never changed. It is stored permanently.
• The cleanup is a separate version, offered to the person who gave the testimonial, before they submit it. It is used only if they choose it over their own.
• The cleanup is checked before they ever see it. If it contains any word they did not say, drops or changes a number, removes a “not” or a “never”, or changes the length by more than about a third, it is thrown away and their own words stay.

The business using Akteora cannot edit a customer’s words either. Not in the dashboard and not through the API. They can correct a job title or a company name, and that is all.

Why a narrow AI turned out better

• It is easy to check. Timestamps are numbers, so we can test them against the transcript automatically. Checking whether a rewritten sentence is “faithful” is far harder.
• It is cheaper and faster. The model reads text and returns a few numbers. It never watches the video and never writes long answers.
• It is easy to explain. We can tell customers exactly what the AI does in one sentence.

The trade-off is real: Akteora will never write you a catchy quote. That is on purpose.

The bigger lesson

For a few years the question was “what can AI do?”. Anywhere trust matters, such as reviews, health, money or legal documents, the better question is “what should AI be allowed to do?”.

Give the model the smallest job that is still useful, check every answer against real data, and keep people in charge of what is true.

Try Akteora free, or read how we built it.