I use AI in the edit. It doesn't do the editing.

Clients have started asking me whether AI can just do this now. It's a fair question and it deserves a straight answer rather than a defensive one, so here it is from inside the edit.

I do use AI. Most editors I know do. But it does none of the editing, and the distinction matters if you're the one paying for the result.

What it's genuinely good at

The honest list is longer than people who sell editing like to admit:

• Transcription. Two hours of interview becomes searchable text in minutes. This used to be an afternoon.
• Rough selects. Point it at ten interviews and ask where somebody talks about pricing, and it will find the moments.
• Audio cleanup. Room tone, hum, a bad lavalier — these are solved problems now.
• Masking and rotoscoping. Work that was genuinely tedious and is now mostly not.
• Subtitle timing. Still needs checking, but the first pass is close.

Notice what those have in common. They're all labour. None of them is a decision. Every one is a job I was happy to stop doing by hand, and none of them is the reason a client hires me.

What it can't do

Decide what the film is about. Ten interviews don't contain one story until somebody chooses which one to tell. That choice isn't in the footage — it comes from understanding what the client needs the film to achieve, which is usually not what the brief says.

Know which take carries it. Two takes can be identical on paper. One of them lands and the other doesn't. The difference is a half-second of hesitation before an answer, or an eye-line that reads as honest. A model scoring transcripts picks the clearest sentence. The clearest sentence is often the least true one.

Hold a pace. Knowing when to sit on a shot two seconds longer than is comfortable, and when to cut away before the viewer is ready — that's the whole craft. It's felt, against a specific audience, in a specific context.

Be accountable. When a piece goes out under a brand's name, someone has to have made the calls and be able to defend them.

Where it actually goes wrong

The failures are rarely dramatic. Nothing explodes. It's that the tool is confidently slightly wrong, in ways you only catch if you already know what right looks like.

Transcription outside English. I work in English, Portuguese and Spanish, and the gap is obvious. English transcripts come back near-perfect. Portuguese comes back readable but wrong in the places that matter — names, industry terms, and anything where a speaker switches languages mid-sentence, which in my work happens constantly. It doesn't flag uncertainty. It writes a plausible word and moves on. If you cut from the transcript without watching, you'll cut a sentence the person didn't say.

Selects that optimise for clarity. Ask a tool for the best answer to a question and it returns the most articulate one. But in a testimonial, the most articulate answer is often the most rehearsed, and rehearsed doesn't persuade. The take you want is usually the one where somebody pauses, corrects themselves, and then says the true thing. On a transcript that looks like the worse option.

Auto-reframe. Useful for turning a landscape cut into vertical, right up until the moment two people are talking and it decides which one matters. It follows movement, not meaning, so it will drift off the person listening — and in an interview, the reaction is frequently the shot.

Noise reduction pushed too far. It's excellent at removing hum. It is also happy to remove the room, and a voice with no room around it sounds like a voice in a box. Nobody can say why the video feels cheap; it just does.

Automatic colour matching. It will make your shots consistent by making them average. If the look was deliberate — warm, cool, deliberately flat — average is exactly wrong.

None of this makes the tools bad. It makes them tools. Every one of these is fine when somebody is watching the output and knows what they're looking for.

How I use it

Sparingly, and always pointed at something. A transcript so I can navigate an interview by text instead of scrubbing. Noise reduction on a take that's otherwise the best one. Never for the structure, never for the selects, never for the pacing.

The tool is fast at the parts that were slow. It's useless at the parts that were hard.

If you're the one hiring

An AI-assembled edit tends to look fine the first time you watch it and hollow the second. The cuts are on the beat, the information is present, and nothing quite lands. If your video exists to persuade somebody — to trust a brand, to take a training seriously, to book something — then pacing is the persuasion, and that's precisely the part nothing has automated.

Use the tools. Just don't confuse the labour with the work.