Machines for the plumbing. Humans for the work.

Key takeaways

• DOT does not use AI for creative work. No AI-generated scripts, concepts, storyboards or brand copy.
• We do use it for infrastructure — transcription, rough asset sorting, technical clean-up. Work where the output is measurable rather than judged.
• Client footage is never used as training data, never licensed to model providers, and never fed into generative tools as reference.
• Synthetic video is getting more recognisable, not less. Audiences are calibrating faster than the tools are improving.
• The commercial risk is asymmetric: AI-generated brand video saves a modest amount and can cost you the trust the film existed to build.

What counts as AI-generated video?

AI-generated video is moving image produced by a generative model from a text or image prompt, rather than captured with a camera or built by an animator. The term is used loosely, and that looseness is where most of the confusion in brand conversations comes from.

It is worth separating three different things, because they carry completely different risks:

• Generated footage — a model produces the imagery itself. Nothing was filmed.
• Generated creative — a model writes the script, the concept or the brand copy. Something may have been filmed, but the idea was not authored by a person.
• AI-assisted post — real footage, real creative, with machine assistance on technical tasks: transcription, object removal, upscaling, rough sorting.

DOT uses the third. We do not use the first two. That is the whole position, and everything below is the reasoning behind it.

Where does AI genuinely help in video production?

In the parts of production where the correct answer is objective and a human doing it adds nothing except hours. There is a real amount of this in any production, and refusing to use tools here would be posturing rather than principle.

• Transcription and logging. Six hours of interview footage becomes a searchable transcript in minutes. This genuinely changes what is possible in an edit, because you can find the moment you half-remember instead of scrubbing for it.
• Rough asset sorting. Grouping takes, flagging usable audio, sorting by shot type. Nobody misses doing this by hand.
• Technical clean-up. Removing a boom shadow, a stray cable, a reflection in a window. Retouching that has been possible for years, now faster.
• Upscaling and stabilisation. Rescuing a shot that would otherwise be unusable.
• Subtitle drafting. A first pass a human then corrects, which is meaningfully quicker than starting from nothing.

What connects these: there is a right answer, and you can tell whether the machine got it. The transcript is accurate or it isn't. The cable is gone or it isn't. Nobody's judgement is being outsourced.

Where does AI actively cost you?

Everywhere the output is judged rather than measured. A concept, a script, a performance, a cut — these have no correct answer, only a defensible one, and defensibility is exactly what a generative model cannot supply.

The practical failure is more specific than "it looks fake". Generative models produce the median of what they were trained on. Ask for a brand film about craftsmanship and you get the average of every brand film about craftsmanship ever published. That is a competent, familiar, entirely forgettable film — and forgettable is the one outcome a brand film cannot afford, because the whole point was to be distinguishable from competitors making the same claims.

There is a second-order problem. Your competitors have access to the same models and the same prompts. Differentiation built on a tool everyone can use is not differentiation; it is a temporary head start that closes in months.

Audience trust requires actual human authorship. The synthetic alternatives are becoming more obviously synthetic, not less.

Are audiences actually able to tell?

Increasingly, yes — and the direction of travel matters more than the current state. Detection is not primarily technical. Viewers who cannot articulate what is wrong with a synthetic clip still report that something is, and they report it faster the more synthetic media they have seen.

That is the asymmetry people miss. Model quality improves gradually. Audience calibration improves continuously and for free, because every synthetic video anyone watches is a training sample. A film that reads as authentic today may read as generated in eighteen months without a frame of it changing.

For campaign content with a six-week flight, that risk is tolerable. For a brand film meant to serve for three years, it is not. The longer an asset is supposed to live, the worse a bet synthetic imagery becomes.

What is the actual commercial trade-off?

Set aside the principle and the maths still does not work for brand-level video. The saving is real but modest: generative tools compress some production stages, but concept, direction, edit, sound and approval remain where the time goes.

Against that, the downside is not "a slightly worse film". It is a film that fails at the specific job it was commissioned for. A recruitment film that reads as synthetic actively repels candidates who were already sceptical. A customer story that looks generated does not just fail to build trust, it subtracts from the trust the customer's own name lent you.

The pattern is consistent. Where the content's job is to be believed, synthetic imagery is a poor trade at any saving. Where its job is to be understood, it matters far less — which is why animation and explainer work sit differently in this argument, and why we are not absolutist about it.

What happens to your footage?

A question we are asked more often every quarter, and one that deserves a direct answer rather than a policy page nobody reads.

Material we shoot for you is not licensed to model providers, not supplied as training data, and not used as reference inside generative tools. It is not uploaded to services whose terms claim rights over what passes through them. That applies to raw footage, to finished films, and to anything your people said on camera.

This is worth asking every supplier you work with, and worth asking specifically. "Do you use AI?" is too vague to be useful — almost everyone does, somewhere. The questions that separate suppliers are: what happens to our material, which tools does it pass through, and what do those tools' terms permit.

How should a brand decide?

Not on principle, and not on cost. On what the specific piece of content has to achieve.

1. Ask what the content's job is. Be believed, or be understood? That single question resolves most cases.
2. Ask how long it has to live. Six weeks tolerates risk that three years does not.
3. Ask who the audience is. Engineers, clinicians and senior candidates detect synthetic media faster than general consumers, and discount it harder.
4. Ask what happens if it is spotted. For most brand work the answer is not "mild embarrassment" — it is that the claim the film was making stops working.
5. Ask your supplier the specific questions about your material, not the general one about their AI policy.

Our own line follows from that reasoning rather than from a philosophical objection: machines for the plumbing, humans for the work. We are happy to have that argued with, and we would rather state it plainly than let a prospect find out mid-project.

Related reading

• The modern playbook for corporate video production in the Netherlands
• What separates B2B video marketing that works from video that just looks good
• Why boutique video production agencies are outpacing the giants
• Brand film production