AI policy
Where we use AI, where people decide, and how you are told.
We sell AI video production, so you are entitled to know exactly what that means before you buy it. This is that answer, in writing, rather than a sentence in a contract.
AI is a production capability, not a positioning claim
The reason to produce with AI is not that it costs less. It is that a concept can be tested in days instead of quarters, in twelve variants instead of one, before anyone commits a budget to a shoot.
That is a real advantage and it is narrow. It does not make the creative better, it does not replace the person who decides what to test, and it does not make a bad hypothesis worth running twelve times. We use AI where it makes iteration possible, and people where it makes the work good.
Read this page alongside the AI Video service — every commitment here is one that page’s mechanism lines already make.
Where AI is used
- 01
Variant generation
Turning one approved concept into the twelve versions a paid account actually needs — different hooks, framings and objections. This is the main reason we use it: it makes iteration possible at a speed a shoot cannot match.
- 02
Synthetic presenters
Generated on-camera performers, where a concept calls for one and a live shoot is not justified by the spend. Always labelled to you, and never presented as a real customer.
- 03
Product scene composition
Placing a real product into generated environments, so a single product shoot can support many settings.
- 04
Localisation and voice
Language variants and voice-over for markets where producing separately would not be justified by the media budget behind them.
- 05
Analysis and drafting support
Summarising account data and drafting first-pass copy. A specialist decides what any of it means; the tool never decides anything.
Where it is not, under any circumstances
- Anything presented as a real customer, a real review or a real testimonial. Ever.
- Anything implying an endorsement, a partnership or a person who does not exist.
- Generating a claim about a product. Claims come from the client and are their responsibility.
- Deciding where budget goes. Automation inside the ad platforms is supervised by a named specialist, which is most of what our media practice is for.
- Recreating a real person’s likeness or voice without their written permission.
Where a human reviews
Every asset gets an editorial pass
No generated output reaches an ad account unreviewed. A person watches it, in full, and is accountable for it going live.
The concept is human before the tool touches it
Variants differ by hypothesis — hook, offer framing, objection — and the hypothesis is written by a creative strategist. AI executes variants of an idea; it does not supply the idea.
Claims are checked against source
Anything a generated asset asserts about a product is checked against what the client has told us is true, because a generative model will produce a confident sentence either way.
A named person signs off
The AI Video Producer owns the pipeline; final editorial approval sits outside that discipline, on purpose. The boundary is published with every other one.
How AI-generated creative is disclosed
Our disclosure philosophy is short: the client always knows, the platform always knows where it requires to, and the audience is never deceived about whether a person is real.
- You are told which assets are synthetic, which are live-action, and which are a mix — asset by asset, not as a blanket note in a contract.
- Platform disclosure requirements are respected on every platform we run. Where a platform requires an AI-content declaration, it is made.
- A synthetic presenter is never labelled to your audience as a real customer, and never dressed as testimonial footage.
- If you would rather nothing synthetic ran on your account, say so and we will produce without it. It costs more and moves slower, and that is a legitimate trade to make.
Data and privacy principles
Your data is not training data
We do not submit your account data, your customer data or your unreleased creative to a model provider for training. Where a tool offers a training opt-out, it is switched off before the tool is used on client work.
Least data, not most
Analysis is done on the smallest extract that answers the question. There is no reason for a model to see a customer list, so it does not.
Personal data stays out
Customer records, email lists and anything identifying an individual are not put into a generative tool at all.
You own the output
Assets produced for you are yours, on the same terms as everything else we make. What we cannot do is warrant that a generative model produced something wholly original, because no one honestly can — so anything at legal risk gets a human-produced alternative.
This is a description of how we work, not a certification. We hold no AI audit, no compliance accreditation and no regulatory approval, and we will not imply otherwise. What we will do is answer a specific question about a specific tool, in writing, whenever you ask one. The data processing addendum covers the contractual side.
Ask us something specific.
If there is a tool, a model or a disclosure rule you need a straight answer on before you can work with us, ask. We would rather answer it now than in a contract review.