Principles for using AI on client work

Every engagement at Entatan uses AI. Models write the first pass of most code and copy, generate tests, review pull requests, and produce reports. That is the reason we can do what we do at the size we are. It also raises fair questions from partners: what is the model doing with our data, who is accountable when it is wrong, and how do we know what was made by a person.

We wrote down the rules we follow. They are published in full on the principles page. This is the short version, with the reasoning.

1. A person is accountable for every deliverable

Nothing ships to a partner without a named person on our team who reviewed it and will answer for it. Models are tools. Tools do not take responsibility. If an agent introduces a bug, that is our bug. If generated copy makes a claim the product cannot back, that is our claim to retract.

2. Your data trains nothing

Partner code, content, customer data, and ad performance are used to do your work and for nothing else. We do not fine-tune models on it, we do not pool it across partners, and we choose model providers whose terms match that commitment. Where a partner needs a specific provider excluded, we exclude it.

3. Generated is disclosed when it matters

Inside a codebase, whether a line was typed or generated is irrelevant once it has been reviewed. In public, it is not. Where generated imagery or video is used in a partner's advertising, we disclose it in the asset metadata and tell the partner, and we follow the disclosure rules of each platform we buy media on.

4. Agents act within a written boundary

Every agent we deploy has a document that says what it can do, what it must ask before doing, and what it may never do. Support agents draft; a person sends, until the partner decides otherwise. Agents that touch money, customer records, or public channels start in approval mode and earn autonomy with a track record.

5. We measure the model like we measure everything else

Generated work is held to the same number as human work. If an ad variant a model wrote underperforms, it gets cut. If an agent's replies drive lower resolution rates than the queue it replaced, it gets rolled back. We keep evaluation sets for every agent in production and run them on every change.

6. We tell you what we do not know

AI systems fail in ways that are new. When we are unsure whether an approach will work, we say so and run a small test before committing your budget. When something breaks in a way we did not anticipate, you hear it from us first, with what we changed.

Why publish this

Partly because partners ask. Mostly because writing rules down is how a small team keeps them when it gets busy. These principles will change as the tools change, and we will note every change on the page with a date.

The question is never whether to use the model. It is who is standing behind the output.