AI at SANE/REBELS · 01
How we use AI
The first principles everything else follows from.
The first principle
Where we start
Three truths that everything else follows from.
We feed the systems we rely on instead of fighting them. Today a platform prices its auction better than we could by hand, and a model carries a method more faithfully than a summary. Both do their best work when we hand them the right inputs and limits. We focus on the inputs, and if something smarter comes along, we switch to it.
Where there's a right answer, code does the math, not AI. Numbers, checks, and carrying out a change are handled by tested code. We use AI only where there's no clear right answer: which idea, which copy, what comes first.
AI has become cheap. Good judgment hasn't. AI produces analyses and copy in minutes. The decision is what stays valuable, and a human makes it. In a controlled trial, experienced developers with AI even came out 19 percent slower while feeling faster, with most of the time going to reviewing and correcting the output.
That's exactly what the next three pages show.
Two kinds of work
Every task falls into one of two classes. The dividing line decides whether tested code does it or a model does, and it's the most important design choice we make.
Has a right answer
Code does this
Doing the math, running a check, carrying out a change, repeating a routine. That belongs in tested code: the same every time, logged, and reversible. A model has no place here; it would only add variance where there must be none.
Has no right answer
A model does this
Which idea, which line, what comes first, what a pile of data actually says. Here a model is fast and useful. What it produces is always a proposal or an explanation, never the final word.
Between them sits the approval
Between understand and act sits the approval. Nothing goes live without a human approving it, and a change counts as done only once a follow-up check confirms it.
Three controls carry every use. Check the inputs first, then a human approval, then the follow-up check. Nothing goes live without a human confirming the exact change, and a thing counts as done only once a check confirms it.
Grounded
Context beats cleverness
A model is only as honest as its source. Instead of letting it talk from memory, we give it one fixed, versioned source: your profile in the Brand Hub, the things no model knows on its own.
It answers from your material. From the structured corpus, not from the model's general knowledge.
What it can't back, it flags. An unbacked number is a working theory, not a fact.
The voice stays yours. A check runs before every output: banned words, AI filler, length. If a line reads like generic AI, it doesn't ship.
It discloses that it's AI. Always, not only where a law requires it.
One profile, shared by everything. The same Brand Hub grounds a built twin, an ad, and the daily operation. One source, reviewed and versioned. How it steers the day is in AI in Ops.
Data security & processing
Three layers: what's built into the system, what we commit to contractually, and the foundation the models run on.
How the system handles data
Decision outside the model. What goes live is deterministic code, not an LLM.
Human in the middle. Paused by default, every live action approved individually, nothing automated.
Data minimization. Only the necessary sources, read-only where possible. End-customer PII is never touched.
Traceable. Every action is logged and auditable.
What we commit to contractually
AI disclosed, with oversight. AI use is stated in our T&C (§ 11); you are the provider toward end users (§ 11.2), and the final compliance review of published content stays with the client.
DPA under Art. 28 GDPR. If we process personal data on your behalf, it runs under the data processing agreement included as Annex A (§ 9.2).
Purpose limitation & subprocessors. Processing only to fulfill the contract; subprocessors contractually bound to data protection.
The foundation: the model provider
No training on your data. API data doesn't flow into model training.
GDPR. DPA with EU standard contractual clauses available.
Retention. Per the provider's standard terms; zero data retention available as an option.
GDPR role: S/R acts as a data processor; the data processed is marketing, product, and account data. Details in our T&C (§ 11 AI, § 9.2 data protection, DPA as Annex A) and the privacy policy. Provider details per the Anthropic Trust Center, as of June 2026. This page is information, not legal advice.
Up close
Where this goes
The same principles, three places. Each one shows how the loop actually runs there.
02
Our take on agentic coding
How we turn an expert's method into software that reproduces it faithfully.
03
Our AIs speak Google Ads
Paid media in a checked loop: data-driven, every live change approved by a human.
04
Meta meets AI
Paid social where the creative carries it: AI drafts in your brand voice, a human does the upload.
05
AI in Ops
The operating system behind the day: deterministic code for the safe part, AI for the judgment.
Sources
What this stands on
Our approach doesn't stand alone. It lines up with the current research and practice on AI-assisted development.
“Generation is solved. Verification, judgment, and direction are the new craft.”
- 1Andrej Karpathy, “Vibe coding,” the original post (2025)
- 2Osmani, Saboo & Kartakis, Google, The New SDLC with Vibe Coding: from ad-hoc prompting to agentic engineering (2026)
- 3METR, Measuring the impact of early-2025 AI on experienced open-source developer productivity (2025)
- 4Anthropic, Effective context engineering for AI agents (2025)
Cite thisIf you use or build on these ideas, cite us as the source: SANE/REBELS (KNUS GmbH), “How we use AI,” sanerebels.com/about/ai, 2026.
See it on a real example?
30 minutes, the whole loop live. You see what gets proposed, what a human approves, and what lands in the end.
Request a callRelated content
Last updated June 2026
Citation rule: if you use or build on these ideas, credit SANE/REBELS (KNUS GmbH) and link to sanerebels.com/about/ai.