--- title: "Human centered. AI first. What we mean by that" lang: en type: blog canonical: https://www.arocom.de/en/blog/human-centered-ai-first alternate: https://www.arocom.de/blog/human-centered-ai-first author: "Axel Roth" published: 2026-06-02 last_updated: 2026-10-31 --- # Human centered. AI first. What we mean by that "Human Centered. AI First." is arocom's guiding principle for digital projects: every task is checked first to see whether AI can support it. Goals, judgement and responsibility stay with people. This article explains why neither "AI replaces people" nor "AI is just a tool" holds up as an attitude, spells out the three rules behind the principle (human approval, a named person as the sender, transparency about AI use) and names the tasks we deliberately do not delegate to AI. ## Human centered. AI first. What we mean by that Two terms that seem to pull in opposite directions: *human centered* and *AI first*. We placed them side by side on purpose, because the right attitude to AI lives in that tension. If you only shout "AI first", you optimise processes and lose people. If you only say "human centered", you risk an organisation that is no longer competitive in five years. Both together is harder, but right. The consequences of this stance, from pricing to jobs, have been set out in our [AI manifesto](https://www.arocom.de/en/ai-principles) since September 2026. ## Why "replacement" and "just a tool" both fall short Before we explain what we mean, it is worth looking at the two sentences that dominate the debate. The first: "AI will replace people." The second: "AI is just a tool." In conversations with managing directors we hear both regularly, sometimes in the same meeting. Both sound decisive, and both fall short. The replacement claim treats work as a sum of individual tasks. A model can write a proposal draft. It cannot vouch for the proposal being correct and being honoured. Responsibility, trust and judgement are tied to people and cannot be installed. If you cut jobs first and think later, you lose exactly the people who can still judge AI output. The tool claim sounds more reasonable, but mostly it reassures. A tool in the classic sense waits until you pick it up. A system that drafts texts, writes code and weighs options by itself changes how work is distributed, whether you call it a tool or not. In many companies, "just a tool" translates to: "We would rather not engage with this seriously yet." In 2026, hardly any organisation can afford that postponement. Our guiding principle therefore holds on to both ends. "AI first" describes the standard we set for how we work, "human centered" the order of responsibility. The following sections show what that looks like in everyday work. ## AI first: the default question, not the default answer For us, "AI first" means this: for every task, we first ask whether and how AI can do it better, faster or more thoroughly. That applies in client projects as much as in our own agency work. The question is mandatory, the answer is open. Often it is "yes, with review". Sometimes it is "no": in contract negotiations, in personnel decisions, in everything where trust is the product. That is different from the AI actionism we currently see in many companies. A chatbot gets introduced because the competitor has one: no task, no metric, no owner. "AI first" is a test question before every task, not a procurement programme. Our own editorial routine shows what the test question looks like on a small scale. Research summaries, first outlines, comparing variants: yes, with review. The decision on which topic we take on and which position we argue: no, that stays at the table. Both answers come from the same question, and that is exactly what makes them reliable. ## Human centered: a person remains the sender Three rules make the principle concrete. **1. AI produces drafts, people approve.** No text, no code, no design decision leaves our company without human review. We can show from our own daily work what that review finds. One draft for a technical article cited a study that does not exist; the source sounded plausible and did not survive a second look. Another draft missed the tone: factually correct, but bolder than the client would ever sound. And again and again we cut wording that carries legal risk, such as statements that could be read as guarantees. Published, each of these findings would have caused damage. The review is therefore not distrust of the tools. It is a promise to our clients: if you buy from us, you get human judgement, accelerated by AI. **2. A person is the sender.** This blog post carries my name, so I stand behind every statement in it. That holds regardless of which tools were involved in research and drafting. This rule has two practical sides. Externally, Google and AI search systems assess who is behind a piece of content; a named author with a traceable track record counts for more there than an anonymous editorial team (the keyword is E-E-A-T). Internally, the name enforces care: if a reader challenges a statement in this text, no machine answers, I do. We recommend the same rule to every client for their own communication. **3. Transparency about AI use.** We talk openly about where AI works at our company, and this blog is itself the example. Drafts are written with AI support, a documented editorial guide governs language and evidence, and the named author reviews and approves every version. We describe the tool chain behind it in [Inside arocom](https://www.arocom.de/en/blog/ai-in-agency-work). Transparency does not mean disclosing every keystroke. It means being able to say, when asked, exactly where AI was involved and who reviewed the result. We advise clients against hiding their AI use. Users and search engines recognise both, and credibility is the scarcer resource. ## What this does not mean: tasks we deliberately keep with people "AI first" is not AI maximalism. There are tasks we deliberately keep with people, even though models could deliver usable results there. Personnel decisions are one of them. A model may summarise documents, but at our company a decision about a person is made by a person who has sat across from them. The same goes for contract negotiations: there, the conversation is the product, not the document at the end. The final acceptance of client projects also stays manual, because that is where our name is on the line. And when something goes wrong in a project, no model writes the apology. It could deliver the draft. We do without it, because an apology only carries weight if the sender writes it and means it. For us, the line does not primarily run along what AI can do technically. It runs where judgement, trust or a relationship is at stake. When in doubt, we decide in favour of the person, even if it is slower. ## What this means for your projects In client projects, the principle translates into three questions. We put them at the start of every AI initiative: - **Who does it help?** An AI feature that improves nobody's work is decoration. We start with the people who work with the system every day (editorial, sales, support), not with the model. - **Who owns it?** Every AI process needs a human owner who checks results and develops the system further. "The AI did it" is not an acceptable answer to your customers. - **How do we measure it?** Hours saved, faster turnaround, fewer errors: defined before the start, reviewed after three months. Sounds unspectacular? That is intentional. Those who are future-proof are not the ones who deploy the most spectacular model, but the ones who build new tools into their work quickly, critically and permanently. The first step there: put the three questions above to your next AI idea.