HafenPixel 創辦人分享如何在自家業務中使用 AI 智能體

Anyone who advises companies on AI should use it themselves. Not for a demo, but day to day, with real data and real consequences.

I run HafenPixel, my web development business in Hamburg, with my own back office in which AI agents do the groundwork. This article covers what they take on, what I still decide myself – and the rules I have set for it.

My five rules at a glance

Nothing goes out without approval – internally the AI works freely, externally I decide

No figure without a source

Client data only when it is cleared – the AI does not read my inbox

Every project has a budget and a clear criterion

The AI may disagree – and so may I

What the AI takes on

The agents work for me, not for themselves. Specifically:

Analysis: bringing together search data, visitor numbers and keyword data and saying what matters. What used to take half a day of spreadsheets is now an overview within minutes.
Research: competitors, trade media, legal changes – with sources, so I can check.
Drafts: texts, posts, page concepts and, when needed, emails. A draft is always a proposal for me, never a result.
Checking my own visibility: regular tests of how AI assistants such as ChatGPT or Gemini talk about my business. This has become one of my most important channels for enquiries.
Tasks and reminders: whatever follows from an analysis lands in the system as a task with a date – not in my head.

What I decide myself

The line is simple: anything that goes outside or costs money, I decide. This is how the work is split:

Task	The AI	Me
Analysis	brings data together, highlights what stands out	decide what follows from it
Research	collects, with sources	check the sources
Texts and pages	writes drafts	cut, change, approve
Marketing	plans content, prepares posts, measures the effect	decide topics and tone, approve every post
Emails	writes a draft when needed	read my inbox myself and approve every email
Client relationships	–	conversations, assessments, bad news

And there is one thing no table captures: how it sounds. Whether a sentence sounds like me, I notice when I read it. No AI can do that for me.

An everyday example: marketing

This shows most clearly in marketing, the area that is most likely to slip alongside client projects. The agents take on much of the groundwork – and I decide what of it goes outside.

The cycle behind my marketing looks like this:

Measure – search data, visitors, enquiries by source
Plan – actions with reasons and dates
Draft – pages, posts, research
Approve – I decide what goes out (the only step that always stays with me)
Measure again – does it work, or not?

In practice:

Checking visibility: the agents analyse search data and compare it with my plan. Whatever does not hold up is dropped instead of me continuing to optimise in the wrong place – rule 5 has an example.
Keeping an eye on AI assistants: regular tests check what ChatGPT or Gemini answer when someone asks for a provider like me. Where the answers have gaps, that becomes a task for the website. Because around a third of my enquiries come this way, this is now as important as Google.
Source of every enquiry: every enquiry is recorded with its source. That shows me which channels actually lead to projects – and which only bring clicks.
Research for guest articles: the agents find suitable trade media, with sources. Which ones I contact and what I write there is my decision.
Content on request: texts are only created when I order them – rule 4 below shows how that is set up.

What I deliberately do not automate in marketing: nothing is posted automatically, and no editor or contact receives an automatic email. Which topics I cover and how I sound stays with me.

My five rules

1. Nothing goes out without approval

The line runs between inside and outside. Internally the agents may work freely: analyse, research, write drafts, create tasks. Anything that goes outside – an email, a post, a change to my website – only goes out after my approval. That costs a few seconds per item and prevents exactly the mistakes you cannot take back.

2. No figure without a source

Language models sound confident even when they are guessing. So the rule is: every figure I use has to go back to a source – my own data, a study, a document.

An example from working on my website: I wanted to write that "more than a third" of my enquiries come via AI assistants. The analysis of my lead tracking showed just under a third. So the page now says "around a third".

3. Client data only when it is cleared

What a client entrusts to me does not end up in a text, a reference or an analysis without asking. Some projects I mention only anonymously, some not at all. The agents know these rules and stick to them – and I still check.

What the AI never gets to see: the AI does not read my inbox. It only writes an email draft when I need one – with the information I give it for that.

4. Every project has a budget and a clear criterion

AI costs real money per request. Every project in my back office therefore has a fixed monthly budget and a success criterion I measure it against. This is what it looks like for a real example:

Project: Texts on request
Success criterion	Every agreed text is delivered as a checked draft – no texts nobody ordered.
Budget	fixed limit per month
Pace	only on request, not continuously
Result	draft for approval, never published directly

The criterion stops an agent from busily producing things nobody needs.

5. The AI may disagree – and so may I

A good agent says when an assumption is wrong. When reviewing my own plan for more visibility, the analysis showed that two of my assumptions were wrong. In turn, I reject suggestions that look good on paper but do not fit me. Both are normal and intended.

What works well – and what does not

Works well: anything that needs a lot of material and little judgement – bringing data together, writing drafts, research with sources, preparing routine tasks. This saves me hours every week.

Works less well: anything that needs a feel for things – whether a reply suits a client, whether a text sounds too smooth, whether an idea fits the brand. The AI then gives a plausible but interchangeable answer.

Surprisingly important: good foundations. The more clearly my own rules, positioning and wording are documented, the better the drafts get. Most of the work was not in the technology, but in writing down how I actually work.

What this has to do with the AI literacy obligation

Under the European AI Act, companies that use AI have to support the AI literacy of their staff. The five rules above are essentially exactly that: clear agreements on what AI is used for, where its limits are and who decides what. In a team of several people, that includes everyone knowing and practising these rules. I have written a separate article on what the obligation actually requires: AI literacy under Article 4 AI Act.

FAQ

Does AI replace employees in your business?
No. HafenPixel is a one-person business, and AI does not change that. It takes on groundwork so there is more time for the actual work for clients.

Do clients get texts written by AI?
Clients get results I take responsibility for. Drafts may be created with AI, but every text and every piece of code only goes out after I have checked it.

Does the AI read your emails?
No. The AI does not read my inbox. It only writes an email draft when I need one, and every email goes out only after my approval.