Please don't AI-ify everything

The Sunday edition: a longer read for the second coffee, one fundamental topic every Sunday. Last week it was passwords, personal data and AI. Today it is a word we almost made our claim.

This week I sat in a branding session — more precisely: I was thinking about claims with my AI agent, the story has to be that honest. A word emerged that has been stuck in my head ever since: AI-ify. AI-ify everything. AI everything. It sounds good, it sticks, it would make a fine company claim.

We decided against it. And the reason has turned into a better text than the claim would ever have been — because «AI-ify everything» is exactly what is happening everywhere right now. And in that blanket form, it is the wrong goal.

The reflex

You cannot blame anyone for it. The tools are impressive, the promises bigger, and the pressure is real: every week someone at the pub or on a stage explains that their company now has «AI in everything». So the reflex kicks in: AI into the contact form, AI into the quotes, AI into the mails, AI into the minutes, AI into the books — and right now, before the competition does it.

I am currently rebuilding my entire digital life around an AI agent. It is hard to accuse me of being a sceptic. Which is exactly why I allow myself the objection: the reflex mistakes the tool for the goal. Nobody ever «digitalised» a company by buying software at random — and nobody makes a company future-proof by pouring AI into it at random.

Three filters instead of one reflex

What belongs in the hands of an AI and what does not can be sorted with three questions. They are unspectacular, and that is precisely why they work.

Filter one: what happens when it goes wrong — and does anyone notice? AI work is fast, but not flawless. So the decisive thing is not whether a mistake can happen, but whether it is noticed and can be undone. An agent writing code that must pass tests and a review before going live: uncritical — the mistake gets caught in the net. An agent sending unchecked quotes to customers: a different calibre. The rule: the harder a mistake is to take back, the more checking belongs between AI and effect. In my setup this rule is literal — every change runs through a documented, audit-proof process, and the final sign-off before anything goes out is a human.

Filter two: what data does the tool see along the way? That was last week’s Sunday edition in one sentence: define data classes, and decide per class which tools may see them and where they are processed. «AI-ify everything» implicitly means: show everything. Run the reflex through this filter and you quickly notice that a part of «everything» belongs on your own infrastructure only — or in no AI’s hands at all.

Filter three: is the process even ready? The most uncomfortable filter. An AI that takes over a chaotic process produces chaos at higher speed. If the knowledge lives in heads instead of documents, if every case is «a special case», if nobody can say what the rule actually is — then the next step is not AI, it is tidying up. It sounds like the more boring answer. It is the cheaper one.

AI-ready instead of AI-everything

Which means the counter-proposal is already on the table. Not «AI into everything», but in this order: document the processes. Choose systems that can be steered. Set the data rules. And then — where all three filters show green — let agents take over. If you proceed like that, you may well end up AI-ifying a great deal. But controlled, traceable, in the right order — and with the calm certainty of knowing at any time what the AI sees and does.

That, by the way, is the difference between two sentences that sound almost identical: «We use AI everywhere» and «We can use AI everywhere it pays off». The first is a reflex. The second is a capability. The capability is called AI-ready, and it is the actual goal.

What I deliberately do not AI-ify

So this does not stay abstract — a small, honest list from my own rebuild. Things that no AI touches in my setup, on purpose:

• The recovery codes of the most important accounts. They live on paper, outside every system. Not because I distrust the AI, but because a skeleton key should not sit in any attackable system — operable by nobody also means stealable by nobody.
• The final sign-off. No post goes online, no code onto the main branch, no mail to third parties without a human saying yes. The agent prepares, checks, documents — the responsibility stays here.
• The judgement over tone and stance. My agent writes drafts, and it writes good ones. But whether a text sounds like me, whether a phrase is honest or merely clever — no machine decides that. This week it learned that my texts contain no emojis. From me, not from itself.
• Relationships. No agent of mine answers customer mails to the end. It may pre-sort and compile facts. The conversation is mine.

You can see the pattern: it is not the «difficult» tasks that stay human — the agent can do difficult. It is the places where irreversibility, trust or responsibility are at stake.

The Sunday question

In case Monday already has an AI plan for your company, here is the question to take along: Which of your processes would profit most from AI — and would it pass all three filters: mistakes get noticed, the data may go there, the process is documented? If yes: AI-ify it, wholeheartedly. If no: the path to yes is the actual project — and it is shorter than it looks, walked in the right order.

And how does that work, technically? For anyone who wants the exact mechanics, the second filter has an appendix: the technical drawing — a single figure showing which data may enter which AI tool and where the line falls for the crown jewels. No computer science degree required, promised.

Questions, objections, a reflex of your own? Write to me — I answer personally. Next Sunday: the third edition, a more personal one.