Everyone moved to Claude. On gut feeling. With the company data along

For a while there was a real shift from GPT to Claude.

Which Claude model? Often that was not even mentioned. The reason? “Much better.” Subscriptions were cancelled and new ones taken, and mostly that happened on gut feeling.

The company data moved along with it. GDPR, anyone?

I am not going to pretend that was a stupid choice. The models genuinely got better, and it is entirely logical that people want to use the best one available.

The point is a different one. No choice was made. Somebody had a feeling, and the organisation followed.

The question is not which model is best

The question should no longer be which chatbot is best. It should be which model suits which job.

One is excellent at writing, another at reasoning, and a third at getting things done. That last category is where it got interesting. The models released this year are not so much faster at answering. They are better at finishing work:

• breaking tasks down and planning them
• understanding and changing code
• analysing files
• producing reports
• checking assumptions
• and above all: finishing things

In a mid-sized business the practical effect shows up straight away. Loose emails, notes and spreadsheets become structured output. A codebase gets understood instead of guessed at. Ticket analyses, management reports, finance checks, incident reports. It feels less like a chatbot and more like a model with judgement.

But you always have to experience this kind of thing for yourself. Much of what you read about it comes from benchmarks and press releases, and those measure something other than your work.

What did not happen

Back to that switch. Look at what did not happen before the company data moved.

Nobody established what exactly went into the prompts. Personal data? Customer data where somebody else is the controller? That changes everything about what has to follow.

Nobody looked at where the processing happens. On consumer plans that is rarely inside the EU, and on several of them the setting “use my conversations to improve models” is on by default.

No processing agreement was signed. And nobody agreed who carries the risk if something comes out that should not have.

That is four questions. Together they cost half an hour. They were not asked. Not because people are careless, but because the switch did not feel like a decision. You cancel one subscription and take another. You do the same with a streaming service.

Except: with a streaming service, no customer file walks in with you.

What I make of it

The everyday user does not want to know any of this. No reproach; it is simply how it is. Somebody who has a quote to write wants a tool that picks the right model underneath and otherwise keeps quiet.

Which is why a layer that does that for you is a logical choice. Not because one vendor has the best models, because that changes every quarter, but because the choice then gets made once at organisation level instead of two hundred times on gut feeling. The actual win is the processing agreement and the data boundary that come with it.

What you do have to do is make that one choice for real. Not name a vendor once and never look again, because then you have only moved the gut feeling to procurement.

I worked the four questions from this piece into something you can run through in your browser. It names no products, because that changes too fast to put in an instrument, but it gives you the requirements that follow from your situation, and it points out where your own answers contradict each other.

Run the Compliance Check →

A switch does not feel like a decision, but it is one. Either you made it or your gut did.