AI-HOY, AInauts!

Welcome to a fresh issue of your favorite newsletter.

This week, Germany's chancellor shared a stage with Nvidia CEO Jensen Huang at an investment summit in Berlin.

Merz wanted to know how best to regulate AI. Huang's answer: "There's nothing to regulate if the industry doesn't thrive." 😉

In other words: build first, regulate later.

We'll leave the rest of that news to the major outlets, which have covered it plenty already. We'd rather focus on practical things that will, hopefully, help you.

Here's what's in this issue:

Let's get started!

80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.

Read the report

One thing Jensen said at the investment gathering did stick with us: Germany's data is one of the country's resources.

True. And your customer data is yours.

Last week, we looked at who else might be reading along when you talk to AI. Our closing advice: don't put genuinely sensitive information into a free chat. Or don't upload it at all.

The same follow-up came in through all our channels: okay, but where should it go instead? For law firms, tax advisers, medical practices and agencies, that is the question.

Language note: our anonymization tool and several provider interfaces shown in this section are in German. The screenshots illustrate the workflows described below.

Local LLMs and anonymizing data before it reaches the chatbot are two starting points. We've built a small anonymization tool for the latter.

We're not data-protection experts. We still want to revisit the issue and put the options in context. As always, this is our best understanding, without a guarantee. It isn't legal advice or blanket permission to upload customer data.

We came across two German providers that approach the problem very differently. They bring the underlying question into focus: do you have to trust the provider, or does the technology reduce that need for trust?

Logicc, based in Hamburg, relies on contractual safeguards.

It's a polished team workspace with GPT, Claude, Gemini and Mistral. Data is stored at Hetzner in Germany, and the provider offers a data processing agreement (DPA) and a professional-secrecy agreement. Check the terms for your chosen plan. The plans we compared cost €29.90 to €89.90 per user per month.

But Logicc and the Microsoft and Google infrastructure running the models still process the content in readable form during inference. Storage and transport encryption don't change that distinction. You're trusting the participants to honor their contracts.

Privatemode, based in Bochum, takes a technical approach.

Your request is encrypted on your device and decrypted only inside isolated hardware at the data center. That's called confidential computing. The design is intended to prevent even the service provider from reading it.

The trade-off: it uses open models such as GLM-5.3. In the comparison in our German issue, that scored just below GPT-5.6 Sol but clearly behind Opus 5.5: 45 versus 58 points on the Artificial Analysis Intelligence Index.

For the full protection described here, use the desktop app rather than the browser. You can try it free; usage after that is typically a few euros a month.

Language note: the linked Langdock security page and news report are in German. Another option: Langdock is a more mature choice in the same contractual category, with its own certifications. The company is moving to Europe over CLOUD Act concerns.

New tools call for new processes and workflows. Here's a simplified ladder, without a guarantee, based on the type of data and your requirements. This discussion uses the German/EU business context; local laws and professional obligations may differ elsewhere.

For lawyers, tax advisers and doctors in Germany, a DPA alone isn't enough. Language note: the following statute is in German. Ask specifically about the provider's confidentiality undertaking under Section 203 of the German Criminal Code and have the setup assessed for your obligations.

Again, none of this is legal advice. Ask your provider these questions directly.

A free extra layer: if our anonymization tool isn't official enough for your needs, look at Privacy Guardrail from DFKI. The browser extension replaces names, IBANs and similar details with placeholders before sending a prompt, then restores them in the answer.

It works with ChatGPT, Claude and Gemini. Still a beta, with no guarantee, but a useful second layer.

P.S. An urgent action on a related issue: two critical PHP vulnerabilities surfaced this week. They can allow unauthenticated attackers to execute code on vulnerable web servers and access data, including customer data. PHP-based sites, including WordPress, Joomla and Magento installations, may be affected.

Attacks are already being reported, and EU agencies are calling for immediate patching. Run a website? Check now whether its PHP installation has been patched.

With access to your email and browser, an agent such as Claude or ChatGPT can help identify other possible exposure. We'll go deeper next week. Here's a starting prompt:

Answering a question takes the right article. Doing the work (a refund, a plan change) takes permission. Running Agents in Customer Work, four weekly webinars. Register for All Four

Advertisement: in partnership with HubSpot. Language note: the webinar promoted in this section is in German and aimed at the German-speaking market.

We've been writing a lot about AI, costs and getting more from your limits and your existing tools. From fellow business owners, we're hearing the same thing: companies are watching their AI bills more closely.

That includes the big ones. Last week, 404 Media reported on an internal Accenture memo: from October 1, the firm centrally covers up to $29 per person per month for Claude.

Anything beyond that has to be charged to a project or team budget. Some employees had previously spent as much as $1,000 a month.

Accenture later clarified that this isn't a hard cap. But it changes the question: if you want more, you need a project to pay for it, and someone who can explain the return.

We find that interesting because few companies, if any, make as much money from AI transformation consulting as Accenture.

They're not alone. Uber had used up its 2026 AI budget by April and has since capped coding tools at $1,500 per employee per month.

Big or small, we're seeing companies tighten the reins.

HubSpot surveyed more than 2,800 marketing leaders worldwide on the subject. One finding: 83% are under pressure to prove AI's ROI.

Back in August, we asked: if everyone works faster with AI, why isn't it showing up in the numbers?

Research such as the METR study illustrates the problem. Experienced developers completed real tasks, sometimes with AI and sometimes without it.

With AI, they were 19% slower. They felt 20% faster.

That study was from early 2025. METR has since measured again: by late 2025, developers appeared to be faster, by between 4% and 18%.

But even the researchers say measuring is getting harder. Many participants no longer wanted to do certain tasks without AI. And people running multiple agents work on something else while waiting.

We see its impact on our business every day. We also see that not every AI user gets the same result.

Here are a few ways to demonstrate the effect, whether you're evaluating your own team or being asked to justify the expense.

Many companies don't know how much effort, money and time their processes consume. Without a before, there is no after.

Pick a recurring weekly task, such as preparing a quote, reporting or writing a blog post. Record how long it takes. Then use AI for it and measure again.

HubSpot found that more than half of marketing leaders still spend substantial time checking AI output before publication.

That's sensible (remember the slop grenades). It's also working time. Measure generation alone, and you'll flatter your savings.

Count waiting time too. We recently had a lively discussion about that in one of our communities. We'll return to the topic soon.

Two weeks ago, we showed that price per token isn't price per task. The same applies one level up.

"We spend €400 a month on AI" tells you little. "A quote now costs us €3 and 20 minutes instead of two hours and €500" tells you much more.

Divide AI costs by what actually comes out at the end.

Are you answering more support tickets? Is customer satisfaction rising? Are you producing more creatives while customer acquisition costs fall? Publishing more content and growing organic reach?

Look at the KPIs you were already tracking. Are they moving in the right direction?

Every approach above involves measurement. The old saying still applies: what gets measured gets managed.

It matters. We see plenty of businesses where people spend all day adjusting their Claude setup, building processes and producing attractive AI slop while costs rise and business metrics stay put. Eventually, that becomes a problem.

It's always useful to see what others do differently.

On Wednesday, November 4, HubSpot will share what its 2,800+ survey respondents do differently in a free live masterclass.

Language note: the masterclass, registration page and agenda below are in German. The event serves the German-speaking market.

On the agenda:

Wednesday, November 4, 2026, 10–11 a.m. CET (Germany). For marketing leaders and decision-makers; no prior experience needed. Everyone who registers gets the recording if they can't attend live.

If you need to explain what your AI spend delivers, or expect to soon, take a look.

Let's finish with a recurring question. It feels as if we write about agents moving from laptops to phones every week. Most recently, on Monday.

The questions keep coming anyway. In yesterday's community live Q&A, it came up again: what happens when I close my laptop? And where do my files need to be?

The shift from working locally with an AI agent to increasingly autonomous cloud and mobile workflows is well under way. Providers currently take different approaches, with different capabilities.

Dots, Grok Bot and OpenClaw run on their own machines. They can work around the clock, have a browser and sign in to your tools. Some now even have their own email addresses for creating logins.

A bit like a human employee.

Visual note: some interfaces in the screenshots in this section are in German.

Grok Bot now starts at $30 a month; in September it was $300.

Here in the EU, Dots still requires Business Premium rather than a Pro subscription. And even Dots can access your own computer only while it's online and the ChatGPT app is open.

Scheduled tasks run in the cloud even when your laptop is closed. Since October 6, that also applies to newly created tasks on Claude's Pro plan.

They can access what you've connected through connectors: Gmail, your calendar, Drive or Notion. They can't see your hard drive.

When your computer stays on, you can control it remotely from your phone or another device. An office Mac mini is one example. You get full access, but you're responsible for updates, logins and power.

What your agent can do while you're away depends on where the data lives, not just which model you're using.

Much of this new way of working comes from software development. We have an always-on Mac mini, but we're moving more work into GitHub for cleaner management and version control, especially when several people and their agents work with the same data.

If you've never tried this, here's a place to start. Your email is probably already in the cloud, which makes it a useful first workflow.

Copy this prompt:

The agent that works for you at night already exists in every major tool. Depending on the task, the setup can be straightforward or take more work.

If you want to see where this may be heading, try Grok Bot. It's also the one we're most cautious about when sensitive data is involved.

The real work is deciding which data may go into the cloud and what agents may do with it there. At the start, less is more.

That's it! Today got a little nerdy. We'll return to lighter fare soon.

Thanks for reading.

Reto & Fabian
Your AInauten