Introduction
Welcome to another episode of TJ's Technology Tuesday — and today I have deliberately placed a guardrails sign on it, because when it comes to AI the topic is above all about guardrails: where should AI ideally have its guardrails set?
Two Guardrails for Artificial Intelligence
What are the typical guardrails AI should ideally have? AI should ideally have guardrails around data privacy — and data security as well. In a corporate context I always picture the middle lane on the motorway: which tools do we use, in what way, and what procedures have we put in place for that in the company?
Claude — An Ethically Responsible AI
This topic has become very topical right now. I don't know whether you caught this in today's Handelsblatt, for example — there is this gentleman, perhaps not everyone knows him yet, who used to work at OpenAI, the company that founded ChatGPT. He now runs his own company, and that company is Claude — or rather, the service is called Claude and it belongs to the company Anthropic. And dear Dario Amodei has just had quite a dispute with the Department of Defense. Here is the latest headline: Discussion about the Pentagon's use of AI. OpenAI backtracks. What is the background? The article already says: OpenAI's successor. Dario Amodei has a very clear policy on the topic of AI because he says: We want an ethically responsible AI. We want AI to have clear guardrails — to know what it must not do. And the Pentagon said: We would like to use your AI model, and Mr Amodei, the founder, said there are two decisive no-gos.
No-go number one: I do not want AI to be used to spy on US citizens on a massive scale. No-go number two: I do not want AI to be used in autonomous weapons systems in such a way that the killing of human beings is automated without a so-called human-in-the-loop — without a person who checks and decides whether that is acceptable. In practice this is naturally always somewhat difficult to judge when you look at it. And the topic of defence always raises difficult questions. The key issue is tempo. Take the Iron Dome in Israel, which is currently very heavily in use: when you have a very short rocket flight time, AI-supported systems may need to decide within fractions of a second whether to intercept an object or not. For me, that is still defensible in the context of a purely defensive system — where the question is how quickly something must be intercepted — and it makes sense to say: okay, there are areas where AI can be used. But as I mentioned, Dario Amodei made it crystal clear: No, we do not want that. And the Pentagon said: Then you are out. Interestingly, colleagues at OpenAI are now beginning to backtrack somewhat and saying perhaps it is not such a good idea after all.
Automation with Claude and OpenClaw
On the subject of Claude and OpenClaw — the names sometimes sound very similar. There was the discussion I covered last week, or the week before, about OpenClaw and the idea that everything gets automated. When it comes to automation it is always vital to say: keep the guardrails in place. And that applies not only in the social and Pentagon sphere, but in private life too.
If I look at Claude right now, it released another update this week. Within Claude there is the web interface and there is an application. Inside the application there is the Chat area — similar to what you know from ChatGPT — there is the Code area, which is the deepest level you can go to, all the way down to the terminal level, and there is Cowork, which acts as a wrapper, a surface for carrying out specific tasks.
Let me give a concrete example. Yesterday I received from a customer a folder containing more than 5,800 vehicle appraisal PDFs from the automotive sector. The idea was to extract the data into an Excel list — date, vehicle identification number, costs, value summary, prior damage, repainted panels — and to have all of that in one place. So Claude got to work and produced the whole thing for me.
Then came the next step — and it was quite fascinating. It creates an Excel table and runs through everything completely. When I then tried in a further iteration to say: give me even more fields, I ran into error messages saying: something went wrong, the prompt is too long. That can sometimes be a problem.
What I did next: I went one level deeper into Claude Code. And Claude Code is the mode where I said: I want many more fields here. You use this when you want to work with even greater power. The important thing when working with it: you grant it access to a folder and it works within that folder.
With OpenClaw, when working with this automatic agent, the risk is often that you say: I'll grant it access to my main directory, I'll grant access to everything. Not a good idea. You should always give it very limited access.
But here is an example. In several iterative steps it worked through things and told me — when I clicked "Show More" — a great deal of information about what it was doing. I always recommend that, even for those who say they are not coders: you should always look over the AI's shoulder and see what it is doing. The nice advantage is that in doing so you learn a little about how these systems work.
And then out came the file: 5,803 PDFs in a complete Excel file that I could even edit further. That means — when I open a new Excel table, for instance, if I look at a registration list — Claude even has a plugin. And with this plugin I can do what Microsoft promises at some point with its Copilot. Claude Cowork is already at the point today — please understand I am not opening the file here as it contains customer data — where you can say: I have something here, I had the mileage readings; extract just the kilometre figure from the mileage column. And there was an additional note: read out. Then I said: put just the number into another column. These are exactly the kinds of things that otherwise take a long time to do manually. You can do them brilliantly with this extension. Or when you need to go a little deeper: a tip I received from the US — Shortcut, also an excellent plugin, with which I can do even more things. I switch over to the browser — so here I have the Try Shortcut topic. And Try Shortcut AI is a solution where I either have an Excel plugin or where I can also work standalone in the browser. Those are automations, especially in the Excel area, that are genuinely outstanding.
So: Claude. There are those three areas you can use via the application. I have Chat, I have Cowork — which is essentially a surface where I can do things. Here I can say, for example: tidy up my Downloads folder for me. The idea is that when I create a new task, I say: in which folder do I want to work? And I specify only the Downloads folder, for instance. It then asks: may I do something there? And the important point: please do not always grant permission for everything — just for once, and only within that folder. And the idea is — something I have done myself — to push a PowerPoint file into a subfolder on my hard drive, a presentation of 197 slides, and say: please translate it into another language. These are exactly the kinds of things you can do wonderfully with this tool.
And on the subject of automation — OpenClaw as the big thing. The most important warning I can give you is: never install OpenClaw on your own production system. I would not necessarily install it on a separate Apple device either; instead I have installed it at Hostinger — that is, on a separate, let's say, online system in a so-called Docker container, which is an isolated environment. That way nothing can go wrong. I have several Google accounts, and for this I have taken a Google account on which I have no productive data. But I keep my productive Microsoft 365 mail as far away from OpenClaw as I possibly can.
Testing something like this is a great idea — absolutely. But there have already been several data leaks in this area, so you really need to configure security very, very carefully, including not using the default port — without getting too technical here.
Automation with Manus
On the topic of automation, Manus is also well suited. However, my experience is that Claude Code is now significantly better, or alternatively the Excel integration. But Manus now belongs to Meta — originally a company from Singapore — where I can also automate tasks.
Perplexity Computer
And something brand new, only since yesterday: Perplexity has released its Computer. With Computer, though, you need a Max account, which costs a mere 200 dollars a month. But here it does something similar to what you know from Manus: when I give it a task, it opens an extra terminal, a separate environment. A bit like having your own server, as I described earlier. Both Manus and Perplexity Computer work this way — they do not operate on your own machine but in isolated environments.
And that is one of the key points: if you take only one thing away from this, please remember above all to experiment with these services, to choose providers that are ethically well-formed. I have become, as I mentioned, a great fan of Claude — and not only since the Pentagon episode. Many thanks to dear Nils Bäumer, who drew my attention to that as well. It even passed me by to some extent last week. So: use ethically well-formed providers.
If you use OpenClaw, please always run it on a separate machine. And with Manus or Perplexity Computer, for example, you have the option of doing things within the browser in your own isolated environment. That is significantly safer than using Comet as a browser running on the local system. And it is above all significantly safer than OpenClaw, which in the worst case runs on your own machine with full access rights. That is definitely not something you want.
Conclusion
Yes, a slightly more technical episode today — but I think it is important. I see many people using these automation solutions that can be incredibly productive: reading 5,800 files and transferring them into an Excel table is super, super impressive. But please do this kind of thing in an isolated environment where you clearly state: the data stays only in these defined areas. Before tackling productive tasks, always set the key guardrails very clearly — just as the head of Anthropic set them against the Pentagon.
I hope that was not too deep, but perhaps gave you a little more background — always connecting the societal dimension right down to your own implementation: productive, but safe.
Your personal coach for executives, Thorsten Jekel.
Key Takeaways
- AI tools such as Claude should have clear guardrails — for data privacy, data security, and ethically responsible use.
- Dario Amodei, founder of Anthropic (Claude), rejected the Pentagon request because AI must not be used for mass surveillance of citizens or in autonomous weapons systems without human oversight.
- Claude offers three usage modes: Chat, Cowork (task-oriented surface), and Claude Code (terminal level for demanding automations).
- Using Claude Code, 5,803 vehicle appraisal PDFs were automatically read and transferred into a structured Excel table.
- The Claude Excel plugin enables direct data manipulation in spreadsheets — similar to what Microsoft promises with Copilot, but available today.
- OpenClaw (open AI agent) should never be installed on your own production system — a recommended approach is an isolated server (e.g. Docker at Hostinger) with a dedicated account set up specifically for it.
- Manus (now part of Meta) and Perplexity Computer are further automation services that work in isolated browser environments and are therefore significantly safer than a locally installed OpenClaw.
- Perplexity Computer requires a Max account (approx. 200 USD/month) and opens a dedicated terminal in an isolated environment for each task.
- Productive data — especially Microsoft 365 email — must be kept strictly away from automated agents such as OpenClaw.
- Core principle: AI automation can be extremely productive, but must always run in clearly defined areas with minimal, purposefully granted access rights.
Frequently Asked Questions
What are AI guardrails and why are they important?
Guardrails are boundaries that define what an AI may and may not do. They are important to ensure data privacy and data security and to guarantee that AI is used in an ethically responsible way — both in a corporate context and in private life.
Why did Anthropic founder Dario Amodei reject cooperation with the Pentagon?
Dario Amodei rejected the Pentagon request because he defined two clear no-gos: Claude must not be used for mass surveillance of US citizens, and must not be deployed in autonomous weapons systems where people are killed without a human-in-the-loop reviewing the decision.
What three modes does the Claude application offer?
The Claude application offers Chat (similar to ChatGPT), Cowork (a surface for concrete tasks such as organising folders or translating presentations), and Claude Code (the deepest level with terminal access for complex automations).
How did Claude Code help process more than 5,800 PDFs?
Claude Code automatically read 5,803 vehicle appraisal PDFs and transferred the relevant fields — such as date, vehicle identification number, costs, prior damage, and repainted panels — into a complete Excel table that could then be edited further.
What can the Claude Excel plugin do?
The Claude Excel plugin allows you to clean data directly inside an open Excel spreadsheet using natural language — for example, extracting just the kilometre figure from a column containing mixed text and writing it into a new column — without any tedious manual editing.
Why should OpenClaw not be installed on your own machine?
OpenClaw, as an automatic agent, has access to files and systems, and there have already been several data leaks in this area. Anyone who installs it on their own production system and grants broad access rights risks having sensitive data compromised.
How can you test OpenClaw safely?
OpenClaw should be run in an isolated environment — for example, in a Docker container on a separate online server (such as Hostinger). It is also advisable to use a dedicated Google account with no productive data and to change the default port.
What is the difference between Manus and Perplexity Computer?
Manus (originally from Singapore, now part of Meta) and Perplexity Computer are cloud-based automation services that execute tasks in isolated environments — without direct access to your own machine. Perplexity Computer requires a Max account at around 200 USD per month and opens a dedicated terminal for each task.
What is Try Shortcut AI and what is it best suited for?
Try Shortcut AI is a tool from the US that can be used both as an Excel plugin and as a standalone application in the browser. It is particularly well suited for automations in the Excel area and was mentioned as a recommended addition for spreadsheet tasks.
What is the fundamental principle for the safe use of AI automation in business?
The fundamental principle is: productive, but safe. AI automation should always take place in clearly defined areas, with minimal and purposefully granted access rights. Productive data — especially corporate email or sensitive customer data — must be kept strictly away from automated agents.
Tools & Resources Mentioned
- Claude (Anthropic) — AI assistant with Chat, Cowork and Code modes; known for ethically responsible guardrails
- Claude Code — Terminal-based automation level within the Claude application for complex file and data processing tasks
- Claude Cowork — Task-oriented surface in Claude for working with folders, translating presentations, and much more
- Claude Excel Plugin — Direct integration of Claude into Microsoft Excel for AI-powered data transformation
- Microsoft 365 / Copilot — Microsoft's promised AI assistant in Office; according to this episode not yet as capable as Claude Cowork is today
- Try Shortcut AI — Excel plugin and browser tool for AI-powered spreadsheet automation (tip from the US)
- OpenClaw — Open, automatic AI agent; recommended only in an isolated environment (Docker/separate server)
- Manus — Cloud-based automation service (now part of Meta); works in isolated environments
- Perplexity Computer — New cloud automation service from Perplexity with isolated terminal; requires Max account (approx. 200 USD/month)
- Hostinger — Hosting provider on which OpenClaw can be safely run in a Docker container
