Hi all,
What a week! Astra, Meta’s Muse launch and privacy scandal, Claude’s Science…
It’s Tuesday, and my feeds had told me (several times) that I'm already behind.
Thank God, Your Money, Your AI will never be one more reason to have FOMO. We are here to decide calmly what's worth our time, and what we can safely ignore.
The photo above is me in 2018, at my MBA graduation 👩🎓 in Manchester. Champagne in hand, very sure I had finally learned how business works.
Eight years later, a lot of it is obsolete.
Same with our knowledge about AI, right?
One of you wrote to me last week:
People seem to think AI can do everything, but what are the actual gaps that stop people getting real value from it? n8n was the big thing, and within months it got overtaken.
How do we keep up with AI, and what do we spend time learning?
And that is a great question - THANK YOU for it!
Today’s edition is about my approach.
My take: treat learning as a portfolio
Everything you learn about AI depreciates. Anyway. But! The pieces just depreciate at very different speeds 🚴🏻♀️.
The tool everyone is talking about this month is the asset most likely to go to zero. The more obvious its value, the faster the big platforms absorb it into their own products. Founders even have a word for it: getting "sherlocked."
The skills that hold their value are the unglamorous ones. Knowing what a good money decision looks like for you. Briefing AI properly. Catching a confident wrong answer.
That’s the framework I have in my mind though:
Here are the 4 layers, from the fastest to expire to the one that lasts.
Layer 1: Tools. The specific product and the specific setup. Exciting, useful, - yet the fastest to expire 😢. Examples:
Jasper, the AI copywriting app, was valued at $1.5 billion in 2022. Then ChatGPT launched and did much of the same thing inside a chat window. Within a year Jasper had cut its revenue forecast, its internal valuation and its team, and changed CEO.
OpenAI's Operator, an agent that could use a web browser for you, launched in January 2025. It was shut down that August and folded into ChatGPT itself. Even the platforms' own tools expire.
Closer to home: an automation that stitches your bank export into a spreadsheet.
In 2026 tools last for months. Does it make sense spending time on mastering them? My answer: only if you use this specific tool daily and only if it’s good enough as a product. Full stop.
It gets more interesting though 👇
Layer 2: Concepts. What AI can reach, remember and act on.
The brands change. But the concepts and ideas carry across products, and you learn them once. Some examples here for you:
AI that uses a browser for you. Operator died. The concept didn't: it lives on inside ChatGPT, in Claude's Chrome extension and in Perplexity's Comet browser. You understand it once, and you'll recognise it in every new product that launches with it.
Connecting AI to your own data across tools. In November 2024 Anthropic published an open standard for this, the Model Context Protocol (MCP). OpenAI adopted it the following March, Google DeepMind a month later.
So what? ➡️ You don't need to know the protocol. What you need to know is that "connect your files" now exists almost everywhere, and MCP is one of the most important concepts in the industry. This will stay and is worth learning.
Memory. ChatGPT, Claude and Gemini all remember you across conversations now. Same concept, three brands. (Useful, but with a catch. More on that below 👇)
Every tool is a delivery method for a concept underneath.
When the tool (as Layer 1) disappears or gets replaced, the concept (Layer 2) is what you keep.
Layer 3: Delegation. This is about HOW you work with AI: what you hand over, how you set it up your tools (Layer 1) and apply your concepts (Layer 2), how its work gets checked.
Your approach to delegation lives for years, and this is what I am willing to spend more time on. Examples:
Hand over a whole task, but keep the decision. That’s fairly straightforward, especially for high-stakes tasks.
Brief any tool and agent like a new colleague. Your country, your planning horizon, your goal, what "good" looks like, and what it should never assume. Good news is that you can re-use the same brief for many tools once you come up with one 💪. That is what .md files are for.
Learn how to build a safety net. When one agent does the work for you, apply a second agent with a different brief to check it. Banks have done this for decades and call it maker-checker, or the four-eyes principle: the person who prepares a payment never approves it. Anthropic's own guide to building agents describes the same pattern, one model generating and another evaluating.
💶 For your money: agent one builds a comparison of three ETFs' total costs. Agent two is told, you are an auditor, recompute every fee from the source documents and list every mismatch. If you can, run the checker on a different AI model, so the two don't share the same blind spots.
Ask twice, look where they disagree. Put the same money question to two different AIs, or to the same one three times. Where the answers agree, you can relax a little. Where they differ is exactly where you would need to dig deeper.
Split the job. For instance, for your tax paperwork, one agent collects receipts from your inbox, one sorts them by category, one checks totals against your bank statements, and you review the summary. Three narrow jobs are easier to check than one big one.
Build in the stop sign. Anything irreversible pauses and asks you first. Every time, with no exceptions.
In real life:
Klarna announced layoffs of 700 customer support agents back in 2024 as its AI assistant was ‘extremely efficient’. By May 2025, its CEO was hiring people back. He admitted that delegating was the right call - but delegating everything, fast, wasn't.
Here’s how how I think of 4 layers applied to 3 popular AI use cases: creating visuals, scheduling or automating workflows, and building a personal agent.
Finally, Layer 4: Judgment. Shaping your own knowledge to tell a good answer from a believable but mediocre one. AI makes this layer more valuable, because it produces believable answers at scale. Holds value for decades. Worth spending your time here!
Setting the rules before you work with any tool or build one. AI then works inside your rules instead of inventing or assuming them.
Designing the check. A net of agents is only as good as what you told the checker to look for. Deciding that it must recompute the fees, not just re-read them, is judgment.
Store facts, not opinions - leave opinions to yourself. Researchers at MIT and Penn State found that when a model stores a profile of you in memory, it becomes more likely to agree with you.
🫶🏻 One of my favourite approaches - running a pre-mortem. Before a big decision: it's 3 years from now and this went badly. Why? AI is excellent at answering this. It will rarely ask it for you.
Asking the better question. Instead of "which ETF is best?" your judgment should be knowing the specific question that matters: what will this cost me over ten years, all fees included?
Knowing where the rule came from. For instance, your AI builds a retirement plan on the "4% rule." Judgment is knowing that rule comes from US market history, and asking whether it fits someone who lives in Spain and earns in EUR.
In real life:
OpenAI rolled back a ChatGPT update in April 2025 after it became, in the company's own words, "overly flattering or agreeable." It had been tuned too heavily on users' thumbs-up. Judgment is remembering that the tool is built, in part, to please you.
The same pattern, applied to the job market
Earlier this year, Yoni Rechtman, a partner at the venture firm Slow Ventures, published a piece called "There will only be 4 jobs." (the picture above comes from Yoni, not me btw 😜)
His argument is: once AI can produce code, designs and plans on demand, job titles built around what you produce stop making sense. Teams will organise around how people work instead.
Here are the four, and where each one sits in our stack.
1. The fast builder (he calls them slop cannons; his words, not mine again 😜)
What they do: ship quickly with whatever tools exist, with a commercial instinct for what's worth building.
Where they live: at the top layers, in tools and concepts. They pick up a new tool the week they need it and drop it without regret.
What actually makes them valuable: knowing what's worth building. Their tools change every quarter. That instinct of theirs does not.
2. The infrastructure person
What they do: keep things stable and secure, stitch everyone's work together, clean up behind the fast movers.
Where they live: Layer 3, delegation. They are the checker agent in human form, the second pair of eyes, the person who notices what broke.
What actually makes them valuable: verification. It gets more valuable as more work is produced by machines.
3. The adult
What they do: govern, say no, and stop the company accelerating off a cliff. Think legal, finance, the board.
Where they live: judgment, the bottom and most important layer.
What actually makes them valuable: risk assessment and seeing opportunities. No model release changes that.
4. The hot people
What they do: the human interface or the people’s persons. They hold teams together inside and make customers, partners and hires want in.
Where they live: actually, outside the stack. Why? Because the stack is rational 🤖- and these people are on the emotional side! 💃
What actually makes them valuable: trust and realtionahips between people, which AI can support but can't replace.
What’s in it for us: none of the four jobs is defined by a tool.
Even the most tool-heavy archetype treats tools as disposable.
The other 3 are defined almost entirely by delegation, judgment and trust, the layers with the longest life.
It also gives you a practical shortcut. Decide which archetype you are, or want to be, and you know where your learning hours should go.
A builder can afford to learn tools fast, because she knows she'll drop them.
An adult should spend almost nothing on tools and a lot on the rules she sets before anyone, human or AI, starts working.
And last but not least - your call to action
I am a very result-oriented person myself. So let’s make our time (mine - spent on writing this, and yours - spent on reading as far 🤓) worth it. Six practical steps to make it work:
Audit. List everything you learned about AI this week or month. Label each item with its layer. Most people will probably find almost everything in layer 1. Re-assess. Re-focus your attention. Be disciplined about this. Learning tools is easier than shaping judgement. But easier does not mean durable, as we saw above.
Write your money brief. One page: country of tax residence, planning horizon, account types, goals, your investor profile this year, and your rules. No account numbers. It will then work for you in every tool that exists or will exist.
Pick one recurring money chore and write the workflow. Steps, inputs, the approval gates (e.g. I personally have a 20 EUR approval threshold. If we need to re-order toilet paper on Amazon that is 5 EUR worth, I am fine with this happenning without my attention). Only then choose a tool. If the tool dies, your workflow stays with you.
Build one read-only agent within your GPT or Claude, then give it a checker. The first one reads and reports. The second one audits its work against the source documents. Neither of them acts.
Ask for the other side. Whenever AI agrees with you on any decision, ask it for the strongest case against. Then run the pre-mortem.
Run the absorption test before learning any new tool. Ask yourself 4 questions, the same ones I'd ask of any investment:
Entry threshold: how many hours to learn it? Is it worth it?
How easy to exit: if it disappears, does my work survive?
What can go wrong: what data does it touch? Is it worth sharing?
The investor question: is this a product, or a feature of the next model release?
And as always - hit reply and tell me what happened when you applied it! I am genuinely interested in what was the most useful for you in today’s edition, and what other questions you have.
Raising a glass of champagne 🥂 to all of your AI victories,
Eugenia
Previous issues available on the website 👇🏼
Did you like this philosophy edition today? 👩🎓
Then I'd love it if you would forward it to your network. I am sure it will save the MBA years (and fees!) for some of your friends or their kids.