On this page
Almost everyone is asking the wrong question about AI.
The usual question is: "Is AI good enough to do this yet?" Good enough to write our content? Good enough to answer customers? Good enough to invest for me?
This week, a bank in Germany answered that question in a way almost no outlet read correctly. And its answer had nothing to do with how good the AI is. It was about what the bank decided the AI is allowed to touch — and what it must never touch.
That's the lesson worth paying for.
August 25: a European bank opens the door to machines
On August 25, 2026, Scalable Capital — the Munich-based online broker — announced "Agentic Investing", billing itself as the first bank in Europe to open its platform to all the major AI assistants: OpenAI's ChatGPT, Anthropic's Claude, and SpaceXAI's Grok.
This isn't a small pilot. According to the official announcement, Scalable has more than one million customers and over €60 billion in assets under management.
Customers can now connect their brokerage account to whichever AI assistant they already use and give instructions in plain language: place buy and sell orders, set up recurring savings plans, manage watchlists, set price alerts, look up stocks, ETFs and warrants, read news, pull real-time and historical prices, and run portfolio analysis (diversification checks, scenario analysis, sector and region breakdowns, risk assessment).
Co-founder and co-CEO Erik Podzuweit called it "the biggest technological shift in finance since online banking."
That's the part everyone read. Now for the part almost nobody noticed.
The 76% figure is being read backwards
Fortune covered the launch on August 26 with a very loud headline: in one test, Claude beat human traders 76% of the time.
The number is real. But it comes from a different study, run at a different time, and that study reached almost the opposite conclusion from the one the headline implies.
On June 17, 2026, Jerry Bell, Victor Haghani and James White of Elm Wealth published "Do AIs Make Good Traders, and Do They Make Good Traders Better?". They handed Claude, ChatGPT, Gemini and Grok the next day's front pages of the Wall Street Journal — news known in advance, before the market had reacted — and had the models trade the S&P 500 and 30-year Treasuries, with leverage, starting from $1 million.
The headline result is genuinely impressive: Claude beat the human players in 76% of rounds, ChatGPT in 63%. Claude's average ending wealth was $2.59 million.
Keep reading the table, though, and the picture changes color:
| Model | Average ending wealth (from $1M) |
|---|---|
| Claude | $2.59M |
| ChatGPT | $1.47M |
| Grok | $0.97M |
| Gemini | $0.49M |
Gemini lost more than half the money. Grok lost a little. Two of the four models — handed tomorrow's news today — still lost money. That is not a portrait of a profession that has been automated. It's a portrait of a deeply lopsided capability.
AI knows what to pick. It doesn't know how much to bet.
The most important part of the Elm report isn't the win rate. It's the reason behind it.
All four models, the researchers found, "took excessive risk" relative to the goals they set for themselves. They went in with 7–12x leverage on equities, producing daily swings of 20–40%. The authors put it bluntly: those bet sizes "could not be justified by any reasonable expectation."
Here's the paradox: the models understand the Kelly criterion for sizing bets by probability — ask them about the theory and they answer fluently — but they can't apply it to the very game they're playing. They know the rule and don't act on it.
In short: AI reads the news very well. It calls direction reasonably well. It has no idea how much money to put behind a conviction.
Elm itself states the caveat plainly: "The performance of any AI model in a simulated, hypothetical exercise does not reflect its ability to generate profits in real markets."
And this is where it touches your business — even if you never trade a share.
The same pattern shows up in every business use of AI: AI is strong at judging direction (what to write, how to reply, what this customer seems to care about) and weak at sizing consequences (how much to spend, what to commit to, what we lose if it's wrong). Give it the first job. Don't give it the second.
What Scalable actually built wasn't AI — it was a door for machines
This is the point I think was missed most in the whole story.
Scalable Capital didn't train a model. It didn't build an "investing AI". What it spent months building was infrastructure that lets other people's machines knock on its door: an MCP (Model Context Protocol) server for cloud-hosted AI assistants, plus a CLI app that installs directly on the user's machine.
MCP is an open standard originally developed by Anthropic and now supported by most of the major AI assistants. It's how a service describes itself to an AI: here's what you can do with me, here's the data you may read, here's how to call it.
Sit with that for a moment.
For twenty years, every business has built exactly one door: a door for people. Websites for human eyes. Buttons for human hands. Forms for humans to fill in. The entire web design industry — my studio included — lives off that door.
What Scalable just did is open a second door: a door for machines. Not so machines can replace people, but because its customers have started living inside a chat box, and the bank doesn't want them to have to leave that box to buy a stock.
It's the same force pushing SEO toward AEO and GEO. You used to optimize so Google understood your page. Now you optimize so a machine can read and answer on the customer's behalf. The next step — the one Scalable just took — is letting that machine act inside your systems.
Whoever builds the door first shows up in the AI assistant's answer. Whoever doesn't, the customer still has to open a browser — and more and more of them won't.
The right line isn't "smart or dumb" — it's "reversible or not"
Now the best part, and the part I'd like you to copy into your notebook.
Scalable lets the AI do a lot: read the whole portfolio, analyze risk, build scenarios, draft orders, propose savings plans. But according to its own announcement, there are things the AI is not allowed to do:
- Every trade and savings plan must be approved by the user before it executes.
- No deposits and no withdrawals through the AI interface — those happen only on the web or in the app.
- Mandatory two-factor authentication, at setup and repeated periodically.
- All activity is monitored in real time, with push notifications and trade confirmations as usual.
- The AI's access can be switched off at any time.
Scalable also says it worked with the regulator BaFin while building this, and still provides pre-trade cost information and the Key Information Document (KID) before each securities trade — in line with current rules.
Look closely at that list. The line they drew isn't based on how smart the AI is. It's based on a completely different criterion:
If a mistake can be undone, the AI does it. If one mistake means real money lost that can't be recovered, a human presses the button.
Misreading a chart — fixable. Suggesting the wrong portfolio — ignorable. But a filled order at 12x leverage has no undo button. And that exact skill — "how much is a sensible amount to bet" — is, as the Elm report shows, precisely where AI is weakest.
Scalable doesn't need to know how good Claude is. It only needs to know which consequences are irreversible, and to put a human right in front of that door.
Chief Product Officer Alexander Seipp told Fortune something refreshingly honest: "Whether you'll find the holy grail of high returns at low risk together with your AI assistant, I think that still remains to be seen." A bank opening its doors to AI while its product chief says "remains to be seen" isn't a lack of confidence. That's the right architecture.
What this has to do with your business
You're not a bank. You sell furniture, run a language school, operate a clinic, own a print shop. So what's the takeaway?
One: change the question. Stop asking "is AI good enough to do X yet?" Ask instead: "If the AI gets X wrong, how long does it take me to fix it?" Five minutes to fix? Hand it over. A lost customer, a lost contract, lost money? Put a human in the middle. It's a filter you can use today, on every process, and it doesn't require you to understand anything about language models.
Two: let AI propose, let people size it. Let AI draft the quote — a person approves the number. Let AI write the content — a person decides to publish. Let AI triage orders — a person handles the exceptions. That's exactly the shape a bank chose after months of working with its regulator. You don't need to be smarter than they are.
Three — and this is the long game — start thinking about the second door. Can your website today tell a machine what you sell, what it costs, how delivery works and when you're open — in a way an AI assistant can understand and answer for you? For most businesses, the honest answer is not yet. Not because it's hard, but because nobody has asked the question.
You don't need to rush into building an MCP server. But structured data, content that answers questions directly, and consistent information across your website and the platforms you're listed on — those are the first bricks of that door, and they're far cheaper than rebuilding later.
Scalable Capital just showed the whole industry where this is heading. The distance between where you stand and that destination isn't measured in technology. It's measured in whether you start early or late.
If you want to know how well your current website "talks" to machines, send me a quick note — it's the cheapest test you can run this week.
Sources
- Scalable Capital becomes first bank in Europe to open platform to AI assistants via Agentic Investing — official announcement, August 25, 2026.
- A major German bank just let Claude and ChatGPT trade for customers. In one test, Claude beat human traders 76% of the time—but there's a catch — Fortune, August 26, 2026.
- Do AIs Make Good Traders, and Do They Make Good Traders Better? — Jerry Bell, Victor Haghani, James White (Elm Wealth), June 17, 2026.
- Scalable Capital Enables Agentic Investing, Opening Platform to ChatGPT, Claude, and Grok — Finovate, August 25, 2026.
