AI and Investments:
Oracle or Adversary?

An investment committee can process more information today than ever before. In a few years, an AI could have read every lease in a real estate portfolio, identified all termination rights, checked tenant creditworthiness, analysed comparable transactions and calculated ten thousand scenarios for interest rates, inflation, vacancy and exit yield – before the meeting even begins.

But does better information also solve the actual problem of investing? The history of recent years suggests a more complicated answer.

On 1 August 2012, Knight Capital deployed new software into its trading system. Then the machine started trading.

Within just 45 minutes, the system sent millions of erroneous orders into the market. Knight Capital lost more than 460 million US dollars.

Particularly remarkable: internal systems had already produced warnings. People received signals but did not recognise their significance in time.

Knight Capital was not modern artificial intelligence. But the case shows a problem that becomes even more important with more powerful systems:

Machines do not only automate correct decisions. They also automate mistakes. And they do both at a speed no human can match.

In 2017, the AI Powered Equity ETF AIEQ was launched. The idea was compelling: a machine can process company data, news and market information on a scale no human fund manager could handle.

More information, however, did not automatically mean more return. For the period from the end of 2017 to mid-2023, AIEQ reportedly achieved an annualised return of around 6.7 per cent, compared with around 11.2 per cent for a broad US equity market benchmark.

Other early attempts at AI-driven investment strategies also fell short of expectations. Sentient Technologies, for example, relied on evolutionary algorithms for investment decisions. The company had raised substantial venture funding. According to published analyses, its fund returned around four per cent in 2017 and practically nothing in 2018; it was then liquidated.

More recent research, however, paints a more nuanced picture. A current study of AI-based investment strategies shows that early adopters were in some cases able to achieve excess returns over comparable non-AI strategies. As the technology spread, this advantage declined.

That would be a very old investment story: an information edge can generate alpha. But once everyone has it, it is no longer an edge.

Real estate is where it gets particularly interesting. Zillow had enormous amounts of data and one of the best-known automated property valuation systems in the world.

With Zillow Offers, the company went one decisive step further. It no longer used its models only to value properties. Zillow started buying houses itself.

In November 2021, the programme was shut down. In the third quarter alone, Zillow had to write down around 304 million US dollars on its housing inventory. CEO Rich Barton explained that the unpredictability of future house prices had been far greater than expected.

The problem was more complex than a “bad algorithm”. Growth pressure, purchase prices, renovation costs, labour and supply chain problems and a changing housing market met models that had learned from the past.

Therein lies a crucial lesson: a model can be excellent on average and still be wrong exactly where the capital was deployed.

Capital does not care about a model’s average error. Capital cares about where the error occurs.

There is another side, too. The German asset manager LAIQON has been using artificial intelligence in its investment process for several years.

According to the company, the LAIQON AI Dynamic Multi Asset achieved a cumulative performance of 66.41 per cent from May 2020 to May 2026, 28.20 percentage points above the stated peer group. A more defensive AI fund also outperformed its peer group, according to the company.

These figures come from corporate communications and should be read accordingly. They do not prove that AI fundamentally makes better investment decisions. But they show something else: AI-supported investment processes can work.

The European Securities and Markets Authority ESMA is now also looking closely at the use of artificial intelligence in asset management. The number of European funds that actually use AI as an essential part of their investment strategy is still relatively small. The experiment has only just begun.

Blackstone takes a different approach. The investment manager describes using artificial intelligence, among other things, for screening investment opportunities and analysing large volumes of documents. AI helps to ask faster:

• Which deals deserve attention?
• Which documents contain anomalies?
• Which risks might have been overlooked?
• Which patterns might humans not see?

But the machine does not replace the investment committee. That is a fundamental difference. AI extends human judgement instead of pretending to make it redundant. And perhaps that is exactly where its greatest strength lies.

Behavioural finance has documented loss aversion, overconfidence, anchoring, herding, confirmation bias and the so-called disposition effect for decades.

One study of professional institutional investors is particularly remarkable. The fund managers studied managed portfolios averaging 573 million US dollars. Their buying decisions showed skill. Their selling decisions were a different story: they even performed worse than random selling strategies.

The researchers found evidence that attention and heuristics explain a significant part of this discrepancy. So the problem was not a lack of information. It was human behaviour.

The usual question is: “Can AI make better investment decisions than humans?” Perhaps another question is more interesting: “Can AI help people recognise their own bad investment decisions?”

Then the role of the machine changes. It goes from oracle to adversary.

An investment team convinced of an asset should not only ask the AI: “Is this a good investment?” But: “Build the strongest possible investment case against it.”

• Which assumptions are unusually optimistic?
• Which information contradicts our thesis?
• Which risks have we played down?
• What happens with 20 per cent less rent?
• What happens with an exit yield of seven instead of five per cent?
• Which assumptions would we judge differently if we had not already spent six months on this transaction?

And perhaps the most important question: what information would make us no longer want to do this deal?

If the answer is: “Actually, none.” – we may no longer have an investment model. We have fallen in love.

Because humans start judging astonishingly fast. In a well-known 2006 study by Janine Willis and Alexander Todorov, just 100 milliseconds were enough for participants to form judgements about an unfamiliar face, including likeability, trustworthiness and competence.

More time changed some of these first judgements surprisingly little. It did, however, increase the confidence with which people held them.

An investment is, of course, not a face. But the psychological mechanism behind it should interest us. Even before due diligence begins, a story may already exist:

• This is an exceptional asset.
• This entrepreneur knows what he is doing.
• This location is going to develop.
• This technology will change everything.

Or the other way round: I don’t like this. Then the analysis begins. And this is where a new danger arises.

Artificial intelligence can analyse millions of pieces of information. But it, too, first has to be asked a question. And that question is formulated by a human.

A human who may long since have warmed to the investment. Who has already worked on it for months. Who likes the seller. Who has visited the asset. Who has presented the deal to the investment committee. Who knows a competitor is also bidding. Who does not want to lose six months of work. And who may already be hoping that the analysis will reach a particular result.

Now this person gets a tool of tremendous analytical power. It can make them more rational. But it can also do something else: it can supply much better arguments for what they wanted to believe anyway.

Confirmation bias with artificial intelligence is still confirmation bias. It just gets a more powerful computer.

While we debate whether artificial intelligence will make investors more rational, companies and investors are pouring enormous sums into artificial intelligence itself: data centres. Chips. Energy. Models. Start-ups. Infrastructure.

The Bank for International Settlements is now even studying the dynamics of a possible AI investment race. If companies expect a few winners to control a large share of future markets, individually rational behaviour can lead to an extraordinary amount of capital being invested overall.

And so, in the middle of the world’s most advanced technology, one of the oldest human investment questions resurfaces: what happens if I am not part of it?

New technology. Old emotion.

AI will change the investment business. It will analyse data rooms faster, uncover risks, compare transactions and calculate scenarios. It may see things that humans overlook.

But precisely for that reason, we should perhaps be careful about turning it into an oracle. Its more interesting role could be that of the adversary.

Not: Confirm my investment thesis. But: Destroy it.
Not: Why should I buy? But: What am I overlooking because I want to buy?

Because even the best artificial intelligence is of little help if we already sympathise with an investment before asking the first question, and then mainly keep the machine busy finding reasons for something our brain has long wanted to believe.

Perhaps we do not primarily need artificial intelligence in investing to get more answers. Perhaps we need it to ask better counter-questions about decisions we have already half made.

— Mag. Nevena Shotekova-Zöchling

• Willis, J. / Todorov, A. (2006): First Impressions: Making Up Your Mind After a 100-ms Exposure to a Face, Psychological Science, 17(7), 592–598.
• Akepanidtaworn, K. / Di Mascio, R. / Imas, A. / Schmidt, L. (2021): Selling Fast and Buying Slow: Heuristics and Trading Performance of Institutional Investors, NBER Working Paper 29076.
• Chen, S. / Sialm, C. / Xu, D. (2026): The Growth and Performance of Artificial Intelligence in Asset Management, NBER Working Paper 35273.
• U.S. Securities and Exchange Commission (2013): Knight Capital Americas LLC.
• Zillow Group (2021): Third Quarter 2021 Financial Results and Plan to Wind Down Zillow Offers Operations.
• ESMA (2025): Artificial Intelligence in EU Investment Funds: Adoption, Strategies and Portfolio Exposures.
• ESMA (2024): Public Statement on the Use of Artificial Intelligence in the Provision of Retail Investment Services.
• LAIQON AG / LAIC (2026): Published information on the six-year track record of the AI-based multi-asset strategies (company information).
• Blackstone: Publications on the use of artificial intelligence in the investment process, in particular deal screening, document analysis and portfolio work.
• Bank for International Settlements, BIS (2026): Research on the economic effects of an AI investment race and possible over-investment dynamics.