AI Belongs in Finance — Just Not LLMs Alone.

Bloomberg's latest article on autonomous AI traders validates what portfolio managers have known all along: the human-in-the-loop is essential. General-purpose Large Language Models trading autonomously are not the future of investment management. AI built specifically for finance, designed to support investment decisions, is.

Across a series of new trading contests between the world's leading Large Language Models (LLMs), including Anthropic’s Claude AI, Google’s Gemini, OpenAI’s ChatGPT and Elon Musk’s Grok, the verdict is unflattering. Most systems lose money. They trade too much. They make wildly different decisions when given identical instructions. (Bloomberg article)

Bloomberg's reporting should not surprise anyone who has spent time in institutional asset management. But that conclusion only applies to general-purpose AI systems designed to replace human judgment. It does not apply to finance AI designed to augment it.

The Problem Is Structural — Not Solvable by a Better Model

A model asked in 2026 how it would have traded in March 2020 already knows what March 2020 looked like. That contamination, known as lookahead bias, has challenged the frameworks underlying academic and quantitative finance for decades.

This is not a bug the next model release will patch. General-purpose LLMs are trained on the full historical record with no boundaries on what the model knows when. When you ask that model to predict market behaviour, it draws on information that would never have been available at the time of the decision.

Jay Azhang, founder of Nof1, which ran the Alpha Arena trading contests, summarized the problem directly: "LLMs can't really make money by themselves. You need basically a very sophisticated harness and scaffolding and data platform in order to even give them a chance."

AI Built for Investing

Most AI systems applied to financial markets are general-purpose LLMs handed a financial prompt. The problems Bloomberg documents: inconsistency, overtrading, poor signal calibration are direct consequences of that approach. This is why aisot does things differently. aisot's AI Insights Platform was designed from the ground up to avoid core challenges of using AI in investing:

• AI-driven quantitative models anchored in financial theory. The foundation of the platform is not a language model asked to think about stocks. It is a suite of AI-driven quantitative models grounded in established financial principles — including the Capital Asset Pricing Model (CAPM) and Arbitrage Pricing Theory (APT). These frameworks define the structural relationships the models are built around: factor exposures, risk premia, and return drivers. On top of that foundation sits a machine learning layer that recognises patterns across large, complex datasets — patterns that theory alone cannot surface, but that theory helps validate and interpret. Models can be run with or without news sentiment, which is where LLMs enter the picture — more on that below.
• Multiple time horizons across different markets: Investment decisions are not all made on the same clock. A tactical reallocation and a strategic portfolio review require different signal lengths, different data inputs, and different model calibrations. The AI Insights Platform offers signals across three time horizons — one month, three months, and one year — covering different markets and asset classes. Portfolio managers can match the signal horizon to their actual decision-making cadence, rather than forcing their process to conform to a single model's output window.
• Mandate-aware portfolio construction with built-in constraints. The AI Insights Platform allows portfolio managers to configure frequency (monthly, quarterly, or annually), volatility or tracking error targets, position size limits, and concentration constraints across sectors, geographies, or individual securities. The platform optimizes within those boundaries, meaning the AI's output is already aligned with the manager's mandate and investment philosophy before it reaches a human decision.
• An AI agent that explains its reasoning. The platform does not simply surface recommendations and expect the manager to take them on faith. The built-in AI agent translates complex portfolio data into clear, actionable insights — including performance drivers, risk attribution, and the explicit reasoning behind each recommendation. The manager understands not just what the AI suggests, but why. The final investment decision always rests with the portfolio manager.

Where LLMs Belong in Finance and How aisot Uses Them Correctly

There is one area where large language models excel in a financial context and it is not autonomous trading. It is the real-time interpretation of unstructured information. But even here, how you deploy them matters.

Markets move on news. Earnings calls, central bank communications, geopolitical developments, analyst commentary. The signal is there, but extracting it at scale, and with the speed institutional decision-making requires, has historically been beyond what human analysts can do alone.

aisot uses LLMs for news sentiment analysis and estimated return forecasts. Our sentiment layer uses LLMs to process and interpret financial news in real time, converting unstructured text into structured sentiment signals that feed into the platform's quantitative models. Managers can choose to run models with or without this layer — the quantitative foundation remains the same either way.

"News sentiment typically accounts for 10 to 15 percent of the estimated return forecast in our models. It is a meaningful signal, but it is one input within a rigorous quantitative framework, not a standalone driver. That distinction matters when you are making real investment decisions.", Dr Nino Antulov-Fantulin, Co-Founder & Head of Research of Aisot Technologies, says.

aisot's approach diverges fundamentally from the general-purpose LLM deployments Bloomberg describes. We use time-boxing and fine-tuning to prevent hallucination and lookahead bias. When processing information for a given date, the model is explicitly bound to what was knowable at that point in time — no future data, no backward contamination. This is technically demanding to implement correctly and it is also the reason aisot's signals are predictive rather than retrospectively calibrated.

The Right Conclusion Is Financial AI with Human in the Loop

The Bloomberg article will be cited by some as evidence that AI has no place in portfolio management. The correct reading is narrower and more useful: AI built for finance has its place as decision-support. It’s purpose-built for institutional investment workflows, grounded in financial theory, with time-boxed and fine-tuned LLMs, configurable mandate constraints, and a human making every final call.

Asset managers who conflate the two will either adopt the wrong version of AI or avoid it altogether. Both are costly mistakes in a market where the information advantage is the competitive advantage.