By Bruce Keith
AI is no longer just in the background; it is reading market data, writing research and increasingly nudging people towards buying and selling.
Research for the FCA’s Mills Review found that around 26% of adults trust tools such as ChatGPT, Claude or Gemini for financial advice, yet only 40% know there is no formal route to recourse if that advice goes wrong.
The stakes rise as these tools start to act, not just advise.
As the Mills Review puts it, “As AI moves from recommending to acting, and firms and consumers delegate more, risks shift from harm within a single firm towards system-wide harms.” So the question is not whether investors will use AI, but how they can tell which AI deserves their trust.
My answer is simple. Investment intelligence must be judged by the consistency of the insights it produces, by how responsibly they are generated and communicated, and ultimately by whether an organisation takes responsibility for its output.
Not all AI is the same
Before you get to trust, you have to have the right tools for the job. LLMs are excellent at reading news, summarising filings and screening sentiment, but there is academic evidence that they are poor at stock selection: what to buy, when to buy, when to sell and what to replace it with. You don’t ask a dentist to perform brain surgery, so don’t ask a chatbot for stock tips when the model was never trained for it and to make matters worse, it is likely to sound plausible.
Prediction needs domain-specific systems. At InvestorAi, we built neural networks that read prices, volumes, fundamentals and corporate events as images, recognising patterns in noisy, incomplete data. This is also different from algo trading, which follows pre-written rules which can be copied. In AI investing, the AI does the research, finds opportunities and builds portfolios by finding its own patterns in the data.
My preferred architecture uses specialist predictive models for stock picking and portfolio construction, an LLM for news and sentiment screening and agents to orchestrate execution within very clear guardrails.
Judge the process, not just the prediction
A good call can be luck, but getting it right on a consistent basis matters more. It’s the difference between betting blind on the Grand National and running masses of data through domain-specific trained models to pick a runner. Neither is guaranteed to win, but I know which one I would side with. How does this translate to investing with AI?
- A track record you can inspect. Unlike a chatbot’s answer, a live model leaves a record. Ours has picked stocks for real investors in India, under regulatory oversight, since April 2021. It won’t get everything right, but you can see exactly what it has done. Look for consistency and repeatable performance over time rather than one standout year.
- Disciplined testing. Treat back-tested results with caution and in the absence of a live track record, make sure models were tested against data they had never seen. In our own research, hold-out data is sealed and opened exactly once, and every failure is logged. Rules like these stop a model from being tuned until it merely looks good on paper.
- Honesty about uncertainty. A trustworthy system says when it isn’t confident. Our models use conviction gating: they act only when their confidence clears a learned threshold; otherwise they stay out.
Accountability cannot be outsourced
Technology changes how advice is produced, not who is responsible for it. Regulators in different markets are reaching the same conclusion. In the UK, the Treasury reports strong industry support for the FCA’s “tech-neutral, outcomes-focused approach”. The Chairman of SEBI (India’s securities regulator), Tuhin Kanta Pandey, is quoted as saying, “Technology cannot dilute accountability”. It’s impossible to disagree.
That responsibility cannot be passed to a model, or to an insurer. Disclaiming “model behaviour” is not good enough; providers must own outputs, not just intent. Human oversight, clear processes and ownership of outcomes are a far better guide to how responsibly an organisation deploys AI. Human-in-the-loop needs to be more than a box-ticking exercise, and for us, every AI-picked strategy we offer runs with expert oversight. As Mills notes, “Oversight can also weaken if reviewers are overloaded, lack the right information or rely too readily on model outputs.”
As AI moves from recommending to acting, guardrails matter as much as models. Agents need kill switches, position limits and the ability for a human to override them at any point. Ultimately, regulation sets a floor. AI is moving faster than any regulator can write rules, so it falls to the industry to go beyond that floor rather than wait for it to catch up.
Intelligence, with accountability
AI can close information gaps that have long disadvantaged ordinary investors. In the UK, only around 9% of adults use traditional financial advice, and it’s lower in India. Because AI’s marginal cost falls with scale, good investment research no longer has to be a privilege of the few. That is what AI is for.
That promise holds only if the intelligence is earned: the right model for the job, tested honestly, explained clearly, and with a regulated party answering for the outcome. Trustworthy AI investing is not the AI that sounds most confident. It is the AI that can demonstrate consistency, with someone standing behind it.
(Bruce Keith is the CEO and Cofounder of InvestorAi. He is based in Mumbai and London and the views are his own)
The post Column: What trustworthy AI investing should look like appeared first on Invezz
