AI in investing: keep control of rules, strategy and monitoring

Artificial intelligence can organise information, challenge assumptions and make a portfolio easier to understand. It can also sound convincing when its data are stale, incomplete or wrong. Anyone using AI for investing therefore needs a verifiable process, not an autopilot: establish the strategy first, set clear limits, then monitor the portfolio for deviations.

The decisive question is not whether a model appears intelligent. It is whether you can verify its claim and whether the resulting decision still fits your own plan.

What BaFin and ESMA actually say

Since 29 July 2026, BaFin has been the market-surveillance authority for AI systems directly connected with financial activities under its supervision. Its authority is based on Germany's implementing legislation for the EU AI Act and covers, in particular, banks, insurers and other supervised financial companies. It does not turn a freely available AI chat tool into a regulated investment adviser or a reliable source of personal recommendations.

For individual investors, ESMA's 2025 warning is more direct. Publicly available AI tools are often neither authorised as investment firms nor supervised by a financial regulator. They have no general duty to act in their users' best interests and may rely on stale, incorrect or incomplete information. ESMA therefore advises investors not to use these tools as their only source, to seek multiple perspectives and to avoid entering personal or financial data casually.

In 2024 ESMA had already clarified the position for regulated investment firms: when they use AI in investment advice or portfolio management, their MiFID II duties continue to apply. A firm's management remains responsible when a model is involved. That is different from an individual investor using a general online tool. Whether a particular service is regulated depends on its provider and what the service actually does.

Practice is also less autonomous than the phrase “AI investing” suggests. In ESMA's 2025 survey, published in 2026, 77 percent of reported AI use cases operated with low or no autonomy: a person had to approve the output, or the system only provided suggestions. Human oversight is not merely an interim measure; it remains the prevailing arrangement among the financial firms surveyed.

Why an AI output is not an investment strategy

A language model produces an answer to a particular prompt. A slightly different wording, a different data cut-off or missing context may change the result. The output is therefore a hypothesis first, not an investment thesis and certainly not an instruction to trade.

Four risks deserve particular attention:

A sound explanation exposes its evidence, acknowledges uncertainty and can be disproved. A weak one replaces missing evidence with a confident tone. An AI idea should never move directly into an order for precisely that reason.

The investment mandate comes first

A personal investment mandate translates objectives into rules set in advance. It states what the portfolio is meant to achieve, which risks are tolerable and when an intervention is allowed. The framework is written in a calm period, not after a market move or an AI suggestion has captured your attention.

A useful mandate covers at least:

• the objective and time horizon
• financial risk capacity and personal tolerance for loss
• target allocation with deliberately chosen bands
• limits on individual positions and concentrations
• a rebalancing rule
• a rule for unplanned decisions, such as a cooling-off period

A band of plus or minus five percentage points may be an example, but it is not a universal standard. The right width depends on the asset class, portfolio size, costs and taxes, and the chosen rebalancing method.

A stress test can expose contradictions early. If a hypothetical 30 percent equity decline would force a sale or create a burden the investor could not tolerate, that is a reason to reconsider risk capacity, liquidity and allocation. It does not automatically prove that every equity allocation is “too high”.

Set limits before an idea becomes an order

Guardrails need to be measurable. “Well diversified” and “not too risky” are not sufficient. Define the change an idea would make to the portfolio and the conditions it must meet before implementation.

Check, in particular:

1. Does the instrument belong to a permitted asset class and fit the target allocation?
2. What would its total weight be after purchase, including indirect holdings inside ETFs?
3. Would sector, country, currency or issuer exposure exceed your limits?
4. Can you verify the source, date and central assumptions behind the AI claim?
5. Which counterarguments or scenarios would invalidate the thesis?
6. Which costs and tax consequences could the transaction create?
7. Does a fixed waiting period apply before an unplanned order?

A minimum of eight or ten positions does not prove diversification. Ten highly correlated technology shares may carry more concentrated risk than a single broad global fund. What matters is the underlying economic exposure, not the number of lines on the portfolio statement.

Monitoring: measure what the AI idea changed

Monitoring starts with a clean baseline. If you later want to assess whether an AI-assisted idea helped, record at least the date, exact output, data sources, assumptions, expected effect, risks and your own reasoning before deciding. Keep this decision log separately when the portfolio tool you use does not capture the origin of an idea.

The later assessment should not focus on return alone. A positive result may be luck, while a negative one can follow a sound process. A more informative set of measures includes:

• deviation from the target allocation and mandate rules
• concentration after fund look-through
• additional turnover, spread, fees and potential tax
• the position's contribution to total portfolio return and risk
• comparison with the documented expectation and an appropriate reference

AI-assisted and other decisions can only be compared fairly when their time periods, risks, costs and market environments are comparable. A handful of ideas over six or twelve months does not establish that one source has a lasting edge.

There is no universal monthly or quarterly review schedule either. A combination of a fixed date and explicit thresholds is more robust. The date prevents neglect; the threshold reveals drift without inviting a trade after every market move.

Where Investboard helps, and where its boundary lies

Investboard lets you record target allocations, bands and behavioural rules in your investment mandate. Plan Alignment checks the recorded portfolio against that framework each day. Fund look-through also shows which companies, sectors and countries accumulate across several ETFs. This makes it possible to see whether an implemented idea has moved the portfolio away from its plan or increased a previously hidden concentration.

The timing boundary matters: Investboard does not place an order or implement an AI recommendation automatically. Before buying, you must test the idea against your mandate yourself. Once the position appears in a connected account or has been recorded, Investboard can put its effect into the context of the whole portfolio.

Nor should a product claim automatic attribution between an “AI decision” and “your own analysis” unless it captures that origin as structured data. Anyone wishing to evaluate the distinction should therefore keep a separate decision log. Investboard supplies the portfolio and plan context, not proof that a particular source of ideas makes better decisions.

A robust six-step routine

1. Capture the idea. Save the actual AI output, not only your later memory of it.
2. Verify the facts. Open primary sources and check dates, definitions and calculations.
3. Apply the mandate. Test the idea against target allocation, bands and concentration limits.
4. Run a counter-check. State the strongest counterargument and model an adverse scenario.
5. Make your own decision. Record the reasoning, expected effect and invalidation criteria.
6. Review on schedule. Measure drift, risk, cost and outcome at the date chosen in advance.

Conclusion: the tool may change, the standard should not

AI can accelerate analysis and expose blind spots. It cannot know your full capacity for loss or accept responsibility. The durable standard is therefore the mandate written in advance: verifiable sources, defined limits, a deliberate decision and monitoring that measures the real portfolio context.