---
title: "AI trading bots: What they can and cannot do"
description: "Assess AI trading bots: backtests, provider authorisation, costs, custody and warning signs. An investor explainer, not a hands-on product test."
canonical: "https://investboard.de/en/wissen/ki-trading-bots-im-test-was-automatisierte-handelssysteme-wirklich-leisten-und-warum-privatanleger-die-kontrolle-behalten-mussen"
language: "en"
published: "2026-09-12"
modified: "2026-09-19"
last_verified: "2026-09-19"
jurisdiction: "DE / EU"
author: "David Bartas"
author_url: "https://investboard.de/ueber-uns"
---

# AI trading bots: What they can and cannot do

How automated trading works, which evidence matters and how individual investors can stay in control.

## In short

AI trading bots automate orders; they do not guarantee profits. Check the provider, permissions, custody and live results after costs. Backtests alone are insufficient. Set limits and stop conditions in advance, and monitor the total portfolio independently. This article explains how to assess the evidence; it is not a hands-on product test.

- Automatic execution does not establish that a strategy is sound.
- Authorisation, custody and protection depend on the actual service.
- Backtests and live results need to disclose costs, risks and losing periods.
- Your mandate determines limits, review dates and stop conditions.

An AI trading bot that operates around the clock, never hesitates and supposedly delivers steady returns is an attractive promise. But automation describes how orders are placed, not whether a strategy is sound. A convincing performance chart does not establish who holds your money, whether the provider is authorised or how much you could lose.

This article examines the technology, evidence and risks behind that promise. It is an explanatory assessment, not a hands-on comparison or a ranking of tested products. The regulatory context is Germany and the EU; readers elsewhere need to check their own jurisdiction.

## What is an AI trading bot?

### Rules, machine learning and marketing

A rule-based algorithm follows instructions such as buying when an indicator crosses a threshold. A machine-learning model estimates patterns from training data. A language model may summarise news or produce a trading idea. A platform can combine these approaches, but none of the labels proves that it can predict markets or generate sustainable profits.

Ask what the system actually does: which inputs it receives, how it produces a signal, when a person intervenes and who can change the strategy. A provider that cannot explain those points has not given you enough information to assess the product. Sophisticated software can still implement a poor investment idea.

### Three services that should not be confused

A **robo-adviser** typically offers automated portfolio management or advice. A **trading bot** executes orders under its configured rules or signals. A **signal service** supplies ideas that someone else may execute. Their legal treatment depends on the activity, instruments and provider, not the marketing name. Software sold for an investor's own use is not automatically equivalent to a regulated investment service.

A further distinction matters: an analysis tool can explain a portfolio without placing trades. See [AI in investing: keep your rules, strategy and monitoring](/en/wissen/ki-in-der-geldanlage-regeln-strategie-und-monitoring-behalten) for that separation of roles.

### Why a backtest is not a live track record

A backtest applies a strategy to historical data. It can help investigate an idea, but it becomes misleading when parameters are repeatedly selected to fit the same sample. The resulting overfitting may describe the past beautifully and fail on new observations. Using information that would not have been available at the time creates another distortion.

Ask for the data period, sources, selection rules and results on data excluded from optimisation. Costs, spreads, slippage and plausible execution delays belong in the simulation. Backtests can model these factors; the question is whether the advertised one actually does. Even a careful simulation cannot reproduce every future market condition.

### What automatic execution changes

A bot may open or close positions while you are unavailable. It may also repeat a faulty instruction before you notice. Understand its permissions, maximum exposure, leverage, error handling and stop procedure before enabling it. Removing hesitation from execution does not remove risk from the decision.

## The regulatory check: verify the provider

### Start with the service and the legal entity

Look beyond a brand name. Identify the contracting company, its address, the authority responsible for it and the precise services covered by its authorisation. An entry in a register must match the entity and activity you are using. Fraudulent websites can also copy the details of a real authorised firm.

Check the [BaFin company database](https://portal.mvp.bafin.de/database/InstInfo/) and [consumer warnings](https://www.bafin.de/DE/Verbraucher/Verbraucher.html), and use the relevant home-state regulator and ESMA registers where appropriate. A foreign provider may operate under an applicable cross-border regime. Conversely, a registration for one activity does not authorise every financial service.

### A practical verification sequence

1. Record the legal entity, website domain and claimed authorisation number.
2. Open the regulator's register independently, rather than relying on a badge or link supplied by the provider.
3. Compare the authorised activities and contact details with the actual offer. If uncertain, contact the firm through details in the register.
4. Search current warnings for the company, domain and related names. An absence of warnings is not an endorsement.
5. If the legal basis for a service remains unclear, resolve that uncertainty before transferring money.

### AI does not replace investor-protection obligations

[ESMA's May 2024 statement](https://www.esma.europa.eu/press-news/esma-news/esma-provides-guidance-firms-using-artificial-intelligence-investment-services) explains that investment firms using AI remain subject to applicable MiFID II obligations, including acting in their clients' best interests. A supervised firm cannot treat a model as an excuse to abandon its responsibilities.

That is different from using a public online AI tool. [ESMA's 2025 investor warning](https://www.esma.europa.eu/sites/default/files/2025-03/ESMA_Warning_on_the_use_of_AI_-_EN.pdf) highlights inaccurate information, misleading recommendations, privacy risks and the danger of relying on these tools alone. Do not infer that an app is supervised merely because it uses AI or refers to European regulation.

## The risks behind the performance chart

### Market and model risk

A strategy that worked in one market regime can fail when volatility, liquidity or correlations change. A model trained on ordinary conditions may behave poorly in a shock. Testing several scenarios helps reveal weaknesses; it does not establish a maximum possible loss.

### Execution, costs and operational failures

The price when a signal is generated may differ from the executable price. Spreads, commissions, financing costs and slippage can materially change the result, especially with frequent trading. Software errors, unavailable APIs and connection failures introduce further risks. Ask what happens to existing positions when the bot or exchange connection stops working.

Compare live results after costs with an appropriate benchmark, over a meaningful period and at a comparable level of risk. A screenshot of a winning account leaves unanswered questions about deposits, withdrawals, unsuccessful strategies and whether the record was independently verified.

### Counterparty and custody risk

Identify where cash and securities are held and which claims you would have against each entity. Deposit protection, custody rights and fund-asset segregation are different mechanisms. They do not guarantee investment returns or protect against ordinary market losses. Not every security is legally a fund's segregated asset, and not every balance shown in an app is a protected bank deposit.

Do not assume that a provider's use of a familiar bank makes the entire arrangement protected. Read the account and custody terms. If assets are missing or the provider is fraudulent, recovery can be difficult even where legal claims exist. The absence of supervision does not mean that victims have no legal rights; it can make enforcement and recovery much harder.

### Concentration risk

A bot can change your total exposure while the rest of the portfolio appears unchanged. It may add technology shares already held indirectly through ETFs, concentrate positions in one currency or increase leverage. Review the combined portfolio, not only the bot account. [ETF overlap](/en/wissen/etf-ueberlappung) explains why several holdings do not necessarily provide several independent sources of risk.

### Behavioural risk

Automation can encourage investors to stop checking. A polished interface and repeated profitable trades may create unjustified confidence. The reverse is also possible: frequent intervention after losses can destroy the logic of a strategy. Write down when to review, what to measure and which events require action before emotions enter the decision.

## How the available approaches differ

The useful comparison is between activities and evidence, not between products displaying the word “AI”. None of the categories below is inherently profitable.

| Approach | What it does | What to check |
| --- | --- | --- |
| Trading bot | Executes a configured trading strategy | Provider status, permissions, live results, costs and failure handling |
| Robo-adviser | Provides portfolio management or advice, depending on the service | Relevant authorisation, suitability process, fees and custody |
| Rules-based or factor ETF | Tracks a published index methodology | Index rules, holdings, costs, liquidity and market risk |
| Signal or analysis tool | Supplies information or suggestions | Data quality, conflicts of interest and whether the specific service is regulated |

### Five questions before considering an offer

- Which legal entity provides the service, and what authorisation does the activity require?
- How was the strategy developed and tested outside its training sample?
- Is there a verifiable live record after all relevant costs, including losing periods?
- Does the provider benefit from more trades, higher leverage or particular products?
- Who holds the assets, and how can you stop the service and withdraw or transfer them?

There is no universal number of backtest years that proves a strategy is robust. What matters includes the independence of the test data, the variety of conditions, the economic explanation and the gap between simulation and actual execution. A transparent answer can still reveal that the product does not fit your objectives.

## Guardrails: keep control of the portfolio

### Write the investment mandate first

Define your objective, horizon, liquidity needs, target allocation and tolerable losses before choosing a tool. State whether automated execution is allowed at all, which instruments it may use and which permissions it must never receive. [The investment mandate](/en/wissen/das-anlage-mandat) provides a framework for documenting those choices.

### Set exposure limits deliberately

Choose limits for total capital, individual positions, leverage and correlated exposures. A fixed allocation such as ten percent is not a universal safe threshold. A suitable limit depends on the strategy, your circumstances and the potential loss, and may be zero. Money needed for essential spending should not depend on an experimental strategy succeeding.

### Decide what monitoring should answer

Record an initial baseline and a review schedule. Measure returns after costs, drawdowns, exposure, mandate deviations and operational incidents. Select a benchmark suited to the strategy's assets and risk; a global equity index is not automatically the right comparison for every bot.

Your review should answer whether the strategy still behaves as intended and fits the overall portfolio. A positive return alone does not answer either question. Keep the monitoring sufficiently independent of the seller to cross-check what the seller reports.

### Define stop conditions before a crisis

Possible triggers include unexplained positions, changed permissions, a regulatory warning, missing statements, an unexpected strategy change or losses beyond your predetermined risk budget. Numerical thresholds are personal controls, not guarantees that losses will stop there. Gaps, leverage and unavailable systems can defeat an intended exit price.

Document how to disable new orders, revoke access and manage positions already open. Understand whether stopping the software also closes positions; those are not the same action. Test operational procedures without assuming that a successful small withdrawal proves the platform is legitimate.

### Separate trading from oversight

Investboard does not execute trades or manage assets. Its role is to help you analyse recorded holdings, look through ETFs and compare the portfolio with your own mandate. Where a bot account is not included or its data are incomplete, that view is incomplete too. No portfolio dashboard can independently certify a bot's safety or prevent all losses.

## Checklist: warning signs before money moves

### Guaranteed returns

Claims of guaranteed trading profits or unusually high returns without corresponding risk deserve strong scepticism. Ask what contractual guarantee is actually being offered, by whom and with what exclusions. An advertising slogan is not evidence of protection.

### Unverifiable identity or authorisation

An unexplained company name, a mismatch with register details or copied credentials are reasons to stop and investigate. A licence logo alone proves nothing. Confirm the entity and the scope of the permission independently.

### Selective or opaque performance evidence

A rising curve without methodology, dates, costs and losing periods is insufficient evidence. Be cautious when the seller substitutes testimonials or screenshots for a complete record. Ask who verifies the numbers and what the verification actually covers.

### Pressure and deposit bonuses

A countdown, repeated calls or a bonus contingent on an immediate deposit can discourage proper checking. A sound decision does not require accepting the seller's deadline. Step away if pressure replaces clear answers.

### Obstacles to withdrawals

Unexpected payments demanded to “unlock” an account or release a withdrawal are serious warning signs. Do not keep sending money simply because the app shows a profit. Preserve correspondence and transaction records, contact your bank promptly and consider reporting suspected fraud to the police and the relevant regulator. Be wary of anyone promising recovery for another advance payment.

## Conclusion: automation needs a clear mandate

AI can accelerate analysis and automate execution. It cannot establish your objectives, make an unsuitable risk acceptable or take responsibility for your portfolio. The essential checks remain the same: understand the strategy, verify the provider, assess credible evidence and retain practical control over access and exposure.

Start with the mandate, examine results after costs and keep an independent view of the total portfolio. If the provider, custody arrangements or strategy cannot be understood well enough to assess them, there is no need to proceed. Automation is optional; an informed decision is not.

This article provides general information, not individual investment or legal advice. Investments can lose value, and past or simulated returns do not reliably predict future results.

## Frequently asked questions

### Is this article a hands-on product test?

No. It explains how automated trading systems work, their risks and criteria for checking providers. It contains no hands-on product tests, ranking or return guarantee.

### Does a good backtest prove that a bot works?

No. Overfitting, unsuitable data and unrealistic cost assumptions can distort a backtest. Independent test data and verifiable live results provide additional evidence, but do not guarantee future profits.

### Is ten percent of a portfolio a safe bot allocation?

There is no universally safe allocation. Limits depend on the strategy, potential losses, liquidity needs and personal circumstances. An appropriate limit may be zero.

## Sources

- [Using Artificial Intelligence for Investing: What you should consider](https://www.esma.europa.eu/sites/default/files/2025-03/ESMA_Warning_on_the_use_of_AI_-_EN.pdf) · ESMA
- [Guidance to firms using artificial intelligence in investment services](https://www.esma.europa.eu/press-news/esma-news/esma-provides-guidance-firms-using-artificial-intelligence-investment-services) · ESMA