Key takeaways
- An AI agent has no legal personality. Duties attach to you, the agent provider and the licensed firm that executes.
- The label decides the duty: information, group guidance and personal advice differ. UK targeted support went live on 6 April 2026.
- In the EU, Article 50 of the AI Act has applied since 2 August 2026: you must be told you are dealing with an AI system.
- Claiming AI that is not there is a breach. On 18 March 2024 two US advisers agreed to pay US$400,000 in total penalties.
- Your approval covers the action as the preview described it, not a wrong preview or an undisclosed limitation.
- If something goes wrong: stop the access, keep the evidence, complain in writing, escalate, then take independent advice.
This guide is one of the AMBR guides to AI agents in finance. It covers responsibility and regulation only. For the mechanics, read what an AI finance agent is. For settings and limits, read about the controls that limit what an agent can do.
The short answer
An AI agent is software, not a legal person, so every rule cited in this guide places its duties on a person or a firm. The EU AI Act speaks of providers and deployers, and ESMA said on 30 May 2024 that investment firms using AI are still expected to meet their existing obligations to clients.
In practice responsibility sits in three layers:
- You answer for your instructions, the access you grant and the actions you approve on an accurate description.
- The firm that provides a regulated service, such as executing an order or holding assets, answers for how that service is delivered under its licence.
- The agent provider answers for what it promised and disclosed, and for whether what it showed you was right.
This is general information, not legal advice. In a real dispute the outcome depends on the contract you accepted and on local law.
Terms used in this guide
- Legal personality: the capacity to hold rights and duties, to own property and to be sued. People and companies have it. Software does not.
- Provider and deployer: the EU AI Act's words for the body that develops an AI system and places it on the market, and the body that uses it under its own authority.
- Licensed firm: a company that holds a regulator's permission for a named activity, such as dealing, custody or advice.
- Approval step: the point at which a person must confirm an action before it is sent, also called a human in the loop.
The glossary has AI washing and other terms defined.
Who is in the chain
Five parties usually stand between your instruction and a completed transaction, and responsibility tends to shift at five hand-off points. Figure 1 numbers them. One company can fill more than one role, so the first question in a dispute is which legal entity did what.
You
You choose the agent, connect the accounts, set what it may do and approve or reject what it prepares. You answer for those choices, and most terms also make you responsible for keeping passwords and API keys secure. Errors you had no way to see, such as a preview that differs from the order sent, start with the firms in the chain.
The agent provider
The agent provider is the company whose name is on the product. It decides what the agent says, what the preview shows and which actions need your approval. It answers for its marketing claims, for the AI notice and for the accuracy of what the agent displays. It may also hold a licence and execute, but often it does not.
The model provider
The model provider supplies the AI model under a contract with the agent provider, not with you, so a model error reaches you through the agent provider. Two regulators say plainly that a firm cannot hand this duty to a supplier: the Central Bank of the UAE in February 2026 for outsourced AI functions, and Hong Kong's SFC in 2024 for delegated ones.
The licensed firm that executes
The licensed firm is the broker, exchange, custodian or payment institution that carries out the transaction. It holds the regulator's permission and the conduct, safeguarding and complaint-handling duties that come with it. Your regulator and your complaint route follow this entity, not the brand on the agent. Check its legal name on the public register.
The data aggregator
The data aggregator connects your brokerage, exchange and wallet accounts and passes balances and transactions to the agent. Stale or wrong data produces a wrong signal. The type of connection matters too: a read-only connection cannot place an order, while a key with trading or withdrawal rights can.
Information, guidance or advice: why the label decides the duty
The same sentence from an agent can count as information, guidance or advice, and each label carries a different duty. The UK shows the middle rung most clearly: the FCA's targeted support regime went live on 6 April 2026.
- General information. Facts that are the same for everyone, or plain facts about your own accounts, such as the share of your portfolio held in one asset. The duty is to be accurate, clear and not misleading.
- Guidance for groups of similar people. A suggestion built for people in circumstances like yours, without a full review of your situation. The FCA describes targeted support as "suggestions designed for groups of consumers with common characteristics", and a UK firm needs a specific permission to provide it for pensions and retail investments.
- Personal advice. A recommendation based on your own circumstances and goals. Giving it is typically a licensed activity: Dubai's VARA, for instance, lists Advisory Services among the activities that need a licence.
The FCA created the middle rung because it estimates that around 23 million UK consumers are underserved by advice and guidance.
For liability, a label on a website does not by itself settle the matter. What the product actually says to you does. A tool that only informs carries different duties from one that advises, so a provider has to keep its product inside the label it claims. That is why AMBR describes Ambre's output as information and monitoring, and why it matters how Ambre separates information from execution.
What the rules say, by jurisdiction
Eleven published rules and statements from six jurisdictions already bear on AI agents in finance. The table gives each one with its date and a link to the source. None of them makes the software itself answerable.
| Jurisdiction | Rule or statement | What it requires or says | Date |
|---|---|---|---|
| European Union | AI Act, Article 50(1) | Providers must design AI systems that interact directly with people so that those people know they are dealing with AI, unless that is obvious. | 2 Aug 2026, applies |
| European Union | AI Omnibus, Regulation (EU) 2026/1744 | Moves high-risk obligations for stand-alone Annex III uses to 2 December 2027. The Article 50(1) notice date is unchanged. | 27 Jul 2026, in force |
| European Union | ESMA statement on AI in retail investment services | Firms using AI with retail clients are expected to comply with MiFID II and act in the client's best interest. | 30 May 2024 |
| United Kingdom | FCA PS25/22, targeted support | A firm needs a permission to give suggestions designed for groups of consumers who share common characteristics. | 6 Apr 2026, live |
| United States | SEC press release 2024-36 | Settled charges against two investment advisers for false and misleading statements about their use of AI; US$400,000 in total civil penalties. | 18 Mar 2024 |
| United States | 17 CFR 4.41(b), a CFTC rule | Simulated or hypothetical performance may be presented only with a prescribed limitations statement. | current rule |
| UAE, Dubai | VARA list of licensed activities | A licence is required before any of eight listed virtual asset activities, including Advisory Services and Management and Investment Services. | undated page |
| UAE, onshore | CBUAE Guidance Note on AI and machine learning | Institutions the Central Bank licenses should tell customers when they are dealing with an AI application, disclose in Arabic and English and stay responsible for outsourced AI functions. | 11 Feb 2026 |
| Hong Kong | SFC circular on generative AI language models, as summarised by Clifford Chance | Investment recommendations, advice and research are high-risk uses; users should be told they are interacting with AI. | 12 Nov 2024 |
| Singapore | MAS Guidelines PS-G02 | Digital payment token service providers are expected not to promote those services to the general public in Singapore. | 17 Jan 2022 |
| Singapore | MAS consultation P017-2025 on AI risk management | Proposed guidelines for all financial institutions on oversight, risk management and life-cycle controls for AI. MAS responded on 7 October 2026. | 13 Nov 2025 |
Each rule links to the document it was read from. Last reviewed .
European Union
Article 50 is the EU rule most likely to touch a finance agent today. Since 2 August 2026 providers must design systems that interact with people so that those people "are informed that they are interacting with an AI system", clearly and no later than the first interaction. The AI Omnibus moved the Annex III high-risk obligations to 2 December 2027 without moving that notice. Whether an agent is high-risk depends on its use, and this guide does not classify any product. ESMA, for its part, names four risks of AI in retail investment services: biased algorithms and poor data, opaque decision-making, overreliance, and privacy and security.
United Kingdom
The UK regulates by activity. Giving suggestions to groups of similar consumers about pensions or retail investments needs the FCA's targeted support permission, live since 6 April 2026. In December 2025 the FCA and the Financial Ombudsman Service published a joint statement on complaints about targeted support. An AI agent changes none of this: the permission and the duties sit with the firm.
United States
US regulators have acted on claims rather than on agents as such. The SEC's action of 18 March 2024 is covered in the AI washing section below. A separate CFTC rule, 17 CFR 4.41(b), lets commodity pool operators and commodity trading advisors show simulated or hypothetical results only with a statement that they "have certain inherent limitations". The rule looks at what is presented, not at the tool that produced it.
Dubai and the wider UAE
Two regulators matter here, with different remits. Dubai's Virtual Assets Regulatory Authority (VARA) requires a licence before any of eight listed virtual asset activities, and both Advisory Services and Management and Investment Services are on the list. An agent that advises on or manages virtual assets in or from Dubai therefore points to one of them. Onshore, the Central Bank's Guidance Note of 11 February 2026 also says the board and senior management of each institution should be accountable for AI outcomes.
Hong Kong
The SFC's circular of 12 November 2024 is addressed to the firms it supervises. For high-risk uses it expects a person to review AI output before it reaches the client, and a firm that delegates part of the work remains responsible for compliance. This paragraph and the table row rely on Clifford Chance's summary of the circular.
Singapore
Singapore reaches AI agents from two sides. PS-G02 limits how an agent linked to a digital payment token service can be marketed to the public there. On governance, MAS consulted from 13 November 2025 to 31 January 2026 on Guidelines on Artificial Intelligence Risk Management and published its response to the feedback on 7 October 2026.
What changes when a person approves each action
Approval moves part of the decision to you, but only for the action the preview actually described. Hong Kong's SFC, in its circular of 12 November 2024, lists human review of AI output as a control for high-risk uses while keeping the firm responsible.
Three principles follow. They are principles, not predictions of how a court or an ombudsman would rule.
- Approval covers what you were shown. Asset, direction, amount, price or limit, fees and destination belong in the preview. What was not shown was not approved.
- Approval does not cure a defect upstream. A misleading prompt, a wrong preview or a limitation that was never disclosed, such as an account the agent cannot see, stays with the firm that caused it.
- Standing approvals shift more to you. A rule that lets the agent act without asking each time is still your instruction, so give it tighter limits.
Worked example, with invented figures
You hold a portfolio worth US$10,000 and tell an agent to keep any single asset below 30% of it. One asset grows to US$3,600, which is 36%. The agent prepares a sale of US$600 to bring the share back to 30% and shows you a preview. You approve it.
| Case | What happened | Where the question starts |
|---|---|---|
| A. Accurate preview | US$600 was sold, as shown. The price rose afterwards. | With you. The preview was accurate and the decision was yours. |
| B. Wrong order | US$6,000 was sold, which is US$5,400 more than you approved. | With the agent provider and the executing firm. The figure changed after your approval. |
| C. Wrong data | US$600 was sold, but a stale balance had overstated the holding. Its true share was 28%, so no sale was needed. | With the agent provider and the data aggregator, and with what you were told about how fresh the data was. |
The figures are invented to show the reasoning. They do not come from a real account, product or decided case, and your terms may allocate each gap differently. The companion guide explains the controls that limit what an agent can do, including read-only access and spending limits.
AI washing: when the claim itself is the breach
AI washing is a firm claiming AI capabilities it does not have, and the SEC has penalised it since 18 March 2024. On that day it announced settled charges against Delphia (USA) Inc. and Global Predictions Inc. for false and misleading statements about their use of AI. The SEC Chair said the firms told clients they were using AI "in certain ways when, in fact, they were not". Together they agreed to pay US$400,000 in civil penalties.
The charges were about what the firms said, not about a trade that went wrong. Before you rely on an agent, check that its AI claims are specific and testable:
- What exactly does the AI do, and what is done by people or by fixed rules?
- Which of your data does it use, and how fresh is that data?
- Which actions need your approval before anything is sent?
- Does the provider publish limits, known failure modes and an incident record?
A phrase such as "AI-powered" with no description behind it is a reason to ask more questions. Performance claims deserve the same scrutiny: this site publishes no backtests and no performance figures for Ambre.
What contracts usually say, and what to look for
Terms of service usually allocate a loss before any regulator is involved, and six clauses do most of the work. Read them before you connect an account, not after a dispute starts.
- The contracting entity. Look for a full legal name and a country, not only a brand.
- Accuracy disclaimers. Expect statements that outputs can be wrong and are not advice. They do not by themselves remove duties that a licence imposes.
- Liability caps. Many terms limit what the provider will pay, often by reference to the fees you paid, and exclude some kinds of loss.
- Who executes. The terms should name the firm that carries out transactions, its licence and whether it is a group entity or a third party.
- Governing law and forum. This decides which country's law applies and where a dispute is heard. One example: the terms of use for MIA, AMBR's marketing agent, are governed by the laws of the Hong Kong SAR.
- The complaint route. A named address, a response time and an external body you can escalate to.
This is a general description. Read the actual terms of the product you use, because they override anything in this guide. AMBR's own are set out in its terms of use and AI disclosure.
If something goes wrong: five steps
Five steps, taken in this order, protect your position after a costly error. The first should take minutes, because an agent with live access can keep acting while you investigate.
1. Stop the access
Pause or switch off the agent, then revoke its access at the source: delete the API key at the exchange or broker, disconnect the data aggregator and change any password you shared. Removing a connection inside the agent does not always remove it at the venue.
2. Keep the evidence
Save what you saw before it changes: your instruction, the preview, your approval, the confirmation, order references and timestamps. Export the chat history if you can, and keep a copy of the terms and the marketing page as they stood that day.
3. Complain to the provider
Write to the provider's published complaints address. State what you instructed, what the preview showed, what happened, the loss and the remedy you want, and ask for a reference number. AMBR's page on how to make a complaint shows the route for its own entities.
4. Escalate to the firm that executed and its regulator
If a different firm carried out the transaction, complain to it too. The route beyond that depends on which entity served you and where it holds its licence: each regulator's complaint or ombudsman channel covers only the firms it supervises. Check the entity on the public register, because the name on the app is not always the name on the licence.
5. Take independent advice
For a significant loss, speak to a lawyer or a consumer body in your own jurisdiction. Time limits vary and can be short. This guide cannot tell you whether you have a claim.
How AMBR states the split for Ambre
AMBR states the split for Ambre, its AI agent for personal finance, in one sentence on the live product page, read on 11 October 2026.
The scope sentence, quoted in full
"Where you choose to pursue a transaction or other financial activity, relevant services may be provided within AMBR's licensed scope or through applicable third-party providers."
ambre, as stated by AMBR
- does
- Gives information: a portfolio view, market signals and monitoring tasks.
- does not
- Give personal investment advice. Ambre is an AI system and its outputs can be wrong.
- executed by
- A licensed AMBR group entity or a third-party provider, not the AI.
In the terms of this guide, AMBR is the agent provider, and execution belongs to a licensed entity or a third party. The public registers showed this on 11 October 2026:
- Dubai
- Amber Premium FZE is licensed by VARA for Broker-Dealer Services, Management and Investment Services and Lending and Borrowing Services: VASP Licence VL/26/03/003, issued 2 April 2026 (VARA register entry). Amber Premium FZE is permitted to serve Institutional Investors and Qualified Investors. The entry does not list Advisory Services.
- Singapore
- Sparrow Tech Private Limited, trading as Amber Premium Singapore, is licensed by MAS as a Major Payment Institution for digital payment token services (MAS directory entry).
- Hong Kong
- Whalefin Markets Limited is a virtual asset trading platform applicant that is deemed to be licensed under Schedule 3G of the Anti-Money Laundering and Counter-Terrorist Financing Ordinance. It is not licensed by the SFC.
- Listed parent
- Amber International Holding Limited (Nasdaq: AMBR) claims no financial licence of its own.
- Not yet public
- Which entity serves Ambre users in each market, and under which licence. AMBR's site names no licence or registration in the United Kingdom, the European Union or the United States. Ask which entity you are contracting with before you act on a signal.
The full record is on the page of licences held by AMBR group entities, with register links. The approval model is described under security, approvals and incident log, and the About page covers the company and its regulated entities.
Not financial advice. Crypto is highly volatile and may drop in value significantly. You may lose the amounts you invest and your investments do not benefit from any form of financial protection.
This article is for information only and is not investment, legal or tax advice.
Frequently asked questions
Is the AI agent itself legally liable?
No. An AI agent is software with no legal personality, so it cannot be sued, fined or made to pay. Duties attach to the people and firms around it: you, the agent provider and the licensed firm that executes. Which of them answers in your case depends on the contract and on local law.
Who pays if an AI agent makes a wrong trade?
It depends on where the error arose. If you approved an accurate preview, the result usually follows your decision. If the preview was wrong, or the order differed from the one you approved, the question moves to the agent provider and the executing firm. The contract, including any liability cap, and local law decide the outcome.
Is it legal for an AI agent to trade for me?
It can be, when the firm behind each step holds the permission that step needs. Executing orders, managing assets and giving advice usually need a licence; the software needs none of its own. Dubai's VARA, for example, requires a licence before any of eight listed activities. Whether a service is lawful where you live depends on its contract and on local law.
Does the EU AI Act treat finance agents as high-risk?
It depends on what the agent is used for: the Act classifies uses, not product names. High-risk obligations for stand-alone Annex III uses apply from 2 December 2027. The Article 50 duty to tell people they are dealing with AI has applied since 2 August 2026. How the Act treats one product depends on its use, its contract and local law.
Does a disclaimer remove the provider's responsibility?
No, not by itself. A disclaimer tells you what the service is and is not, and it can limit some contractual claims. It does not switch off duties that a licence or a regulator's rules place on a firm, and it does not excuse a misleading claim. How far a disclaimer reaches depends on the contract and on local law.
Sources
Eight regulator and official documents carry the main claims, all read on 11 October 2026. Three more are linked in the jurisdiction table where they are used: 17 CFR 4.41, Clifford Chance's summary of the SFC circular and the MAS consultation page.
- EU Artificial Intelligence Act, Regulation (EU) 2024/1689, Article 50, artificialintelligenceact.eu.
- European Commission, "AI Act" policy page with the application timeline, last updated 3 August 2026, digital-strategy.ec.europa.eu.
- European Securities and Markets Authority, "ESMA provides guidance to firms using artificial intelligence in investment services", 30 May 2024, esma.europa.eu.
- Financial Conduct Authority, PS25/22, rules for targeted support, published 11 December 2025, updated 27 February 2026, fca.org.uk.
- US Securities and Exchange Commission, press release 2024-36, 18 March 2024, sec.gov.
- Virtual Assets Regulatory Authority, "Licensed Activities", vara.ae.
- Central Bank of the UAE Rulebook, Guidance Note on artificial intelligence and machine learning, 11 February 2026, rulebook.centralbank.ae.
- Monetary Authority of Singapore, "Guidelines on Provision of Digital Payment Token Services to the Public [PS-G02]", published 17 January 2022, mas.gov.sg.