The regulatory perimeter for AI in asset management

The Act sorts AI into four buckets: prohibited, high-risk, limited-risk (transparency obligations only), and minimal-risk.

Most asset management AI lands in minimal or limited risk. That is the key nuance many people miss. Portfolio optimization, factor models, and trade-signal generation are generally not automatically high-risk under the Act's Annex III list.

Where it gets sharper:

• Creditworthiness assessment of natural persons is explicitly high-risk. If your platform assesses a retail client's credit to offer a margin lending product, that component is high-risk.
• Prohibited practices include manipulative or subliminal techniques and certain exploitation of vulnerabilities. A "nudge engine" that exploits a client's behavioral biases to increase trading volume is in dangerous territory.
• General-purpose AI (GPAI) models, meaning large models like the LLMs behind a research-summarization tool, carry their own transparency and documentation duties on the provider.

Practical takeaway: a quant equity model is usually low-risk under the Act, but a retail credit-scoring or client-vulnerability component can pull a product into high-risk. Classify component by component, not app by app.

The European Commission's official summary is a clean starting reference: EU AI Act overview.

What "high-risk" actually requires

If a component lands in the high-risk tier, expect: a risk management system, data governancedata governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.View full definition → (documented training data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.View full definition → and bias testing), technical documentation, logging, human oversight, and accuracy and robustness testing. This is essentially formalized model risk management. If you already run a solid model validation function, you are most of the way there.

MiFID II: suitability is where advice models bite

MiFID II's suitability obligation requires that any personal recommendation or discretionary portfolio decision match the client's knowledge, experience, financial situation, and objectives (including risk tolerance and, increasingly, sustainability preferences).

An AI advice engine does not lower this bar. It raises the documentation burden.

Concrete failure modes:

• A model recommends a complex structured product to a client whose questionnaire flagged low product knowledge. Suitability breach.
• The model drifts: markets change, and its risk-tolerance mapping no longer reflects the questionnaire logic that was approved. Undetected drift is a suitability breach waiting to surface.
• The firm cannot explain why a client received a given recommendation. MiFID II expects a suitability report. A black-box "the model said so" does not satisfy it.

The European Securities and Markets Authority (ESMA) has issued guidance emphasizing that firms using AI in investment services remain fully responsible under MiFID II, with senior management accountable. AI is a tool, not a defense.

The SEC: conflicts inside predictive models

The US route is different. The SEC in 2023 proposed rules on the use of predictive data analytics (PDA) by broker-dealers and investment advisers. The core concern: when a firm uses AI to optimize for its own revenue (steering clients toward higher-fee products or more trading), that is a conflict of interest, and the firm must eliminate or neutralize it, not just disclose it.

As of early 2026, the specifics of that rulemaking have been contested and reshaped, so treat the exact final text as unsettled. What is durable is the SEC's underlying position: existing fiduciary duty and anti-fraud rules already apply to AI-driven advice. A registered investment adviser owes a duty of care and loyalty regardless of whether a human or a model produces the recommendation.

The SEC has also warned publicly about "AI washing": overstating AI capabilities in marketing. Claiming your fund uses "proprietary AI" when it runs a simple rules engine is a disclosure problem, and the SEC has brought enforcement actions on this theme.

A worked classification

Take a hybrid robo-advisor serving EU retail clients and US clients.

| Component | EU AI Act tier | MiFID II | SEC |

|---|---|---|---|

| Risk-tolerance questionnaire scoring | Limited/minimal | Suitability core | Duty of care |

| Portfolio optimizer | Minimal | Indirect (via suitability) | Fiduciary |

| Product recommendation engine | Limited (transparency) | Suitability + report | Conflicts, PDA concern |

| Margin credit check on client | High-risk | N/A | Fiduciary |

| LLMLLMA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → research summarizer | GPAI provider duties | N/A | AI washing risk |

One product, five different regulatory footprints. This is why "is our AI compliant?" is the wrong question. The right question is "which component, which regime, which tier?"

Guardrails to run before deployment

Regulators converge on the same practical controls. Build these before you ship, not after.

1. Inventory and classify

Maintain a model inventory. For each AI component, record: purpose, EU AI Act tier, MiFID II touchpoint, data sources, and owner. Regulators increasingly expect this on demand.

2. Human oversight that is real

"Human in the loop" must mean a human who can understand and override, not a rubber stamp. For high-risk components the Act requires it explicitly. Define the override authority and log when it is used.

3. Explainability proportionate to impact

A suitability recommendation needs a client-facing rationale. A back-office trade-cost estimator does not. Match the explainability effort to the client harm potential.

4. Drift and bias monitoring

Set thresholds and alerts. A minimal logging pattern for a recommendation service:

# Log every recommendation for audit and drift detection
log_record = {
    "client_id": client.id,
    "risk_score": client.risk_score,      # from suitability questionnaire
    "recommended_product": rec.product_id,
    "product_risk_level": rec.risk_level,
    "model_version": model.version,
    "timestamp": now_utc(),
}
# Compliance flag: recommendation riskier than client tolerance
if rec.risk_level > client.risk_score:
    log_record["suitability_flag"] = True
audit_store.write(log_record)

That single suitability_flag is the difference between catching a MiFID II breach internally and explaining it to a regulator later.

5. Marketing review

Have compliance sign off on any external claim about AI. "AI washing" enforcement is a genuine, current risk in the US.

The direction of travel for 2026

The EU AI Act's high-risk obligations phase in through 2026 and 2027, so build documentation now. ESMA and national regulators are actively supervising AI in investment services. In the US, expect enforcement under existing fiduciary and anti-fraud rules to lead, with the exact PDA rulemaking still in flux. Firms operating across both jurisdictions should design to the stricter standard, which is usually the EU's.

Key Takeaways

• Classify by component, not by product. A quant model is usually low-risk under the EU AI Act, but a retail credit-scoring or client-vulnerability feature can pull the same product into high-risk.
• MiFID II suitability is the sharpest edge for advice AI. You must be able to explain and document why a client got a recommendation; "the model said so" is not a suitability report.
• The SEC applies existing fiduciary and anti-fraud rules to AI now, with a specific focus on model-embedded conflicts and on "AI washing" in marketing.
• The high-risk checklist is really model risk management: inventory, data governancedata governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.View full definition →, human oversight, logging, and drift monitoring. Build it before deployment.
• Operating across the US and EU? Design to the EU standard, which is generally stricter, and keep dated records because the rules are still phasing in through 2027.