
October 15, 2026. The Financial Stability Board (FSB) published the final version of its 12 sound practices for the responsible adoption of artificial intelligence by financial institutions.
The timing matters. AI no longer just writes. It assesses risks, monitors transactions and prepares analyses. Increasingly, it also acts directly inside business software.
The FSB draws a clear conclusion: the questions are now governance, traceability and control.
The report creates no obligation for French wealth advisors (CGPs, conseillers en gestion de patrimoine). It does, however, offer a useful framework for any professional who hands tasks over to AI. What does it contain, and what should an advisory firm take from it?
In short. On October 15, 2026, the FSB published twelve non-binding sound practices for AI adoption in finance: four on governance, eight on managing risks across the AI lifecycle. The final version strengthens documentation, traceability and the treatment of agentic AI. For a wealth advisor, it is a voluntary reference, useful for checking calculations, tracing recommendations, limiting AI access and choosing vendors.
Contents
Why is the FSB focusing on artificial intelligence?
Because AI can amplify existing risks and create new ones, both within individual institutions and across the financial system.
The FSB coordinates national financial authorities and standard-setting bodies at international level. It brings together authorities from 24 countries and jurisdictions and reports to the G20.
Its charter states that its work creates no legal rights or obligations.
The report opens with the benefits: fraud detection, compliance, credit analysis, productivity. It then distinguishes two levels of risk.
Level | Risks identified by the FSB |
|---|---|
Institution | Weak governance, unauthorized tools ("shadow AI"), poor data quality, lack of explainability, fabricated outputs, ineffective human oversight, cyber attacks, vendor dependency |
Financial system | Vendor concentration (cloud, models), correlated behavior when everyone uses the same models, cyber risks, model and data risk |
Source: FSB, final report, October 15, 2026.
The 12 practices target individual institutions. The FSB monitors systemic risks separately, and notes that ignoring AI carries its own risk: a weaker ability to detect fraud and cyber threats.
What do the FSB's 12 sound practices say?
They form two groups: four practices on organization-wide governance, and eight on managing risks across the AI lifecycle.
No. | Sound practice | In plain terms |
|---|---|---|
1 | Strategic direction and oversight | Leadership aligns AI with the business model, risk appetite and strategy. |
2 | Governance and accountability | Clearly defined roles and responsibilities. |
3 | AI risks in the risk framework, documentation | AI risks are built into existing frameworks, and every use is documented. |
4 | Organizational adaptability | Governance and skills evolve with the technology. |
5 | Materiality and risk assessment | Each use case is assessed at inception, then reassessed. |
6 | Selection process | Model choice reflects needs and risks. |
7 | Data governance | Data that is accurate, complete, consistent, reliable and secure. |
8 | Explainability and transparency | More explainable AI where feasible, compensating controls otherwise. |
9 | Performance management | Testing and ongoing monitoring, proportionate to risk. |
10 | Human oversight | Human control matched to risk, autonomy and complexity. |
11 | Cyber and ICT risk management | AI attack scenarios built into tests and exercises. |
12 | Third-party AI risk management | Managing vendor dependency: performance, data, concentration, continuity. |
Source: FSB, final report, October 15, 2026. Summary by Apana.
Three ideas run through the whole report: proportionality, traceability of every use, and accountability that stays with the institution, even when the tool comes from a vendor.
One case in the report illustrates this. At an asset manager, an AI tool produces an initial risk assessment of investment proposals, with each finding linked to the applicable rules and source documents.
The tool neither approves nor rejects anything. The decision stays with the risk manager.
What changed since the consultation
Put out for consultation on June 10, 2026, the draft received around 160 responses. The final version keeps the 12 practices but strengthens several points:
documentation (practice 3) must allow actions to be reconstructed, especially for agentic AI;
transparency is reinforced for high-stakes uses (practice 8);
independent validation and real-time monitoring are spelled out (practice 9);
vendor management now covers AI embedded in third-party software and exit strategies (practice 12);
considerations on agentic AI were added, which the FSB describes as still emerging.
The status is unchanged: this is not a binding international standard.
Why does agentic AI raise the stakes?
Because an agent no longer just produces an answer. It acts inside systems, sometimes over hundreds of steps, faster than a human can check.
Three levels of AI should be distinguished:
generating (writing, summarizing): the risk is an error in the text;
analyzing (extracting, comparing, spotting anomalies): the risk lies in the reliability of the result;
acting (connecting to software, chaining tasks, triggering operations): the risk becomes operational.
For this third level, the FSB lists specific risks: unauthorized or erroneous actions, data leaks, disruption to connected systems, and oversight that is hard to scale.
Among its recommendations:
give each agent a unique identifier and trace every action: what triggered it, which tools and data were used, whether permissions were respected;
strictly separate the agent's goals, permissions and rules (such as when to seek human approval) from the data it can access;
recognize that risks compound as an agent's access expands.
Take a typical case. An assistant retrieves a client's data, runs a retirement simulation, then prepares tasks in the CRM (client relationship management software).
Each step raises its own question. Is the assistant allowed to read this file? Can the calculation be verified? Are the prepared operations approved before they run? And six months from now, who will know what the agent did, and why?
What does this mean for wealth advisors?
No new obligation, but a very practical set of questions for any firm using AI.
The FSB addresses financial institutions. French wealth advisors remain subject to their own framework: MiFID II (Markets in Financial Instruments Directive), the IDD (Insurance Distribution Directive) and the GDPR.
For AI, the EU AI Act also applies, as we covered in what changes on August 2, 2026. The mapping below is our own analysis, not the FSB's text.
Question for the firm | FSB practices | Possible application (Apana's reading) |
|---|---|---|
How do you check a tax calculation produced with AI? | 8 and 9 | Run it through a deterministic engine with identified rules, and check the result before presenting it. |
How do you keep a record of a recommendation? | 3 | Keep the input data, assumptions, scenarios compared and the approved version. |
How do you protect client wealth data? | 7 and 11 | Know where data flows, limit what the tool can read, never paste a client file into a consumer chatbot. |
How do you supervise an AI agent? | 5 and 10 | Grant minimal access, and require human approval before any action on a client file or contract. |
How do you choose a vendor? | 6 and 12 | Ask where data is hosted, which model is used, how it changes and how to retrieve your data. |
One point bears directly on the duty to advise. According to the FSB, AI used in automated advice can produce recommendations unsuited to a client's profile when data is incomplete or outdated.
Why does architecture matter as much as the model?
Because the reliability of a critical task depends on the tools AI calls and the controls around it, more than on the model itself.
That is our reading of the report. The FSB puts it its own way: the most complex model is not always the most appropriate.
A general-purpose model excels at understanding a request and orchestrating tasks. For a tax calculation or a compliance document, it should rely on robust business systems.
Our article on training, fine-tuning and RAG explains why a sourced answer beats one produced from memory.
This is the approach we take at Apana:
Copilot captures and structures client interactions (meetings, documents, emails) into wealth data;
Advisor provides deterministic calculation engines, simulators and sourced scenarios;
Officer brings together CRM, compliance, account aggregation, document management and e-signature.
The model chooses which module to call. Calculations, data and records stay in the business tools. The advisor validates every step.
Which tasks can be entrusted to AI?
Those whose outcome can be checked, traced and owned.
The FSB report marks a shift in maturity. The question now is under what conditions business operations can be entrusted to AI.
For advisory firms, the advisor's role is changing. Less data entry. More supervision, judgment and explanation. And tools assessed as much on their architecture as on their intelligence.
Which firm will be able, tomorrow, to show a client exactly how AI contributed to their advice?
FAQ
Are the FSB's sound practices mandatory for wealth advisors? No. The FSB states that its 12 sound practices are not a binding international standard and create no rights or obligations. French wealth advisors remain subject to their own regulatory framework, including MiFID II, the IDD, the GDPR and the EU AI Act. The report serves as a voluntary reference.
What does the FSB recommend for agentic AI? The FSB recommends assessing each agent according to its autonomy and access, giving it a unique identifier, tracing each of its actions, and separating its goals and permissions from the data it can reach. It notes that these agent-specific practices are still emerging and may evolve.
What changed between the consultation and the final report? After around 160 responses, the final report of October 15, 2026 strengthens documentation and traceability, transparency for high-stakes uses, independent validation and vendor management. It also adds considerations on agentic AI. The 12 practices and their two groups are unchanged.
Sources
FSB, "FSB finalises sound practices for the responsible adoption of artificial intelligence (AI)", press release, October 15, 2026.
FSB, "Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Final report", final report, October 15, 2026.
FSB, "Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Overview of consultation responses", October 15, 2026.
FSB, "Sound Practices for Responsible Adoption of Artificial Intelligence (AI): Consultation report", consultation document, June 10, 2026.
Want to go further? Why is a general-purpose chatbot not enough to build advice? We explained it in why advisers shouldn't use ChatGPT to build advice.
Who did what in a client file prepared with AI? On one of your real cases, we show you where each piece of data goes, where each figure comes from and who validates each step.
