CSSF AI Questionnaire: eDesk Window Closes 31 December 2026
On 17 November 2026 the CSSF opened a new eDesk procedure for its questionnaire on the use of artificial intelligence, run as a joint initiative with the Banque centrale du Luxembourg (BCL). The CSSF AI questionnaire is addressed to banks, investment firms, payment institutions, electronic money institutions, specialised PFS, support PFS, crypto-asset service providers and authorised investment fund managers. According to the CSSF communique, it must be completed and submitted through the form in the dedicated eDesk procedure, with submission open until 31 December 2026.
The communique publishes no question set: the form and the user guide on submission arrangements sit inside the eDesk procedure, so an entity sees what it has to answer only once one of its linked eDesk users opens it. For specialised PFS, support PFS and crypto-asset service providers, this is the first time the joint survey has reached them.
This is the third CSSF and BCL survey of AI use in the Luxembourg financial sector. The first ran from October 2021 to January 2022 and led to a thematic review in May 2023; the second ran from June to August 2024 and led to a second thematic review on 16 May 2025.
Related reading: CSSF AI Communique: Mapping Frontier Cyber Risk to DORA
CSSF AI questionnaire dates and channel, compared with 2024
| Item | 2024 survey | 2026 questionnaire |
|---|---|---|
| Window | 19 June to 7 August 2024 | 17 November to 31 December 2026 |
| Entity types addressed | Banks, investment firms, payment institutions, e-money institutions, authorised investment fund managers | The same five, plus specialised PFS, support PFS and crypto-asset service providers |
| Submission routes named | Form in the eDesk procedure “Specific Enquiries”, or an API solution using the S3 protocol | Form in the dedicated eDesk procedure only |
| Contact named | No dedicated contact specified in the 19 June 2024 communique | edesk@cssf.lu |
The 2026 communique mentions the form alone, so a team that scripted its 2024 upload through the S3 API should confirm the route in the user guide first.
The deadline wording leaves one gap: the communique says “until 31 December 2026” and gives no cut-off time.
Three entity types join the panel
The survey population has widened with each round. The 2021 to 2022 survey covered credit institutions, payment institutions and e-money institutions. The 2024 survey added investment firms and authorised investment fund managers, which more than tripled the panel: 537 institutions were targeted and 461 answered, an 86% participation rate. The 2026 communique adds specialised PFS, support PFS and crypto-asset service providers.
One term needs care. The 2026 communique does not define “authorised investment fund managers”. In the 2024 round, the thematic report scoped that term to management companies subject to Chapter 15 of the Law of 17 December 2010 and authorised AIFMs under the Law of 12 July 2013. AIFMs that are only registered fall outside that wording. Whether the 2026 round uses the same perimeter is something the user guide or edesk@cssf.lu can confirm.
An entity with no AI in use is still an addressee. The 2024 panel was mostly made up of firms that did not run AI in production: 28% of respondents used AI in production or development, and 22% were experimenting or planning to within 12 months. The general sections applied to all respondents, and the report’s figure that 36% deny staff access to public generative AI tools covers AI users and non-users alike. The 2026 communique sets out no exemption for entities reporting no AI use, so those entities should inspect the eDesk form and submit the information requested for their entity; the public notice does not establish whether that submission will be short.
What the 2024 form asked, and what the 2026 communique leaves open
The 2026 communique does not describe the questions. The best public indication is the structure of the 2024 online questionnaire, which the May 2025 thematic report sets out in four sections:
- General information: company size, IT team size, IT outsourcing, use of freely available generative AI tools, any generative AI policy and the overall status of AI adoption.
- Digital strategy: current and future investment in AI and distributed ledger technology, including tokenisation and crypto-assets, and the expected revenue or cost effects.
- AI questionnaire: benefits and challenges, organisation, data and governance, security and robustness, the machine-learning development lifecycle, technical infrastructure and generative AI usage methods.
- AI use cases: one entry per use case through dedicated tabs, covering development aspects and trustworthiness.
Two details from 2024 shape the preparation work. Use-case submission was optional and not every detailed question within a use case was required, so the report’s use-case analysis rests only on the answers received. The questionnaire also separated local from group activity: 46% of respondents reported investments in innovative technologies (AI and DLT) at group level only, and 55% of the entities using or experimenting with AI had their data science team at group level.
If the 2026 form keeps that structure, the internal inputs are an AI use-case inventory with status (production, development, experimentation), technology and internal or client-facing use, the policy on public generative AI tools, and a local versus group split of teams and spending. That is my reading of a repeat of the 2024 questions; the eDesk form decides what 2026 actually asks. Our piece on AI model risk in Luxembourg prudential reporting sets out AI model inventory fields that can feed those answers.
The AI Act classification answers the authorities questioned
The 2024 survey asked respondents to place each use case in an AI Act risk level: unacceptable, high, limited (transparency), minimal or no risk, or not yet classified. Of all use cases reported, 5% were classified high risk, 16% limited risk, 50% minimal or no risk, and 29% were not yet classified. None was classified as unacceptable.
The CSSF and BCL then checked those answers against Regulation (EU) 2024/1689, the AI Act, and found that they reflected the entity’s perception of business risk more than the legal classification. Credit scoring, one of the few financial use cases listed as high risk, was classified high risk in only three of six cases.
The report records high-risk labels assigned by respondents to two ‘AML/fraud detection’ cases, one IRB credit risk model and two other credit risk models. It refers to Recital 58, which discusses AI systems provided for by Union law for detecting fraud in the offering of financial services and for calculating credit institutions’ and insurance undertakings’ capital requirements for prudential purposes. Annex III, point 5(b), expressly excludes financial-fraud detection from its natural-person creditworthiness category. Whether a particular AML, fraud-detection or credit-risk system is high risk depends on its intended purpose, assessed against the Article 6 and Annex III criteria. Use cases for cyber security, process automation and chatbots also received high-risk labels although the AI Act does not list them as high risk.
Annex III, point 5(b) covers AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score. A model that scores corporate borrowers falls outside that point. If the 2026 form repeats the classification question, my working assumption is that the answer should come from the same Article 6 and Annex III analysis the entity runs for its own AI Act inventory, which our EU AI Act compliance guide for financial institutions walks through.
Opening the procedure in eDesk
The form is only reachable by an authenticated eDesk user linked to the entity. The CSSF’s eDesk user guides set out the chain. Every user needs a valid LuxTrust certificate and an eDesk account. The account then has to be linked to the entity through a “New entity link request”, which an advanced user of that entity accepts or rejects. Each entity using eDesk needs at least one advanced user, and a candidate becomes one by sending a “New advanced user request” with a mandate and supporting documents, which the CSSF verifies before accepting.
That last step is outside the entity’s control. A support PFS or crypto-asset service provider that has rarely used eDesk should check now whether its advanced user is still active. Crypto-asset service providers are also among the entities covered by our note on CSSF prudential reporting for PIs, EMIs and CASPs moving to eDesk.
Some eDesk procedures are restricted to specific roles that the advanced user assigns. The CSSF remuneration reporting guide, for example, restricts its online procedures to users holding the “Human resources responsible” role. The AI questionnaire communique names no role, so whether one applies will show once the advanced user opens the procedure list.
Where the answers go
In the 2024 round, responses were aggregated, anonymised and analysed to produce the thematic report, which compared results with the first survey only after filtering to credit institutions, payment institutions and e-money institutions. The 2026 communique states no publication plan and no date for any report. It also cites no legal provision as the basis for the request; its operative wording is that the questionnaire must be completed and submitted.
Frequently Asked Questions
Can group IT or an outside provider fill in the form for the Luxembourg entity?
eDesk accounts are linked to the entity or entities the user works for, and the entity’s advanced user approves each link. The communique says nothing about delegation, so the user guide inside the procedure is the place to check.
Does submitting the questionnaire count toward AI Act or DORA obligations?
The communique links the questionnaire to no other instrument. The AI Act high-risk rules for systems under Article 6(2) and Annex III apply from 2 December 2027 under Article 113(c)(i), as amended by Regulation (EU) 2026/1744; assessing classifications before that date is preparatory. Those rules and the DORA register of information for ICT third-party arrangements, required by Article 28(3) of Regulation (EU) 2022/2554 and maintained using the standard templates in Commission Implementing Regulation (EU) 2024/2956, are separate obligations with their own addressees and formats. My reading is that the questionnaire answers should be consistent with those records, without standing in for them.
Will the 2026 answers be compared with what our entity said in 2024?
The published 2025 report compared results only in aggregate, by entity type. It contains no entity-level comparison, and the 2026 communique does not say how individual answers will be used.
Related Articles
- EU AI Act Compliance for European Financial Institutions: risk tiers, provider and deployer duties and incident reporting under the AI Act.
- AI Model Risk in Prudential Reporting: what Luxembourg banks should document for AI components that feed prudential returns.
- CSSF Prudential Reporting for PIs, EMIs and CASPs: The Move to eDesk: the timetable for moving these prudential returns onto eDesk.
- CSSF AI Communique: Mapping Frontier Cyber Risk to DORA: what the CSSF’s July 2026 AI communique expects DORA ICT risk management to evidence.
Key Takeaways
- Check this week that the entity has an active eDesk advanced user; a new advanced-user request needs CSSF verification.
- Have that advanced user open the procedure and its user guide before assigning work, since the question set is not public.
- Specialised PFS, support PFS and crypto-asset service providers have no earlier round of answers to reuse.
- Label group-level AI teams, tools and spending as group-level.
- Take any AI Act risk level from the Article 6 and Annex III analysis, not from perceived business risk.
Sources and References
- CSSF, communique “New eDesk procedure – Questionnaire on Artificial Intelligence”, 17 November 2026: cssf.lu
- CSSF, communique “New eDesk procedure – Artificial Intelligence Survey”, 19 June 2024: cssf.lu
- CSSF, press release “Second thematic review on the use of Artificial Intelligence in the Luxembourg financial sector”, 16 May 2025: cssf.lu
- CSSF and BCL, “Thematic review on the use of Artificial Intelligence in the Luxembourg financial sector” (second edition, May 2025), executive summary, chapters 2 to 4 and 7, and the AI Act classification analysis on pages 41 to 43 (PDF): bcl.lu
- CSSF, “Thematic review on the use of Artificial Intelligence in the Luxembourg Financial sector”, 3 May 2023: cssf.lu
- Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744; consolidated text of 27 July 2026 (Articles 6 and 113; Annex III, point 5(b)); original Act (Recital 58): EUR-Lex
- CSSF, “User Guide for XBRL reports – Bank prudential reporting”, version 1.3, section 3.2.1 (eDesk prerequisites) (PDF): cssf.lu
- CSSF, “User Guide for remuneration reporting”, version of 29 May 2026, section 4.2.1 (eDesk prerequisites and roles) (PDF): cssf.lu
- Regulation (EU) 2022/2554, Article 28(3): EUR-Lex
- Commission Implementing Regulation (EU) 2024/2956: EUR-Lex
Filing the AI questionnaire before 31 December
The work in the next six weeks is mostly access and assembly. Confirm the advanced user, open the “Specific enquiries” procedure for the Artificial Intelligence Questionnaire, read the user guide, and route the use-case inventory, the generative AI policy and the local versus group split to the person who will key the answers. Set an internal sign-off date that leaves room before submission closes on 31 December 2026.
Disclaimer: The information on RegReportingDesk.com is for educational and informational purposes only. It does not constitute legal, regulatory, tax, or compliance advice. Always consult your compliance officer, legal counsel, or the relevant supervisory authority for guidance specific to your institution.
