Listed companies must publish annual financial reports in machine-readable ESEF/iXBRL format, and banks and insurers must submit hundreds of supervisory templates per period. Both require mapping every number to a taxonomy element and keeping text and figures consistent. AI suggests tags, flags inconsistencies, pre-fills templates and drafts replies to supervisors, while validation rules and accountable humans retain control of the filing.
Problem & value
- Today: finance teams finalise the report in Word/InDesign, then an XBRL specialist or service provider tags thousands of facts, often in a compressed pre-filing window. Banks map ledger accounts to COREP/FINREP cells with spreadsheets and hard-coded rules; every taxonomy release means rework.
- AI value: (1) suggest taxonomy elements for new or changed disclosures from label, context and prior-year tagging; (2) detect mismatches between numbers in tables, notes and management report; (3) explain failed validation rules; (4) pre-fill templates and draft explanations of variances for supervisors; (5) shorten the closing-to-filing window.
- Why it matters: ESEF tagging is mandatory for issuers on EU regulated markets, and auditors give an opinion on compliance of the digital report format, so errors have audit and enforcement consequences [1][2]. Published vendor claims on time savings exist but are marketing figures; measure your own baseline.
Variants
- Financial report drafting & ESEF/iXBRL tagging: drafting notes and management report from ledger data and prior-year text, checking figures against text, then tagging and building the iXBRL package (core of this use case).
- Bank prudential and financial reporting: COREP (own funds, credit risk, large exposures) and FINREP, plus AnaCredit-style granular credit data, via EBA Data Point Model (DPM) [3].
- Insurance: Solvency II quantitative reporting templates (QRT) and EIOPA taxonomy [4].
- Consolidated groups: mapping many subsidiary charts of accounts to one group reporting taxonomy.
- Supervisor Q&A: drafting replies to supervisory queries and plausibility-check rejections.
How it works
- ledger / consolidation + draft report
- mapping layer (account
- taxonomy element, learned from prior year)
- LLM / classifier tagging suggestions for tables and narrative facts (with confidence)
- human review in tagging tool (accept / correct, extension elements only when needed)
- deterministic validation (XBRL formula, ESEF rules, EBA/EIOPA filing rules)
- LLM consistency check (numbers in text vs tables, sign, scale, period, rounding)
- package build (iXBRL / XBRL-CSV)
- sign-off by CFO / board
- submission to OAM / supervisor
- Tagging suggestion: embedding retrieval over taxonomy labels, definitions and last year's tags, re-ranked by an LLM; a fine-tuned classifier suffices for stable table structures.
- Consistency check: LLM extracts numeric claims from prose ("revenue rose 4.2% to EUR 1.3 bn") and a script recomputes them from tagged facts; the LLM only flags, arithmetic is deterministic.
- Validation is never delegated to the model: XBRL formulas and supervisory validation rules are authoritative; the LLM explains failures and proposes fixes.
- Human-in-the-loop: every tag change is reviewed; filing approval and management representation stay with the responsible officers.
Data preparation
Needed are the general ledger and trial balance, sub-ledgers, consolidation data, the previous year's reports, the chart-of-accounts mapping and the official taxonomies (ESEF/IFRS, EBA, EIOPA). They live in the ERP and consolidation tool, in Excel closing workbooks, in Word or PDF annual reports on file shares, and in the regulator's templates. Typical gaps: account mappings kept only in a controller's head or spreadsheet, manual adjustments outside the ledger, inconsistent entity and currency treatment, report text and figures maintained in different files, and no version history of past taxonomy mappings.
Steps: (1) freeze a reconciled data set per period with the trial balance as single source (PREP-005); (2) document and clean the account-to-taxonomy mapping as a governed table (PREP-012, PREP-011); (3) convert prior-year PDFs and Word reports into structured text (PREP-007, DOC-018); (4) collect prior-year mappings and filings as gold examples (PREP-016, PREP-020); (5) fix ownership of each source (PREP-004). Effort is moderate, about 30 to 40 percent, as the data is mostly structured. Check with GUIDE-003.
Implementation paths
Which path for whom
- Enterprise: Path A, because reporting suites for ESEF tagging and supervisory templates (COREP, FINREP, QRT) are certified, follow taxonomy updates and are a commodity function with strong audit expectations.
- Developer: Path C, because vendors and integrators building reporting tooling benefit from open taxonomies, Arelle and open-weight models they can run and audit themselves.
Path A — Buy (recommended)
- Disclosure/XBRL platforms: Workiva, Toppan Merrill, DFIN; EU/DACH specialists: AMANA (DE), Lucanet (DE), CCH Tagetik (Wolters Kluwer), Iris Business (UK). Ask which ship AI tagging assistants.
- Prudential reporting: Regnology (DE), Wolters Kluwer OneSumX, Moody's, AxiomSL/Adenza (Nasdaq), Vermeg (regulatory suites), plus national tools of savings/cooperative bank IT providers (e.g. Finanz Informatik, Atruvia) in Germany.
- Choose A when: statutory filings are mission-critical and need taxonomy updates, audit trail and vendor-maintained validation. AI is typically an add-on assistant inside these tools.
Path B — Build on cloud APIs
- Frontier LLMs with structured output (Claude, Gemini, GPT, Mistral — EU regions or EU providers such as Mistral for sensitive pre-release financials) for mapping suggestions, narrative checks and supervisor-reply drafts; embeddings for taxonomy retrieval.
- Open-source XBRL tooling for validation and building: Arelle (validation, iXBRL viewer), the ESMA-endorsed conformance suite, and the Arelle ESEF plugins; use them as deterministic gates.
- Orchestrate with workflow tools; keep LLM output as suggestions written into the tagging tool.
Path C — Self-host (open source)
- Models: open-weight LLMs (Qwen3, Llama 4, Mistral Large/Mixtral-class, gpt-oss) behind vLLM for unpublished financials that must not leave the perimeter; embeddings such as BGE or multilingual E5; Arelle for validation. Pre-publication figures are inside-information, so on-prem is attractive.
- Hardware: HW3 for 70B-class models; HW2 enough for a classifier plus ~30B quantised LLM.
Path D — Custom training
Fine-tune a classifier on your own multi-year tagging history (tens of thousands of tagged facts) and on supervisory template mappings. Worth it for large issuers, tagging service providers and banks with many legal entities. Labels come directly from approved filings.
State of the art (2026)
- ESEF reporting requires iXBRL tagging of primary statements in IFRS consolidated reports, block- tagging of notes, and a taxonomy updated annually by ESMA; the legal basis is Delegated Regulation (EU) 2019/815 (RTS on ESEF) [1][2]. Check the current ESMA Reporting Manual for the taxonomy year in force.
- Research on LLMs for XBRL tagging shows that tagging with large taxonomies is hard: models do well on common elements but fail on rare elements and extension decisions; benchmark work such as FiNER-139 (financial numeric entity recognition with XBRL tags) is the standard reference [5].
- EBA supervisory reporting uses the Data Point Model with XBRL-CSV transmission and a regularly updated reporting framework; fixed templates make deterministic mapping dominant and AI mainly assists with mapping maintenance, reconciliation and explanations [3].
- Fast-moving: single-submission points (EU ESAP, European Single Access Point) will centralise public financial and sustainability data, increasing demand for clean machine-readable reports [6].
- Limits: hallucinated element names, wrong sign/scale/period contexts, silent drift when taxonomies change, weak handling of company-specific extensions; no vendor claims should be accepted without a back-test on your prior-year filing.
Shared resources & combinations
- Same stack: DATA-002 (report narratives), DATA-012 (CSRD/ESRS tagging uses the same digital reporting format and tooling), DATA-024 (FP&A commentary), DOC-012 (filing analysis from the consumer side), LANG-020 (horizon scanning for new taxonomy/rule changes).
- Upstream/downstream: ANOM-006 (ledger anomaly checks before closing), AGENT-006 (close and reconciliation). One on-prem LLM server and one taxonomy vector index can serve all of these.
Effort, cost & KPIs
- Path A: weeks to a first filing season for ESEF; months for prudential suites with data model integration. Cost drivers: licences per entity/template set, taxonomy updates, integration.
- Path B/C: 3–6 months for a tagging assistant that integrates with an existing tool; ongoing evaluation against each year's filing.
- KPIs: share of facts auto-suggested and accepted unchanged, tagging errors found by validation vs. found by auditor, hours from final numbers to filing, number of extension elements, consistency findings per report, supervisory resubmissions.
People & change
Involve the CFO or management as sponsor, the head of financial reporting as process owner and the reporting specialists as end users. Legal and compliance (and, in regulated entities, the supervisory contact) confirm that tooling meets filing duties; the auditor should hear early how AI-assisted tagging is controlled. IT and information security handle access to closing data, which is highly confidential; data owners of the ledger and consolidation confirm the mappings.
Works council: usually no, because the system processes financial figures, not employee behaviour. §90(1) no. 3 BetrVG still requires timely information if reporting workflows change with AI; check §87(1) no. 6 if the tool logs individual user activity. The DPO has a small role, as personal data is largely absent (remuneration disclosures are the exception); sign an AVV (Art. 28 GDPR) with cloud vendors.
Keep humans in control: every AI-proposed tag, mapping and consistency finding is approved by a qualified accountant, and the filing remains the company's responsibility. Pilot on a past period or the interim report, compare against the filed version, and go live only when differences are explained. Typical concern: "an AI error in a filing". Answer with the four-eyes principle, a retained audit trail and parallel runs. Plan training for the reporting team (GUIDE-006) and a controlled pilot (GUIDE-009); GUIDE-011 covers monitoring after go-live.
Risks, regulation & compliance
- EU AI Act: minimal risk — internal reporting support is not an Annex III use case (credit scoring or insurance pricing are different use cases). AI literacy (Art. 4) still applies.
- Filing responsibility: the issuer and management remain responsible; auditors assess ESEF compliance. Keep an audit log of AI suggestions versus final tags.
- Confidentiality and market abuse: unpublished results are inside information (MAR); use EU-hosted or self-hosted models with strict access controls and no training on inputs.
- GDPR: mostly aggregate data; remuneration reports, related-party notes and loan-level data (AnaCredit) can contain personal data — apply data minimisation.
- Operational resilience: DORA applies to financial entities for ICT providers including AI vendors [7].
- Mitigations: deterministic validation gates, four-eyes review, back-testing, version pinning of taxonomies, no auto-submission.
Getting started
- Inventory your obligations (ESEF, EBA/EIOPA templates, national) and current tagging workflow.
- Re-tag last year's report with an AI suggestion tool and compare to the filed version.
- Add an automated figure-to-text check on the draft report before tagging starts.
- Wire Arelle or vendor validation as a mandatory gate and log all suggestions.
- Plan taxonomy-update regression tests for each new release.
Sources
- ESMA — European Single Electronic Format (ESEF) — https://www.esma.europa.eu/policy-activities/corporate-disclosure/european-single-electronic-format
- Commission Delegated Regulation (EU) 2019/815 (ESEF RTS) — https://eur-lex.europa.eu/eli/reg_del/2019/815/oj
- EBA — Reporting frameworks — https://www.eba.europa.eu/risk-and-data-analysis/reporting/reporting-frameworks
- EIOPA — Solvency II reporting and taxonomy — https://www.eiopa.europa.eu/tools-and-data/supervisory-reporting-dpm-and-xbrl_en
- Loukas et al. — FiNER: Financial Numeric Entity Recognition for XBRL Tagging (ACL 2022) — https://arxiv.org/abs/2203.06482
- Regulation (EU) 2023/2859 establishing the European Single Access Point (ESAP) — https://eur-lex.europa.eu/eli/reg/2023/2859/oj
- Regulation (EU) 2022/2554 (DORA) — https://eur-lex.europa.eu/eli/reg/2022/2554/oj
- Arelle open-source XBRL platform — https://arelle.org/
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj