OPINION

Is XBRL tagging for sustainability reports "archaic nonsense" — or a chance for digital reporting and AI? If you are an NLP researcher and want to discuss this with us, join NLP4Climate on 12 August.

Is XBRL Tagging for Sustainability Reports "Archaic Nonsense" — or a Chance for Digital Reporting and AI?

· 4 min read

Is XBRL Tagging for Sustainability Reports “Archaic Nonsense” — or a Chance for Digital Reporting and AI?

When we speak with CEOs of sustainability startups, a familiar claim comes up: mandatory XBRL tagging for sustainability reports — machine-readable digital reporting under ESRS / CSRD — is “archaic nonsense”1, because AI pipelines can pull the same information out of PDFs. The same argument shows up in public LinkedIn debate: tagging as “nineties” tooling that AI has made unnecessary2.

It is an attractive story. AI tools are improving quickly, and they do have real potential for analysis of financial and sustainability disclosures. The harder question is whether that potential already justifies treating structured, tagged source data as optional in EU digital reporting.

From an engineering perspective, those are different jobs. Analysis can sit on top of messy documents. Comparability, audit trails, low-cost reuse by civil society and academia, and reproducible benchmarks work much better when sustainability data is published in a machine-readable form at the source. Claiming that AI already makes XBRL digital tagging obsolete is, in our view, overstated — at least on the evidence we see so far.

What we know about accuracy so far

  • ChatReport (UZH / ETH, 2023): about 83% on 11 broad TCFD questions against annotator consensus — useful, but not a substitute for tagged source fields, and “correct” already excludes cases where human analysts disagree.
  • Climate Finance Bench (2025): best RAG configuration answered about 62% of 330 report questions correctly, with roughly another 10% incomplete — so a sizeable share remains wrong or unsupported.
  • Forster et al., Nature Communications (2026): strong average agreement (β = 0.885, adj. R² 0.91–0.93) — but Supplementary Fig. S2 Spearman correlations vs Refinitiv span ~0.17–0.90 (median ~0.78; GHG reduction % at the bottom with n = 40), and Table S5 sMAE ranges from ~0 on several emissions metrics to >0.8 on e.g. female top-management share.
  • MSCI Institute (2025): once set up, XBRL extraction from India’s digital sustainability filings was up to 10× faster than AI PDF extraction, at a fraction of the cost, with higher reliability for tagged fields.

We prefer a different framing: mandatory XBRL for sustainability reporting is a chance for sustainability and for AI. Structured digital disclosures make sustainability information usable beyond a few organisations that can afford PDF extraction pipelines. And years of tagged reports would be exactly the kind of high-quality substrate the field needs to train and evaluate better models — rather than treating PDF archaeology as the permanent default.

Relevant work we point to includes ChatReport, Climate Finance Bench, Forster et al. on LLM ESG extraction in Europe, Benchmarking the Benchmarks, our AI Benchmark for Sustainability Report Analysis, and our summary of the MSCI Institute’s When old tech beats new tech analysis.

We are collecting research and practitioner evidence through our Call for Evidence on mandatory XBRL for ESRS vs AI claims. The sustainability-report benchmark is one concrete activity that can test extraction quality and limits rather than assert them.

A fuller academic treatment of the evidence is in preparation for journal publication. This post is the short version: the claim is in circulation, it matters for EU digital sustainability reporting, and we want the NLP community in the conversation.

Discuss this at NLP4Climate — 12 August

Is XBRL tagging for sustainability reports “archaic nonsense” — or a chance for digital reporting and AI?

If you are an NLP researcher and want to discuss this with us, join NLP4Climate on 12 August. You can also submit evidence via the project form.

Footnotes

  1. Public LinkedIn comment arguing that digital tagging is “archaic, time wasting nonsense” because AI and bots can scrape ESRS reports without tags — comment on Donato Calace’s post on EFRAG’s VSME iXBRL template. ↩

  2. Jérôme Cloetens (CEO, Palau), LinkedIn post arguing that XBRL “is the nineties,” that manual tagging is slow specialist work, and that AI can now query sustainability reports without tags — “I told our community call last week that XBRL is the nineties.” (12 March 2026). ↩

Share: