CSRD reporting automation after Omnibus I

A lot of companies spent 2024 and 2025 building CSRD reporting pipelines for a regime that no longer exists in that form. The Omnibus I package, published in the Official Journal on 26 February 2026 and in force since March 2026, cut the directive’s scope back to the largest undertakings: EU companies with more than 1,000 employees and net annual turnover above €450 million, with the non-EU threshold likewise raised to €450 million generated in the EU. Member States have until 19 March 2027 to transpose it.

The standards moved too. On 3 July 2026 the Commission adopted the revised ESRS delegated act, which it describes as shorter and clearer, cutting mandatory datapoints by more than 60% and total datapoints by more than 70%. EFRAG’s work on the amendments describes a 61% reduction in mandatory datapoints, with voluntary datapoints removed from the mandatory standards and more weight placed on materiality and judgement. Companies apply the revised standards from financial year 2027.

So the 2026 automation question is no longer how to build a collection engine for a thousand datapoints across four reporting waves. It is how to stabilise first-wave reporting under the current framework, shrink the data architecture to match the narrowed regime, and get through a standards transition without collecting everything twice.

Who still reports, and when

The timeline now splits into two populations. First-wave companies, reporting since FY2024, keep reporting, with the Quick Fix delegated act extending reliefs through reports published in 2026 and 2027: all first-wave companies may omit the anticipated financial effects disclosure, and those with fewer than 750 employees may additionally omit Scope 3 and total GHG emissions, ESRS E4 on biodiversity, and the S1–S4 social disclosures. These companies then migrate to the revised ESRS for FY2027 reporting.

Everyone else moved. The stop-the-clock mechanism delays the later waves by two years; one implementation summary puts wave 2 companies, who would have published in 2026 for FY2025, at a first report in 2028 for FY2027. Combined with the higher thresholds, many companies that were mid-build have left scope entirely.

The planning error we keep seeing is teams executing 2024-era project plans on autopilot. Before any tooling decision, re-derive your own position from the amended thresholds and dates. A wave 2 company still in scope has gained two years and should spend them on data quality, not on rushing a report nobody will file. A company out of scope should decide deliberately which parts of the pipeline earn their keep for other consumers.

Shrink the pipeline, keep the spine

The revised ESRS support a much leaner data model, and the temptation is to celebrate and delete half the architecture. The better move is to keep the structural spine and cut the datapoint sprawl. The design pattern that survives the transition has five layers:

• A source registry for entities, sites, and suppliers, recording where each raw figure originates and who owns it. This layer is scope-independent and is the part worth keeping even if your reporting obligation vanished.
• A materiality gate ahead of disclosure mapping. With the revised standards leaning harder on materiality and judgement, the assessment is now the main filter on what you collect at all, so it belongs in the pipeline as an explicit, versioned artefact rather than a PDF from a workshop.
• An ESRS taxonomy layer that maps each collected item to legacy Set 1 and to the revised standards, with effective dates. This is what makes dual-state reporting possible across the FY2026/FY2027 boundary.
• An evidence repository holding source documents, system extracts, and calculation methodology per KPI, linked from every reported figure.
• A versioning layer tracking definition changes, estimate revisions, and restatements over time.

The operating principle for 2026: controlled, high-confidence collection of the datapoints that survive, instead of broad collect-everything programmes feeding disclosures that are about to become voluntary or disappear. For multinational groups the narrowed scope does not remove the consolidation problem; the entities that remain in scope still need subsidiary data rolled up consistently, so the group-level identifiers and mappings stay load-bearing.

Where AI drafting fits

Nothing in the amended framework prohibits AI-drafted disclosures. The constraint is structural: the report must satisfy ESRS content, materiality, and evidence requirements, and an assurance provider will trace figures and claims back to sources. That places AI naturally on the drafting and structuring side of the line. The uses that hold up in practice are converting structured KPI outputs into narrative drafts, aligning wording across sections, checking consistency between claims, metrics, and policy statements, and generating first-pass cross-references and tag suggestions.

Keep AI out of final sign-off. The workflow should include human review, source traceability, and red-flag checks before anything is published, because assurance-ready reporting depends on auditable evidence rather than fluent text. The practical control is an AI trace log: which disclosures were machine-drafted, which source data fed them, and who approved the final language. It costs little to maintain and answers the exact question an assurance provider will ask.

Assurance is where the cost actually lands

The amended framework adjusts assurance requirements, but reporting remains assurance-relevant corporate reporting, and this is where automation programmes are won or lost operationally. A report you can publish is not the same as a report you can defend, and the gap between the two is the evidence pack: source documents and extracts, calculation methodology for each KPI, the materiality assessment files, change logs for estimates and restatements, and approval evidence from management.

Because the revised ESRS are narrower and more principle-based, expect assurance attention to shift from datapoint volume toward how sound your processes are, how materiality judgements were made, and whether every figure traces cleanly to source. That changes what you automate. The high-value automation is no longer form-filling across hundreds of datapoints; it is lineage capture, evidence linking, and change tracking, the unglamorous plumbing that turns a review from an archaeology project into a lookup. Budget accordingly: teams that spent 80% of their effort on collection UIs and 20% on evidence are inverting that ratio for FY2027, and the ones that built the taxonomy and versioning layers early are the only ones finding the transition boring, which is what a well-run transition should be.