A sustainability report pulls in a dozen departments and months of calendar time, for a document read line by line by an assurance provider. Writing is not the bottleneck: methodological framing, data traceability and consistency across chapters are what consume the time.
These 30 AI prompts for CSRD cover ten stages of report production, from double materiality to answering an auditor's questions. Each one is written out in full, with what to provide and what it cannot do. They work with ChatGPT, Claude or Gemini, and complement our guide of 50 AI prompts for sustainability officers.
Read this before using these prompts. A language model does not know your company: not your sites, not your contracts, not your measurement gaps. Any generated content must be grounded in your actual data, reviewed by the person responsible for the topic, and linked to a supporting document. A CSRD report engages the collective responsibility of the members of the administrative, management and supervisory bodies (Article 33 of Directive 2013/34/EU) and is subject to external verification. No sentence produced by a tool should be published without documented human validation.
What the CSRD requires in 2026
The framework has changed. Directive (EU) 2026/470 of 24 February 2026, published in the Official Journal of the European Union on 26 February 2026, entered into force on the twentieth day after publication, that is 18 March 2026.
It reserves the obligation to publish sustainability information for companies whose net turnover exceeds EUR 450,000,000 and whose average number of employees exceeds 1,000 during the financial year. This scope applies to financial years starting on or after 1 January 2027. Member States must bring into force the provisions needed to comply with Articles 1, 2 and 3 of the directive by 19 March 2027 at the latest, and with its Article 4 by 26 July 2028 at the latest.
Two points condition how AI can be used on this topic. First, the ESRS have been revised: on 3 July 2026 the European Commission adopted two delegated acts, one of which amends Delegated Regulation (EU) 2023/2772 as regards the simplification of certain sustainability reporting standards. The Commission states that these acts are not in force until they are published in the Official Journal, which has not happened yet; the applicable text therefore remains Delegated Regulation (EU) 2023/2772 as currently in force. The EFRAG Secretariat published a 2026 draft list of ESRS datapoints on 28 August 2026, open for fatal-flaw review until 23 October 2026. Second, the models were trained on the earlier version of the standards.
Practical consequence: ask for structure, rewording and review, never for normative references.
Three settings that change the quality of the answers
Prompts 1 to 3 · Double materiality analysis
AI-assisted double materiality remains an in-house exercise, documented and validated internally. Our article on CSRD double materiality covers the method in detail.
Build the long list of topics. Prompt: "You are a sustainability analyst. Based on the activity description below, propose a long list of candidate sustainability topics, classified as environmental, social and governance, and flag those you are proposing by sector analogy rather than from my text." · To provide: activity, sector, sites, processes, typical customers. · Limit: candidates, not a materiality outcome.
Draft impact statements. Prompt: "Rephrase each topic below into distinct impact statements, separating positive from negative impact and actual from potential impact, and specifying the location in the value chain, the time horizon and the people or environments affected. Do not add any topic absent from my list." · To provide: selected topics, value chain scope. · Limit: severity and likelihood remain internal judgment calls.
Have your scoring scale critiqued. Prompt: "Here is my materiality threshold and how I score severity, scope, irremediability and likelihood. Act as a critical reviewer: list the scoring inconsistencies, the cases where two comparable topics would be scored differently, and the questions an assurance provider would ask. Do not propose a new threshold." · To provide: method, scales, two scored examples. · Limit: the threshold is a governance decision.
Prompts 4 to 6 · IRO identification
Derive risks and opportunities. Prompt: "For each impact below, propose the risks and opportunities that could follow from it, distinguishing physical, transition, liability and reputational risk, and describe the mechanism through which the impact becomes a financial effect." · To provide: impacts, summarized business model, main cost items. · Limit: no figures, amounts come from management control.
Assign a time horizon. Prompt: "Classify these IROs according to the short, medium and long-term horizons defined below, indicate for each one the observable fact that would trigger a change of horizon, and flag those whose horizon cannot be decided with the information provided." · To provide: your horizon definitions, depreciation periods, contract cycles. · Limit: do not let the tool decide an undocumented horizon.
Detect blind spots. Prompt: "Compare my list of IROs against the described activity scope and flag only the blind spots: activities, sites, purchasing categories or populations that do not appear in any IRO. Phrase each blind spot as a question, without inventing an IRO." · To provide: list of IROs, description of the consolidated scope. · Limit: the answer is only as good as the description provided.
Prompts 7 to 9 · Stakeholder mapping
Prepare a consultation. Prompt: "Write 12 interview questions to consult this stakeholder category on the impacts of our activity: open, non-leading questions, understandable without CSRD knowledge. For each one, indicate the materiality information it is meant to gather." · To provide: target category, duration, topics to cover, language. · Limit: prepares the consultation, does not replace it.
Summarize verbatims. Prompt: "Group these verbatims by sustainability topic, distinguish expressed expectations from reported facts, and for each group indicate the number of verbatims and quote a representative extract. Do not extrapolate beyond the statements provided." · To provide: anonymized verbatims. · Limit: anonymize before any submission, no personal data in a public tool.
Document the engagement process. Prompt: "Write the description of our stakeholder engagement process: categories consulted, methods, frequency, and how the results changed the materiality analysis. Write 'to be completed' wherever I have not provided the information." · To provide: consultations held, dates, effects on the analysis. · Limit: address every 'to be completed' mention before publication.
Prompts 10 to 12 · Selecting the applicable ESRS
The ESRS standards determine the mandatory content. This is the stage where AI makes the most mistakes, since it does not know the revised version.
Match topics to standards. Prompt: "Match each material topic below to the relevant ESRS standard(s) and corresponding disclosure requirements, one entry per topic with the exact reference. If you are not certain of the reference in the revised version of the ESRS, write 'to be checked' instead of guessing." · To provide: material topics, ESRS version applied. · Limit: every reference must be cross-checked against the text in force.
Justify an omission. Prompt: "Write the omission justification for this topic deemed non-material: assessment method, thresholds applied, evidence, conclusion. Three paragraphs, factual tone, no superlatives, using only my material." · To provide: assessment outcome, thresholds, sources consulted. · Limit: a poorly justified omission is one of the first points raised in assurance.
Scope value chain information. Prompt: "For each requirement below, indicate what comes from our own scope and what requires data from suppliers or customers, and for the latter propose an estimation wording that spells out the method and its uncertainty." · To provide: selected requirements, consolidation scope. · Limit: actual availability must be verified with suppliers.
Prompts 13 to 15 · Data point collection
Turn a requirement into a collection sheet. Prompt: "Turn this disclosure requirement into a collection sheet: information requested, unit, scope, period, frequency, likely internal source, owning department, plausibility check. One sheet per data point, no commentary." · To provide: text of the requirement, simplified org chart. · Limit: the proposed owning department is a hypothesis to confirm.
Write the request to a contributor. Prompt: "Write a 150-word message to the head of this function requesting the data below: what is requested, in what format, by what deadline, what it is used for. Factual tone, no reporting jargon." · To provide: target function, expected data, format, deadline. · Limit: adjust to the recipient's level of reporting knowledge.
Qualify a missing data point. Prompt: "For each missing data point in this list, propose two options: a documentable estimation method, or an unavailability statement. For each one, indicate what will need to be shown to an assurance provider." · To provide: list of gaps, what exists partially. · Limit: choosing between estimating and declaring unavailable is a management decision.
Prompts 16 to 18 · Policies, actions and targets
Write a policy description. Prompt: "Write the description of our policy on this topic from the elements below: objective, scope, level of responsibility, third parties concerned, document availability. Use only my information and list at the end of the answer the sections left blank." · To provide: existing policy, scope, designated owner. · Limit: never let the tool fill a blank section.
Describe an action plan. Prompt: "Describe this action, specifying the topic it relates to, the scope covered, the time horizon, the resources committed and how its progress is tracked. Do not attribute any quantified effect absent from my material." · To provide: internal action sheet, committed budget, tracking indicator. · Limit: announced gains must come from a verifiable internal source.
Check a target. Prompt: "Review this target and verify that it includes a base year, a baseline value, a scope, a time horizon, a calculation method and a link to the associated policy. List what is missing as questions. Do not rephrase the target." · To provide: target wording as it will be published. · Limit: completeness check, not an ambition validation.
Prompts 19 to 21 · E1 climate section narrative
These prompts cover the writing of the climate section, not the emissions calculation. On the quantified part, see our article on the reliability of AI applied to carbon accounting.
Build the section outline. Prompt: "Propose a detailed outline of our E1 climate section. For each part: the disclosure requirement covered, the data needed, the indicative length, the person to consult. Do not write the text yet." · To provide: selected requirements, available data. · Limit: the outline must be cross-checked against the ESRS text in force at the date of your report.
Write the transition plan. Prompt: "Write the description of our transition plan for climate change mitigation from the elements provided: trajectory, levers selected, investment decisions, governance, alignment with strategy. Every statement must point back to an element I gave you; remove any sentence you cannot link back." · To provide: approved trajectory, levers, investment decisions, oversight body. · Limit: the section most exposed to greenwashing criticism.
Describe climate risks and opportunities. Prompt: "Based on our resilience analysis, write the description of the identified physical and transition risks, by site and by time horizon, specifying the scenarios used and their limitations. Do not cite any scenario I have not mentioned." · To provide: analysis results, scenarios used, sites concerned. · Limit: no scenario or projection produced by the tool.
Prompts 22 to 24 · Cross-chapter consistency check
Check terminology. Prompt: "Review these sections and list the vocabulary inconsistencies: same concepts described with different terms, scopes described differently, contradictory time horizons. Give the exact extract and its section for each inconsistency. Do not correct anything." · To provide: relevant sections, internal glossary if it exists. · Limit: process in batches, quality drops on very long documents.
Check internal cross-references. Prompt: "Verify that every internal cross-reference in this document points to an existing section and that the information announced is actually found there. List broken or misleading cross-references with their location." · To provide: full document with numbered sections. · Limit: does not detect references to appendices not provided.
Cross-check narrative against figures. Prompt: "Compare the qualitative statements in this text against the attached values. Flag any statement not supported by the figures, any trend described in the opposite direction from the data, and any value cited in the text but absent from the data." · To provide: written text, published dataset. · Limit: detects apparent contradictions, does not validate any calculation.
Prompts 25 to 27 · Preparing for limited assurance
Directive (EU) 2026/470 postponed to 1 July 2027 the deadline for the Commission to adopt limited assurance standards. In the meantime, the questions asked concern method, scope and evidence.
Anticipate the assurance provider's procedures. Prompt: "You are preparing a limited assurance engagement on a sustainability report. Based on the list of our indicators, propose for each one the procedures the assurance provider is likely to perform and the supporting evidence they will request." · To provide: published indicators, production method for each one. · Limit: anticipation, not an assurance standard.
Build the evidence file. Prompt: "For each indicator below, list the supporting documents to gather, the person who holds them, the period covered and the check that allows the document to be reconciled with the published figure. One entry per indicator, no commentary." · To provide: indicators, source systems, owners. · Limit: gathering the documents remains a manual task.
Describe an internal control. Prompt: "Write the description of the internal control applied to this indicator: who enters it, who checks it, at what frequency, what deviation triggers an investigation, what record is kept. Write in the present tense, with no promise of future improvement." · To provide: actual procedure, alert thresholds, tools used. · Limit: describe only controls actually in place.
Prompts 28 to 30 · Answering an auditor's questions
Simulate the interview. Prompt: "Play the role of a demanding assurance provider on this chapter. Ask your questions one at a time, starting with method and scope, wait for my answer before each next question, and flag the answers you consider insufficient, explaining why." · To provide: relevant chapter, method applied. · Limit: training exercise, no predictive value on the actual questions.
Test an exposed wording. Prompt: "Review this passage and flag any statement that could be read as an unsubstantiated environmental claim: superlatives, comparisons without a basis, promises without a deadline, effects attributed without proof. Propose a more cautious wording for each one at constant information." · To provide: written passage, available evidence. · Limit: does not replace legal advice on claims.
Prepare the response memo. Prompt: "Based on the assurance provider's question and my material, write a structured memo: question received, factual answer, attached documents, acknowledged limitations, corrective action if needed. One page maximum, neutral tone, no defensive justification." · To provide: exact question, response material, available documents. · Limit: the memo engages the company, validation before sending.
FAQ
Can you use ChatGPT to write a CSRD report?
Yes for structuring, rewording and checking consistency. No for producing substantive content without a source. The report is verified by a third party and engages the collective responsibility of the administrative, management and supervisory bodies: every statement must be linked to internal data and a supporting document. A well-built CSRD ChatGPT prompt saves formatting time, not collection time.
Which companies are covered by the CSRD after directive (EU) 2026/470?
Directive (EU) 2026/470 of 24 February 2026 reserves the obligation for companies whose net turnover exceeds EUR 450,000,000 and whose average number of employees exceeds 1,000 during the financial year. This scope applies to financial years starting on or after 1 January 2027. Member States must transpose the corresponding provisions by 19 March 2027 at the latest.
Can AI perform a double materiality assessment?
No. It can prepare the long list of topics, rephrase impact statements and critique a scoring scale. Assessing severity, scope and likelihood relies on knowledge of your sites, your contracts and your stakeholders. It is also the part the assurance provider examines first.
How do you avoid an AI-generated text being challenged during limited assurance?
By linking every statement to a provided element and removing the rest. Three instructions reduce the risk the most: forbidding the tool from filling in missing information, asking it to flag blank sections, and having each section reviewed by the person responsible for the topic before it is included.
Can you share your sustainability data with a public AI tool?
Not without precautions. Personal data, contractual information and unpublished financial elements have no place in a general-purpose tool. The alternatives are anonymization, an offer with a data non-reuse commitment, or deployment in a controlled environment.
Further resources
Kabaun and report production
These prompts handle the text. The report rests on data that must be collected, traced and presented to an assurance provider.
Kabaun covers that part: generation of reports compliant with the CSRD and ESRS standards including ESRS E1, supporting documents linked to every entered data point, a complete and tamper-proof audit trail with timestamp and author identification, time-limited delegated access for auditors and consultants. The monitoring module tracks CSRD, ESRS and EU Taxonomy developments applicable to your scope. The Klem assistant answers in natural language on your data, with every action remaining traceable and reviewable.
Conclusion
Reserve these 30 prompts for the tasks where AI is reliable: structuring, rewording, checking consistency, preparing questions. Keep normative references and materiality judgments out of its reach.
First concrete action: run prompt 24 on your last published report. The contradictions between narrative and figures show up in minutes, and they are the ones an assurance provider will flag.
Kabaun automates the collection and traceability of your sustainability data → [www.kabaun.com/contact](https://www.kabaun.com/en/contact)