Gains:
- Ability to print the report section by section with only verified data and prohibit artificial intelligence from generating new numbers/claims
- Ability to scan for inconsistencies of the same metric in different sections and determine the correct value by going to the source
- Ability to prevent correct text from turning into a misleading summary by maintaining 'partial/estimated/not approved' reservations during summarization
All the data was collected, carbon was calculated, mapped to the framework, stakeholders were listened to; Now all this should turn into a readable sustainability report. A CSRD report contains hundreds of pages, dozens of indicators, many narrative sections. This is where the most visible contribution of artificial intelligence (AI) is: producing draft text, making different sections coherent, summarizing long technical content. But this is where the biggest risk lies: AI is adept at producing fluent, beautiful, proof-free text. In this unit you will learn how to maintain accuracy and auditability while printing the report quickly.
Let's put the basic principle from the beginning: AI writes the text, the human connects it to evidence. Every numerical claim in a sustainability report should be based on an audit trail, and every qualitative claim should be based on a fact. The AI's job is to organize, streamline, and make coherent the expression; not to produce the truth of the claim. “Well-written” and “accurate and substantiated” are two independent qualities.
Where is AI strong and where is it risky in report writing?
Quest
AI contribution
Risk
Producing a first draft
Quickly fills the blank page
Can add claims without evidence
Consistency check
Catches term/unit conflict
Can "correct" errors
summarizing
Shortens the long section
Can omit critical nuance
Tone editing
Provides formal/clear language
May slide into exaggeration
Translation
Generates multilingual reports
Mistranslates technical term
The critical rule is this: when giving the AI a report, give it only validated inputs and forbid it from generating new numbers/claims. Otherwise, the AI derives numbers from previous sentences that “seem reasonable” but have no source.
Tip: When drafting a report, always give the AI this rule: "Do not add any numbers, percentages, dates, or claims other than the data I gave you; if something is missing, write [DATA REQUIRED] there." This single sentence prevents most of the fabricated allegations that have crept into the report.
Step by step: secure report writing
1. Build the skeleton first. Plan the section headings and what verified data will go in each section. AI produces good contents/skeleton.
2. Section by section, feed it with data. Print each partition separately, with only the verified data for that partition. Don't say "make up" the entire report at once.
3. Make a claim-evidence match. Note the source next to each numerical sentence. A sentence without a source will not be included in the report.
4. Check consistency. It is a common mistake for the same metric to appear with different values in different sections; The AI is good at scanning for this, but the human approves the correction.
5. Scan for hype and greenwashing. In the final round, clear the text with an eye for “immeasurable/non-evidentiary statement” (the next unit digs into this).
Weak prompt / Strong prompt
Weak prompt:
Write the climate section of our sustainability report.
No data; The AI makes up a "typical climate episode", inserting numbers and targets into the text that you don't actually have.
Powerful prompt:
Your role: ESG report writer. Write: draft of the “Climate” section of the report. The ONLY data you will use (opt out, add new number):- Scope 1: 1,240 tCO2e (2024, verified)- Scope 2: 2,100 tCO2e (2024, location-based)- Scope 3: partial, 60% covered (including proxy)- Target: to 2030 up to Scope 42% reduction in 1+2 (not SBTi approved, applied)Rules:- DO NOT FAKE number/percentage/date other than the data I give; if missing, write [DATA REQUIRED].- Make clear that Scope 3 is partial and includes a proxy.- Write honestly that the target has not yet been approved, do not say "approved".- Tone: clear, formal, understated.
Prompts for consistency and quality control
Perform a CONSISTENCY check on the report draft below. Identify (mark, not correct):1) Situations where the same metric occurs with different values in different places2) Unit inconsistencies (kg/ton, kWh/MWh)3) Conflict of a claim in one section with another section4) Numerical claims with unspecified sourcesList each finding with section/row reference.Draft: [DRAFT]
The following prompt summarizes a long technical section to the manager; but it forces you to maintain nuance.
Your role: report editor.Task: reduce the technical section below to 150 words for the executive summary.Rules:- Do not change or exaggerate any numbers by rounding.- SKIP reservations such as "partial/estimated/not confirmed", move to summary.- Present uncertainty as "certain".Section: [CHAPTER]
Attention: AI most often throws out "reservations" when summarizing. If phrases like "Scope 3 data is partial", "target not yet confirmed", "data contains prediction" are deleted for the sake of the summary, an accurate text turns into a misleading summary. Maintaining reservations in summaries is as important as preserving numbers.
three mini cases
Case 1 — Leaked fake target. A company had AI write its climate section with a simple prompt. YZ added a "2040 net zero" target that was not actually set because "it was mentioned in similar reports." The editor caught this; The target was not actually taken and would have been a misrepresentation if published.
Case 2 — Inconsistent metric. In the summary section of a report, water consumption was written as "1.2 million m³" and in the detail section it was written as "1.8 million m³". The AI's consistency scan flagged the contradiction; went down to the source, the correct value was combined as 1.8 million m³.
Case 3 — Deleted reservation. An executive summary was written by AI omitting the caveat that “60% of Scope 3 is an estimate” and the report appeared to be fully measured. The pre-audit check added back the reservation; otherwise the summary would be more assertive and misleading than the report itself.
Common mistakes
- Printing a report without providing data. The empty prompt drives the AI to make up numbers and targets.
- Producing the entire report at once. Writing by feeding it data, section by section, maintains control.
- Discarding reservations when summarizing. The words "partial/estimated/unconfirmed" should also remain in the summary.
- Having the AI blindly correct the inconsistency. AI doesn't know which value is correct; man descends to the source.
- Mistaking "fluent" for "correct". If a well-written sentence is without evidence, it still won't be included in the report.
In summary
Report writing is the most visible but riskiest area of AI contribution. AI quickly fills the blank page, keeps sections consistent, and summarizes long content; but it tends to add claims without evidence, delete reservations, and produce exaggerations. Safe method: feed only verified data, forbid it from generating new numbers, print section by section, scan for consistency but let the human do the correction down to the source. A "well-written" report is only a good report if every claim is based on evidence.
Application task
Have 4-5 numerical data items that you assume are confirmed (one "estimated" and one "unconfirmed target"). Ask the AI for a report section draft via the powerful prompt above. Then: (1) verify that no numbers that you did not provide have leaked into the text, (2) check that the "estimated" and "unconfirmed" reservations are retained, (3) reduce the text to a 150-word summary and confirm that the reservations remain in the summary.
checklist
- [ ] I entered only verified data in the report.
- [ ] I explicitly gave the AI the "make up new number/claim" rule.
- [ ] I printed the report section by section, feeding it with data.
- [ ] I linked each numerical assertion to a source/audit trail.
- [ ] I scanned for inconsistencies and determined the correct value by going back to the source.
- [ ] I kept the "partial/estimated/not confirmed" reservations in the summaries.
- [ ] I eliminated the exaggerated and immeasurable expressions in the last round.