AI ESG Report Drafting for Sustainability Consultants

Keep the numbers out of the model's hands: calculate in a spreadsheet, give AI the finished data table, and have it draft narrative in which every sentence cites a data cell. Then trace each number back to its cell and test each wording against the evidence. Anything unsupported is cut or reworded before the client sees the draft.

Two things make ESG drafting riskier than most AI-assisted writing. Figures such as emissions depend on factors, boundaries and dates that a model will happily invent. And a report's sentences get recycled into websites, tenders and adverts, where wording rules apply. Since 27 September 2026 the EU's rules on green claims in consumer marketing have applied, so a sentence that looked harmless in a PDF can become a problem on a product page.

Follow me on Instagram@sagnikteaches

A first report for an event venue, and who does what

A three-person sustainability consultancy is writing a first report for a 14-person event venue that hosts about 60 events a year. Two corporate customers have sent supplier questionnaires, and the owner wants one document that answers both, plus a short page for the website. The report covers calendar 2025 and is voluntary. Everything here is invented to show the method, including the emission factors, which are placeholders for the sums and not for use.

Connect on LinkedInSagnik Bhattacharya

The consultancy works to one rule: a spreadsheet makes the numbers, the model drafts the words, and a register decides which claims survive. Each of the three has one job, and none of them is allowed to do another's.

Subscribe on YouTube@sagnikteaches

Start with a data table the model is not allowed to change

The inputs go in one sheet with the evidence named beside each, so a reviewer can reach any figure in two clicks.

Where data is missing, estimate by a stated method and label it. The two gas months here are filled from the average of the other ten (8,000 kWh a month, so 16,000 kWh), which means 83% of the gas figure is measured and 17% is estimated. The report has to say so. A model will not volunteer that, and if asked to fill the gap it may well invent a plausible figure.

The calculations then sit in a second sheet, with the factor set named for the reporting year and its version recorded:

Scope 1 means emissions from sources the venue owns or controls, and Scope 2 means emissions from the electricity, steam, heat or cooling it buys. A model can help write these formulas and explain them to the client, but a second person recomputes three of them with a calculator before anything goes further. Set a rounding policy now (one decimal for tonnes, whole numbers for percentages) so the same figure does not appear as 107.7 in one place and 108 in another.

The drafting prompt: no cell, no sentence

Give the model both tables, a one-paragraph boundary note and a short list of things it may not do. The boundary note is written by the consultant, not the model: the organisational boundary (the venue's premises and one van), the scopes included (1 and 2), the scope not measured (3), the period, the factor set and version, and the estimation methods.

You are drafting the environmental section of a first voluntary report for an event venue.
Sources: the INPUT table (D1-D9), the CALCULATION table (C1-C8) and the BOUNDARY note below.
Write 150 to 200 words of plain narrative.
Rules:
- Every sentence that contains a number or an assertion ends with its source in square
  brackets, for example [C4].
- Use only numbers that appear in the tables. Do not calculate anything yourself.
- State that the gas figure includes two estimated months.
- No adjectives of degree (significant, substantial, strong).
- Do not use "eco-friendly", "green", "sustainable" or "carbon neutral".
- No comparisons with other years, other venues or equivalents (cars, trees, homes).
- No targets, commitments or plans.
- If something is not in the tables, write [NO DATA] rather than filling it in.

A run on the venue's tables produces something like this (illustrative):

In 2025 the venue's Scope 1 and Scope 2 emissions were 107.7 tonnes of CO2e [C4], or 1.8 tonnes per event [C6]. Electricity accounted for 77% of the total [C5]. The gas figure includes an estimate for two of the twelve months, based on the average of the other ten [D3]. On-site solar supplied 3.9% of the electricity the venue used [C8]. Of 60 tonnes of waste, 22 tonnes (37%) were recycled [C7]. No reportable injuries were recorded among an average headcount of 14 [D9, D8]. Emissions from supply-chain purchases and guest travel are not included [NO DATA].

Two things need a human. The last sentence is a boundary statement, not missing data, so the tag should point to the boundary note and the wording should say these emissions "have not been measured for this report". And the first run of this prompt, before the rules were as tight, produced "significantly lower than a typical venue" and "equivalent to the emissions of 23 cars". Models reach for comparisons and equivalents because they are common in sustainability writing, and a prompt alone will not always stop them. The number trace below is what catches whatever slips through.

The claim register: one row for every sentence that asserts something

The register is a sheet kept next to the draft. Each row holds a sentence that claims something, what evidence would support it, what evidence exists, and a decision. Sentences come from three places: the AI draft, the client's existing website and brochures, and the owner's own wishes in the kick-off call. The last two are where most unsupported claims start.

For the pattern behind this table, outside ESG, see the tutorial on backing up every claim AI writes in marketing. The register is the same habit applied to a report. Its test is blunt: if a row has no evidence in the third column, the sentence does not go in, however good it sounds.

Six wording tests, with the EU green claims rules in mind

The rules that began applying on 27 September 2026 come from Directive (EU) 2024/825, often shortened to EmpCo, and the directive sets no transition period. It amends the EU's rules on unfair business-to-consumer practices, so a report sent to a lender or a corporate customer is not directly in scope. In practice, though, report sentences are copied onto websites, packaging and adverts within weeks. If the client sells to consumers in the EU, run every sentence through these tests as if it were going on a product page.

After the six tests, add one pass that uses the model as an adversary rather than an author. Give it the final draft and the register, and ask it to read as a sceptical customer.

Read the report text below as a sceptical customer who has never met the venue.
List every sentence that a reasonable reader could take to mean more than the evidence
in the register shows.
For each: quote the sentence, say what a reader might wrongly take from it,
and name the register row that limits it.
Do not rewrite anything.

Two items from a run, as an illustration. First, "On-site solar supplied 3.9% of the electricity we used": a skimming reader may assume the venue runs mostly on solar, and register row 2 limits it, so the sentence should add that the rest came from the standard grid supply. Second, "No reportable injuries were recorded": a reader may take it to mean no incidents of any kind, and row 6 limits it, so the definition of "reportable" belongs beside it. The model will also flag some sentences that are fine, and you can ignore those. The pass is cheap, and it tends to surface the sentences that are technically true but easily misread, which are the ones that cause trouble later.

A corporate customer that quotes your client's report in its own consumer marketing is making that claim itself, and it will come back to your client for the evidence. That is another reason to settle the register before the report goes out.

The separate Green Claims proposal is a different instrument. It stalled in June 2025, when the Council cancelled the final negotiating round after the Commission announced that it intended to withdraw the proposal. The EU's own procedure records still list it as ongoing, with no formal withdrawal and no new step since, so check its position before telling any client that something depends on it. The directive above is the one in force.

Tracing every number in the draft back to a cell

Once the register is settled, check the numbers mechanically. Ask the model to list, not to judge: every number in the narrative, the sentence it sits in and the cell it cites. Do the comparison in a spreadsheet, using a formula such as a rounded copy of the cell against the number in the text, because comparison is arithmetic and arithmetic stays out of the chat.

Below is a draft narrative and the cell references it cites.
List every number in the draft, one per row:
number as written | the sentence it is in | the cell it cites (or NONE) | unit as written
Include numbers written as words and percentages.
Do not check the values. Do not correct the draft.

After the spreadsheet comparison, the venue's trace looks like this (illustrative):

Two checks the trace cannot do for you. It cannot tell whether the cell itself is right, which is why a second person recomputes three results from the raw inputs. And it will not notice a unit slip unless you ask, so add one more pass: list every kWh, MWh, kg and tonne in the draft and confirm that each matches the unit in its cell. A kWh quoted as MWh is a thousandfold error that reads perfectly well.

Answering the two customer questionnaires from the same table

The venue's reason for reporting was two supplier questionnaires, and this is where one table pays for itself. Each question is answered from a cell, using the register's wording, so the report, the website page and both questionnaires can never disagree. A sample of the answers:

Notice that two of those answers are plain admissions. A questionnaire invites a flattering answer, and a model asked to fill it in will oblige, which is why the register is open on screen while anyone answers. "No target has been set" is a perfectly good answer; an invented target is a problem the venue will carry into every future questionnaire.

What an assurance provider will still ask for

AI drafting changes how the words arrive, not what a reviewer needs. If the client wants anyone outside to look at the report, expect requests for the source documents (invoices, meter readings, waste tickets), the calculation workbook with the factor set and its version, the boundary and estimation notes, the claim register and management's sign-off. The lower level, limited assurance, involves less testing and ends in a negative conclusion: nothing came to the provider's attention suggesting the information is materially misstated. The higher level, reasonable assurance, involves more testing and ends in a positive opinion that the information is prepared in line with the criteria. At either level the evidence is the invoices and meters. The model's output is not evidence of anything.

Three habits keep that conversation short. Keep the prompts and outputs in the file. Tell any reviewer how the draft was produced before they ask. And never describe the report as assured, verified or audited unless a provider has issued a statement on it, and then only for the scope that statement covers.

Which frame the client reports against is a separate decision. Whether a company must report at all depends on its size and on the EU's continuing simplification of reporting duties, so confirm the position for each client instead of assuming it. Many small clients report voluntarily because a customer or lender asked, and the Commission recommended a voluntary standard for small and medium-sized companies, known as VSME, on 30 July 2025. Match the report's headings and data points to whichever frame the questionnaires actually use. For a related case, see the tutorial on emissions figures for freight tenders, and for turning another client's data table into narrative, turning charity data into impact reports uses the same data-to-words discipline.

Telling the client how you use AI, and protecting their data

Put it in the engagement letter, so it is agreed before any data is uploaded.

• Tools and plans. Name them. Business plans of ChatGPT and Claude don't train on your content by default; on personal plans, switch the training setting off before use.
• What AI touches. Drafting words, tidying tables and writing formulas. Not sourcing data, deciding which claims survive, or giving assurance.
• What is shared. Send the aggregate tables, not the invoices. Account numbers, addresses and customer names have no place in a prompt.
• Who is responsible. The consultancy signs off every statement, whatever tool produced the first draft.

Keep each client's data and drafts in separate chats or projects, and never paste one client's figures into work for another; the tutorial on reusing past client work without leaking client data covers the habits. For the review steps themselves (the second person who recomputes, the register sign-off), setting up human review without slowing down shows how to make them routine.

Hours saved on a first report, and where they come from

Invented hours for a consultancy that drafts by hand today:

The saving, about nine and a half hours, comes almost entirely from drafting. The register gets longer because it now exists, and that is deliberate. As of November 2026, ChatGPT Business and Claude Team are each $20 a seat on annual billing ($25 monthly) with a two-seat minimum, and the spreadsheet you already use does the rest. Next year the picture improves: with 2025 as the baseline, trend claims become possible, and the number trace gets an easy second column to check against. For the wider report-writing workflow around all this, see AI consulting report writing from findings to a client-ready draft.

Questions consultants ask about AI in ESG reports

Can AI calculate a client's emissions?

It can help write and explain formulas, but the calculation should live in a spreadsheet you can audit, with factors, units and sources recorded. Models make arithmetic and unit errors, and will invent a factor if you do not supply one. Use AI for the narrative around finished numbers, then trace every number back to a cell.

Do the EU green claims rules apply to an ESG report?

The directive that began applying on 27 September 2026 covers business-to-consumer marketing, so a report sent to a lender or corporate customer is not directly covered. But a report's sentences are often reused on websites, packaging and adverts. If the client sells to consumers in the EU, test every sentence as if it will be pasted into a product page.

Can we call the report assured or verified?

Only if an assurance provider has issued a statement on it, and then only for the scope that statement covers. An AI-assisted draft and a client's internal review are not assurance. If the client wants external assurance, ask the provider early what evidence they need and tell them how the draft was produced.

Should we tell the client we use AI to draft?

Yes. Name the tools, say which parts they touch (drafting narrative, not calculating or sourcing data), confirm that client data goes only to plans that don't train on it, and record that you remain responsible for every statement. Put it in the engagement letter so it is agreed before any data is uploaded.

Further reads

• AI for Expert Witnesses: Report Drafting With Defensible Checks — Another report-drafting workflow built around defensible checks.
• ChatGPT Business vs Claude Team: Seat Costs for a Small Consultancy — Seat costs and privacy defaults for a small consultancy.
• The AI Tool Stack a One-Person Consultancy Actually Needs — The wider tool stack a small consultancy actually needs.
• Does ChatGPT Train on Client Data? Business vs Free Plans — Which plans train on client data and which do not.
• A Simple AI Risk Register for Small Businesses (With Template) — Record the risks of AI-assisted reporting in one simple register.
• Should You Tell Customers When You Use AI in Your Marketing? — How and when to disclose AI use in customer-facing material.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: text of Directive (EU) 2024/825 (EUR-Lex); EU procedure records for the Green Claims proposal and an EU Parliament press release of 23 June 2025; EU Commission corporate sustainability reporting page (updated 3 July 2026); GHG Protocol scope 2 guidance; an ACCA technical article on ISSA 5000; ChatGPT Business and Claude pricing pages (all checked 4 November 2026).