Positioning
“What can this do that ChatGPT can’t?”
The question comes up in every other first conversation, and it is fair. The honest answer is uncomfortable: on a single, familiar technical question a general model gets close. The difference arises in four situations that nobody tries first.
The difference is not how the answer is phrased but where it comes from. A general language model produces a plausible answer from training knowledge — often right on familiar topics, and you cannot tell from the answer when it is not. CSRD-GPT retrieves passages from a maintained collection in which ESRS (2023) and ESRS (2026) are kept apart, and names the source. Concretely the systems differ on four points: version separation, evidence chain, client separation and professional behaviour.
Point 1
Version separation: the difference that cannot be trained away.
Since 3 July 2026, delegated act C(2026) 5010 final has set out a revised version of the European Sustainability Reporting Standards, amending Delegated Regulation (EU) 2023/2772. For financial years 2026, both versions apply side by side. For financial years beginning on or after 1 January 2027, the revised version is mandatory.
A general language model has seen both bodies of text in training — as the same material, with no dividing line. It therefore cannot distinguish between them and mixes them. This is not a quality defect that OpenAI or Microsoft can fix with the next model; it follows from how the systems are built. Keeping the versions apart has to happen in retrieval, not in the model.
One visible symptom: in the revised standards the water standard is titled simply “Water”. The earlier title “Water and Marine Resources” no longer appears in the new text. A model that completes standard titles from memory still returns the old one — and sounds entirely confident doing so.ESRS E3, as amended by C(2026) 5010 final
Point 2
Evidence chain: a reference is not yet evidence.
A general model can produce a source reference without having read the source — the reference is part of the generated language, not its origin. This is why wrong answers look exactly like right ones.
CSRD-GPT works the other way round: the passage is retrieved from the collection first, then the answer is written on that basis. The citation therefore shows what the answer actually rests on. That is the precondition for checking it — and checking is not optional in an assured report.
Point 3
Client separation: where Copilot drops out.
Microsoft Copilot is the more serious competitor, because it is already present in the enterprise and has access to SharePoint. The distinction therefore lies elsewhere than with ChatGPT.
Copilot sits in the employer’s tenant. For a consultant serving five clients that is not a workable model: client documents must not sit in another client’s tenant, nor travel between engagements. In many advisory contracts, uploading client material into a general AI service is grounds for termination. CSRD-GPT creates a separate knowledge space per client or project; material from one space is not retrieved in another.
Point 4
Professional behaviour: what the system does when it does not know.
A general model fills gaps. It completes lists from the legal text, invents figures where the report would need them, and backs down when contradicted. In sustainability reporting, those are the three most expensive failures there are.
Named placeholders, not invented figures
Where a figure from the company is missing, a named placeholder appears with a task attached — not a plausible number that nobody questions again later in the draft.
Exclusions get documented
Where a topic is assessed as not material, the reasoning belongs in the report. The system prompts for it instead of leaving the gap silent.
The reporting year gets asked for
Because the applicable version depends on the financial year, the reporting year is asked for or derived from the documents before a requirement is named.
Side by side
The four points compared.
| Criterion | ChatGPT (general) | Microsoft Copilot | CSRD-GPT |
|---|---|---|---|
| Origin of the answer | training knowledge plus web search | training knowledge plus documents in the Microsoft tenant | curated collection of legal sources |
| Standard versions kept apart | no — both versions in the same training | no | yes, as separate holdings |
| Evidence in the legal act | reference is generated, not retrieved | cites tenant documents, not the legal act | citation from the retrieved passage |
| Client separation | none | only within your own tenant | separate space per client or project |
| Behaviour when data is missing | keeps writing, fills plausibly | keeps writing | named placeholder with a task attached |
| Maintenance of sources | training cut-off, not steerable | depends on tenant content | superseded legal versions are removed |
| Best suited to | orientation, phrasing help, summaries | working on your own documents inside the enterprise | evidence-bound technical work on CSRD and ESRS reporting |
The assessment of ChatGPT and Copilot refers to the position as at 1 September 2026 and to the generally available configurations without customer-specific integration.
In fairness
Where a general model remains the better choice.
If you want to shorten a text, structure a presentation, draft an email or read into a new subject, you do not need a collection of legal sources. A general model is faster, broader and cheaper for that. CSRD-GPT earns its place where the statement has to be evidenced — and where it matters which version is meant.
Frequently asked
Why can’t I just type ESRS questions into ChatGPT?
For a single, familiar technical question you often get a usable answer. It becomes unreliable where the version matters, where a complete list from the legal text is needed, or where the statement has to be evidenced. A general model has seen ESRS 2023 and ESRS 2026 in training and holds no dividing line between them.
We already have Microsoft Copilot across the group. Isn’t that enough?
Copilot works on the documents in your tenant and on training knowledge. It has no maintained legal source with version separation, no evidence chain into the legal act, and no client separation across organisational boundaries. For a consultant with several clients, the last point rules it out.
Is CSRD-GPT its own model?
No. The difference is not in the model but in what the model gets to see: a curated, version-separated source collection, professionally defined behaviour, and knowledge spaces kept apart per engagement.