Who it is for · Assurance
Hold disclosures against the legal text — not against a formulation.
In the assurance of sustainability reporting the question is never whether a statement sounds plausible but what it follows from. A tool that can produce a source reference without having read the source is not merely useless here; it is dangerous.
CSRD-GPT retrieves the passage and writes on that basis — the citation shows what the answer rests on. For assurance work that means: requirements in their own words with citations, an explicit statement of the version applied, and a named gap list when a client’s disclosures are held against the disclosure requirements.
Typical questions
What actually gets asked in assurance work.
Which version applied to this financial year?
With two versions applicable side by side and an option to choose, this is the first question of any assessment — and the one general models fail.
Is the list complete?
Lists from the legal text in their own words, not as a summary. An eight-item list returned with six looks complete and is not.
Is the materiality assessment traceable?
The requirements on approach and documentation in their own words, as the benchmark against which the approach presented is assessed.
Are the exclusions reasoned?
Where a topic was assessed as not material, the reasoning must be in the report. This is a common finding and easy to test.
How did comparable companies report?
Published reports from listed German companies as a benchmark for the extent and depth of disclosure.
Where does this figure come from?
A metric without derivation in a draft is a finding. Report drafts from CSRD-GPT carry named placeholders at such points rather than figures.
Built for this
What is built for the assurance perspective.
- Requirements in their own words with citations, not as paraphrase
- Explicit statement of the version applied with every answer
- Separate knowledge spaces per client, with no mixing across engagements
- Holding position under mere contradiction; showing the range where interpretation is contested
- No metrics without derivation in generated drafts
- Published reports as a benchmark in the same collection
What CSRD-GPT expressly is not
Not audit evidence, not an audit procedure and not a judgement. The audit evidence is the legal text the answer points to.
The value is in the path there: finding the relevant passage, reading it in its own words, knowing the version. That is the work that costs time today and carries no professional value of its own.
Frequently asked
Does an AI answer count as audit evidence?
No, and that is not the claim. The audit evidence is the legal text the answer points to. CSRD-GPT shortens the path there and makes it traceable — the judgement stays with the auditor.
How does the system behave on contested interpretations?
Where several defensible readings exist, the range belongs in the answer. Contradiction alone is not a reason to withdraw an evidenced statement. That backing down is precisely the weak point of general models.
Does client material stay separate?
Yes. On the Premium plan each client gets its own knowledge space; material is not retrieved across spaces and is not used for training.