How Verdandi's source-anchored Q&A works

The question behind the question

A compliance professional working through EU sustainability regulation rarely starts with a general question. They start with a specific one. What does the adopted CSDDD text actually require in terms of stakeholder engagement. Whether the EUDR due diligence statement obligation applies to a downstream operator or only to the operator who first places the product on the market. What the EFRAG guidance says about materiality assessment for a company with no direct EU operations but a significant EU customer base.

These questions have answers. The answers exist in the regulatory text, in Commission guidance, in EFRAG Q&As, in consultation papers. The problem is not that the answer is unknowable. The problem is that finding it means locating the right document, in the right version, and reading enough of the surrounding text to be confident the answer is not qualified by a clause two paragraphs later.

This is the task Verdandi’s Q&A feature is built to shorten, without shortening it in a way that trades accuracy for speed.

What source-anchored means

Every answer Verdandi gives is generated from a specific set of documents that Verdandi retrieves in response to the question, not from a general model of what EU sustainability regulation says. This distinction matters more than it sounds like it should.

A general-purpose AI tool answering a regulatory question is drawing on whatever it absorbed during training, blended with whatever pattern of language is statistically likely to follow the question asked. It has no reliable way to tell you which version of a regulation it is describing, whether a provision it mentions has since been amended, or whether the guidance it is citing has since been superseded. It will often produce an answer that reads as confident and specific while being wrong in a way that is not obvious from the text alone. For a fuller account of why that failure mode is particularly dangerous in a compliance context, see why source-anchored regulatory intelligence works where generic AI does not.

Verdandi works differently. When a question is asked, the system retrieves the specific documents most relevant to that question from its indexed source library, and generates the answer from those retrieved passages, not from an internal model of the regulation. The system tells you which documents it drew on, and you can open the source and read the passage yourself.

This is the same underlying architecture as Forseti, applied to the EU sustainability regulatory stack rather than financial regulation. The reasoning for building it this way, rather than relying on a general model, is the same in both cases: a compliance answer is only useful if you can verify it, and verification requires a visible link back to the specific source it came from.

The five streams, and asking within one

Verdandi organises its source library into five regulatory streams, and today a question is asked against one stream at a time, chosen by the user before asking it.

Legislation covers binding EU law: regulations, directives, and decisions currently in force. This is where you go for what the adopted text of CSRD, CSDDD, EUDR, CBAM, or the EU Taxonomy actually says right now.

Proposals covers pre-legislative texts: Commission proposals and trilogue documents that have not yet been adopted. This is where you go to understand what a regulation is likely to require once a proposal currently moving through the legislative process is adopted, without mistaking it for current law.

Guidance covers interpretive guidance: EFRAG Q&As and Commission FAQs on CSRD, ESRS, and the EU Taxonomy. This is where the interpretation of a principles-based requirement is most often made concrete, in a form that is not binding but that auditors and enforcement authorities draw on.

Consultations covers EFRAG, EBA, ESMA, and EIOPA consultation papers and draft regulatory and implementing technical standards. This material is pre-legislative in form, though draft RTS and ITS within it are near-certain to become binding law, which is worth keeping in mind when weighing how much to plan around it.

Case law covers CJEU and General Court judgments interpreting EU sustainability regulation. This is binding judicial interpretation that national courts must apply, and it shapes what adequate compliance looks like in ways that go beyond what the adopted text alone specifies.

Asking a question within a chosen stream means the answer is generated only from that stream’s retrieved passages, and is anchored only to sources within it. A question asked in Legislation will not surface a Commission proposal that would change the answer, and a question asked in Proposals will not tell you what the currently binding text says. Keeping the question scoped to the right stream, and knowing to check more than one stream when a question genuinely depends on more than one, is currently the user’s job rather than the system’s.

Bringing these streams together automatically, so that a single question returns an answer that draws on and distinguishes between what is binding, what is proposed, and what is still under consultation, is on our roadmap. It is not yet built, and this article will be updated when it ships.

What happens when you ask a question

The mechanics are worth explaining honestly, because they clarify both what the feature is good at and where its limits sit.

When a question is submitted within a stream, Verdandi searches that stream’s indexed document library for the passages most likely to be relevant, drawing on the full text of the documents rather than summaries of them. It ranks these by relevance to the specific question asked, not just by keyword match. A question asked in Guidance about stakeholder engagement under CSDDD will surface the relevant EFRAG or Commission guidance on the topic, rather than returning only the passage where the word appears most often.

The system then generates an answer from the retrieved passages, structured to address the question directly, with the specific documents it drew on shown alongside the answer. Because the stream is fixed for the question, the answer reflects that stream’s status: an answer from Legislation reflects currently binding text, an answer from Proposals reflects a text that has not yet been adopted, and so on. Understanding what that status means, and why the distinction between binding and proposed matters, is discussed in more detail in why there is no official checklist for EU sustainability compliance.

Because the answer is generated from the specific documents retrieved for that question, rather than from a general model of the regulation, the source is always visible. You can click through to the passage the answer relies on and read it in its original context. This matters because a compliance answer that cannot be checked is not a compliance answer you can defend to an auditor, a regulator, or a court. It is an assertion.

What the feature is good for

The most direct use case is exactly the situation described at the start of this article: a specific question, arising from a specific supply chain fact pattern, that needs an answer grounded in the current text of the relevant regulation or guidance.

It is also useful for a task that is easy to underestimate: checking whether an assumption still holds. A compliance programme built eighteen months ago against the EUDR application dates as they stood at the time may be operating on outdated assumptions, since those dates have since moved twice. A question asked directly against the current source library returns the current position, not the position the compliance team last checked.

It is useful for scoping questions that determine which parts of a regulation apply to a specific business at all. Whether a particular derived product falls within Annex I of the EUDR. Whether a company’s revenue and headcount put it in CSRD wave two or wave three. Whether a specific due diligence obligation under CSDDD extends to indirect business partners or only direct ones. These questions have defined answers in the source text, and the value of asking them directly against that text, rather than relying on a secondhand summary, is that the answer comes with its own verification built in.

What the feature is not

It is not a substitute for legal advice, and Verdandi does not present it as one. An answer grounded in the source text tells you what the regulation says. It does not tell you how a specific fact pattern in your business should be characterised against that text, and it does not carry the professional liability that a qualified legal opinion carries. Where a question turns on a genuinely contested interpretation, or where the consequence of getting it wrong is severe, the appropriate use of a source-anchored answer is as the starting point for that conversation with a qualified professional, not the end of it.

It is also not a way to skip the interpretive work that principles-based regulation requires. As discussed in why there is no official checklist for EU sustainability compliance, a large part of EU sustainability compliance is judgement about whether a specific process is adequate for a specific risk, not retrieval of a defined answer from a defined text. Source-anchored Q&A is well suited to the retrieval part of that task. It does not replace the judgement part, and no tool honestly can.

It is also, today, scoped one stream at a time. A question that genuinely needs an answer spanning what is binding, what is proposed, and what guidance has since said about it currently has to be asked separately in each relevant stream, with the reader doing the work of reconciling the answers. This is a real limitation, not a design choice we think is ideal, and it is the reason cross-stream synthesis is next on our roadmap.

Finally, it is bounded by what has been indexed. Verdandi’s coverage spans CSRD, CSDDD, EUDR, CBAM, and the EU Taxonomy, across all five streams: Legislation, Proposals, Guidance, Consultations, and Case law, and that coverage expands as the regulatory landscape develops. A question that falls outside the current scope of coverage, such as a query specific to a national implementing measure in a domain where Verdandi’s coverage is intentionally limited to EU-level instruments, will not return a useful answer, and the system should not be expected to invent one.

Why this matters more in a principles-based regime

Source-anchored Q&A would still be a useful convenience in a prescriptive regulatory domain where the rules are specific and stable. It is closer to essential in a domain like EU sustainability regulation, where the interpretation of a principles-based obligation is distributed across the adopted text, subsequent guidance, and consultation documents that show where the interpretation is still developing.

A question asked against only the adopted regulation risks missing the guidance that has since narrowed or clarified how a provision applies in practice. A question asked against a general AI model risks getting an answer that blends adopted law, proposed amendments, and plausible-sounding invention into a single confident sentence with no way to tell which parts are which. Anchoring every answer to a specific, visible, current source within a known stream (whether that stream is binding adopted law, judicial interpretation, interpretive guidance, or pre-legislative material) is the mechanism that keeps the convenience of asking a direct question from becoming a liability, even while cross-stream synthesis remains on the roadmap rather than live today.

Verdandi lets you ask questions directly against the official documents that define EU sustainability law, with every answer anchored to a visible, verifiable source. Start for free.