Ravical, the organisational AI platform for full-service accounting and tax firms, is introducing Domain Knowledge: work that relies on legislation is now grounded in the primary source, not the model's memory or a search result.

When work in Ravical touches tax law, legislation or regulatory guidance, the agent does not answer from memory. It retrieves the relevant text from the sources Ravical indexes itself and verifies every claim it makes against the primary source.

A client told us about this recently. They had a question about setting up a liquidation reserve and asked an AI tool for help with the tax rule. They got back a confident answer. The rule had changed recently, and the answer applied the old version. The tool did not know its own information was out of date, so it never said so. That same failure mode was possible for us too. Answers used to rest on two things: the model's own knowledge, and web search. Neither was verified. Web search returns commentary written about the legislation rather than the legislation itself, so errors carried through. The failures that slipped past were the hard ones to spot: a citation pointing to the wrong paragraph, a percentage restated incorrectly. So we changed how the work gets grounded. We now index the primary sources ourselves.

Benjamin Vandermarliere, CPO at Ravical

Verification

Verification is the point of the domain knowledge. In professional work, an answer that is almost right is the hardest kind to catch, so it is not enough for the agent to sound right. It has to be able to prove it. Ravical handles this in two steps.

The agent verifies its own work

Any agent working on a client file can call the knowledge agent when it needs domain knowledge. The knowledge agent drafts its work, then for each citation in that draft it retrieves the article from Ravical’s index and compares the claim against the source text. Claims that hold stay. Claims that do not are dropped, or the agent searches further. The loop runs until the work settles.

You verify it yourself

Every claim comes with a link to the article it was verified against. Click Open source to land on the exact page, and in longer texts, the right paragraph. The source text is also readable inside the session, so checking a single claim does not mean leaving your work.

Types of sources

Primary sources

Original legislative and regulatory texts, held in Ravical’s own index rather than retrieved through web search. This matters because many of these texts are hard to reach through search at all: some official sites block search engines or load their content dynamically, so a third-party copy of the text ranks above the original. Others are large enough that retrieval fails part way through. Indexing them directly means the agent works from the text itself.

Secondary sources

The reference works and professional handbooks your market treats as standard. These are indexed and used for verification alongside the primary texts.

Recently changed topics

Subjects where the rules have moved recently. A model’s own knowledge goes stale, and the failure mode is a confident answer with no verification behind it. When a question touches one of these topics, the agent is required to verify rather than answer directly.

Speed was never the hard part

The value of AI in a firm was never that it works quickly. Work delivered fast but wrong is worse than no work at all, because it carries the confidence of work that is right. What makes the work usable is knowing it can be trusted, and being able to see why. Knowledge & Verification is how Ravical earns that at the level of a single claim: the agent proves what it says against the law itself, and hands the adviser the source to check. The firm keeps its judgement. Ravical makes sure the ground under it is solid.

Domain Knowledge is automatically live today for all clients in Belgium, the Netherlands and UK, on all tiers. Speak to your Customer Value manager if you’d like more info.