Skip to content

How we built an AI assistant for tax and legal regulations

Case study: a RAG system over a huge regulatory database for a large consultancy. Architecture, challenges, phases and results: MVP in 2 months, full system in 5.

Case studies··2 min read

Share
How we built an AI assistant for tax and legal regulations

One of the projects we are most proud of is an AI assistant for tax and legal regulations, built for a large consultancy. The client's name is confidential, but we can share the approach and lessons.

The challenge

Consultants answer complex client questions every day. The answer exists, but it is scattered across a large number of laws, bylaws, ministry opinions and internal texts, with amendments and cross-references. Keyword search did not understand questions asked in natural language, and general AI tools invented regulations, which is unacceptable in this field.

Requirements

  • answers in natural language, in Serbian,
  • a reference to the exact law, article and document with every answer,
  • support for regulation versions,
  • an interface for the internal team and a foundation for later commercial use,
  • data confidentiality.

The solution

We delivered the project in phases:

  1. Knowledge base analysis and architecture: document structure, question types, citation format.
  2. Integration and document preparation: cleaning, splitting by regulatory structure (law, article, paragraph), metadata and indexing for semantic search.
  3. AI layer: hybrid search, re-ranking and answer generation with mandatory citations.
  4. User interface: questions, answers, clickable sources and history.
  5. Testing and optimization: a set of real consultant questions as the quality benchmark.

Key lessons

  • Structure matters more than the model. Splitting regulations correctly by article improved accuracy more than switching models.
  • Citations build trust. Consultants started using the system only once they could verify the source in one click.
  • "I don't know" is a good answer. A system that admits it has no basis is more valuable than one that always says something.

Results

  • MVP for the internal team delivered in 2 months,
  • complete system in 5 months,
  • finding the relevant provision went from tens of minutes to seconds,
  • architecture ready to connect to the Intrascope workspace.

Who this is for

Law firms, tax advisors, auditors, accounting firms, insurers, banks and any company with a large base of regulations or procedures. See the project page and our AI document assistant service, or send an inquiry.

Share

Frequently asked questions

Can AI give legal advice?

The AI assistant helps professionals find and interpret regulations faster, with sources. The final assessment is made by a qualified professional, as defined in the contract.

How are regulations kept up to date?

New and amended regulations are processed and indexed automatically, and the system tracks versions and validity dates.

Tell us what slows you down

We reply within one business day. The first call and the estimate are free.

Send inquiry

Related articles