Using AI as a tax research assistant: what it's genuinely good at

The question CAs ask most often about AI is 'can I trust it?' The more useful question is 'what specific tasks is it reliably good at, and where does it fall short?'

For tax work, the answer is fairly clear once you've used an AI research tool for a month. It's excellent for certain types of questions and unreliable for others. Knowing the boundary is what makes you a better user of the tool.

What the Tax Assistant is built for

The Tax Assistant in DribbleAudit is designed for quick reference questions where the answer exists in the law and a reliable corpus makes it findable quickly:

• GST rate for a specific category of goods or services — with the HSN/SAC code and the applicable notification.
• Definition checks — 'Is this entity an aggregate turnover threshold supplier for composition scheme purposes?'
• Circular references — 'Is there a CBIC clarification on ITC eligibility for staff accommodation?'
• Basic compliance questions — 'Is RCM applicable on legal services received from an advocate?'
• Exemption applicability — 'Is this healthcare service zero-rated or exempt?'

What it's not designed for

The Tax Assistant is not a substitute for the Bench or Counsel features on contested questions. Specifically:

• Cases where the legal position is genuinely unsettled and multiple courts have taken different views — you need the case law research, not a quick answer.
• Fact-specific transaction planning where the wrong answer has significant consequences — you need the full opinion structure with the limitations section.
• Questions where the law is clear but the application to specific facts is non-obvious — 'what is the place of supply for this specific cross-border software licence arrangement' needs reasoning, not a reference.

Why verified sources matter more than fluency

The failure mode of generic AI on tax questions is confident fluency about wrong things. It answers in the precise register of a tax professional, cites the right-sounding sections, and gets the conclusion wrong because it's pattern-matching on training data rather than reasoning from current law.

The Tax Assistant's design constraint is that every answer must cite a verifiable source from the maintained law corpus. If a question can't be answered from the corpus, it says so and routes you to the full research workflow. That constraint makes it less 'conversational' than a general-purpose chatbot — but it means the answers you get from it are ones you can put in writing without checking.

The productivity case

A CA who does GST work answers the same ten questions repeatedly — with slight variations across clients, time periods, and specific facts. What is the rate for this service? Is RCM applicable here? What's the registration threshold for this category? Does this exemption apply?

Having those answers available in seconds with the citation already there — so you can send a client a two-line response with the supporting notification number without going to the portal — is a concrete time saving. Over a week of typical practice, the Tax Assistant handles the quick-reference work and frees the full research workflow for the questions that actually need it.