Most language models are eager to please. Ask one almost anything and it will produce a fluent, confident reply, even when it has no real basis for it. For a casual question that is harmless. For a chartered accountant advising a client on the IT Act 2025, a confident wrong answer is the single worst thing software can do. It does not just waste time. It can lead to a notice, a penalty, and a loss of trust that took years to earn.
So when we started building FicomAI, we set ourselves a strange goal for an AI product. We wanted it to say "I do not know" more often than a typical model ever would. We wanted refusal to be a feature, not a failure.
The problem with a model that always answers
A general model answers from patterns it learned during training. That training data is frozen at some point in the past, it is a blend of many sources, and none of it is the authoritative text of a brand new Indian law. When you ask such a model about a specific section, it reconstructs something that sounds right. Sometimes it is right. Often it is close. Occasionally it is confidently and dangerously wrong, and there is no way for the reader to tell which is which.
For tax and compliance, close is not good enough. A section number that is off by one, a threshold quoted from the old Act, a provision that was repealed in the new one. These are the small errors that turn into real consequences. The reader needs to know not only the answer, but exactly where it came from.
An answer you cannot trace is not an answer. It is a rumour with good grammar.
Retrieval, not recall
The fix is to stop asking the model to recall, and start asking it to read. FicomAI uses retrieval. When you ask a question, we search the actual text of the IT Act 2025, pull the passages that match, and ask the model to answer using only those passages. The answer then carries the section it came from, so you can read the source yourself in one click.
This changes the whole posture of the system. The model is no longer the source of truth. The law is. The model becomes a careful reader that finds the right passage and explains it in plain language, while always pointing back to the page.
It also gives us a clean rule for honesty. If the search finds nothing relevant in the IT Act 2025, the model has nothing to stand on, and it is instructed to say so. That is where the refusal comes from. It is not the model being shy. It is the model being correct about the limits of what the law actually says.
What an honest refusal looks like
When someone asks about a threshold that simply does not exist in the new Act, FicomAI does not improvise one. It tells you it cannot find the provision, points you to the nearest relevant rule, and reminds you to consult a qualified professional. A reply like that feels less impressive in a demo. In daily practice it is the most valuable thing the product does, because it never sends you down a wrong path with false confidence.
What the market taught us
Before writing much code, I spent a long time simply listening. I spoke with practising professionals across firms of every size, from solo practitioners to teams handling hundreds of clients. The same picture kept appearing. Roughly three hundred and fifty thousand chartered accountants in India were facing a law that had been rewritten overnight, with no tool built for it, and a flood of client questions they were still learning to answer.
Three things came up again and again.
- They were spending hours every week hunting through PDFs for answers they used to know by heart.
- They did not trust a generic AI, because they had already been burned by a plausible answer that turned out to be wrong.
- They wanted to show clients the source, not just an opinion, because their credibility depended on it.
That research did not just validate the idea. It set the rules. Speed mattered, but only if the answer was correct. A beautiful interface mattered, but only if every claim was traceable. The market was not asking for a smarter chatbot. It was asking for a trustworthy colleague.
Why this is the foundation, not a feature
Everything we are building on top of the engine, the compliance calendar, the notice intelligence, the reconciliation agent, depends on this same discipline. An agent that acts on your behalf has to be even more careful than one that only answers, because its mistakes are no longer just words on a screen. They become filings, replies, and reminders that real clients depend on.
So we are starting where trust is built, with a model that reads the law, cites its work, and is comfortable saying that it does not know. If we get that right, everything else has somewhere solid to stand.
That is the whole idea behind FicomAI. Not an AI that knows everything, but an AI that is honest about what it knows, and always shows you where it read it.