Tax teams drown in documents. Advising a single client can mean pulling from thousands of pages of IRS tax code, regulations, rulings, and internal memos — spread across manual uploads and synced Dropbox folders, and revised constantly. Somewhere in that pile is the answer to a simple question.

By Width.ai

Find any answer in a mountain of tax code and client documents in seconds - with complete citations and references

A smarter way to search tax authority and internal documents, after the chatbot platforms that promise it fall over.

The challenge

This client, a tax law firm didn't come to us empty-handed. They came with a POC already built by another vendor: a chatbot over a large corpus of IRS tax code and internal documents used for client work, with manual upload and Dropbox sync for getting documents in, standard chat on top, and a web search tool bolted on.

The chatbot was not working at all:

• Answers were incorrect. Not occasionally off — wrong in ways the attorneys caught immediately, which meant nobody trusted anything it said.
• It couldn't tell current from superseded. A low understanding of date relationships meant the system routinely answered from old versions of documents — outdated guidance and stale memo revisions surfacing as if they were current.
• Complex questions came back empty. Any query with a low semantic relationship to the source text — the everyday reality of tax language, where the question and the controlling provision share almost no wording — either failed or returned the wrong material.
• Web search never fired. Even when users asked for the latest guidance directly, the chatbot wouldn't invoke its own web search tool.

The questions themselves were never exotic. They were the questions a tax practice asks all day:

• Does the §179 deduction apply to the improvements in the Henderson file?
• Which of our clients are affected by the latest §174 amortization guidance?
• Do our internal memo and the current regulations agree on the R&D credit treatment?

Answering them meant falling back to hours of manual reading, brittle keyword searches that return lines instead of answers, or trusting a chatbot that guessed. None of those options is fast, affordable, and trustworthy at the same time.

We came in, diagnosed the issues in the architecture, outlined the key issues stakeholders were seeing in plain terms, architected the new system with them, and built the new AI agent architecture on our multi-agent framework.