ALI is an AI assistant for UK crypto tax that I built as lead developer for Crypto Accountants, a UK firm where I also run IT operations and the web platform. It answers questions from HMRC guidance, quotes the section it relied on and links the exact HMRC manual page, so an accountant can verify the answer instead of redoing the research.
The problem
Crypto tax questions in the UK come up constantly and the answers live in HMRC manuals that are long, cross-referenced and written for specialists. An accountant answering a client has to find the right page, read the relevant section and then write a reply. That is slow, and most of the time goes on finding rather than judging.
A general chatbot does not fix this. It produces confident text, but in tax a confident answer with no source is a liability. The firm needed something that would do the finding and drafting while leaving the judgement, and the evidence, in plain view.
Constraints
- Answers must be checkable. Any statement about the law had to come with a quote and a link to the HMRC source.
- The accountant stays in the loop. The tool drafts; a qualified person verifies.
- Real documents, not just questions. Clients send PDFs, so the assistant had to read them and summarise a case.
- One engineer. I run the main site, forum and academy too, so ALI had to be maintainable by one person alongside everything else.
What I built
Retrieval over HMRC guidance
I built embeddings from HMRC material so that every question first retrieves the passages most likely to answer it. The language model then writes its answer from those passages rather than from whatever it absorbed in training.
Answers with quotes and exact links
Each law answer quotes the relevant section and links the exact HMRC manual page. This is the core feature, not a footnote. It turns the output from "trust me" into "here is where it says so".
Document reading and case summaries
ALI reads uploaded PDFs, summarises the case and drafts a reply. That covers the part of the job that is reading and writing, so the accountant can spend their time on the decision.
Launch content
ALI launched with a demo video series showing real questions and the cited answers, so users could see the verification step before trying it themselves.
How it works
- A user asks a question or uploads a document.
- ALI retrieves the most relevant HMRC passages using the embeddings.
- The model writes an answer from those passages, quoting the section it used.
- The answer includes the URL of the exact HMRC manual page.
- The accountant reads the quote, follows the link if needed, and verifies before anything reaches a client.
Decisions and trade-offs
Retrieval with citations instead of a general chatbot. A general model is easier to ship: you write a prompt and you are done. But it cannot tell you where an answer came from, and it will happily blend outdated or foreign rules into a UK answer. Retrieving from HMRC material first narrows the model to the right source, and forcing a quote makes errors visible. The cost is extra engineering on the corpus and the retrieval step, which I think is the only honest way to build this for a tax firm.
Quote plus link, not just a link. A bare link asks the reader to go and find the relevant paragraph. A quote shows the exact words the answer rests on, and the link proves the quote is real. Together they make checking fast.
Human verification as part of the design. I could have positioned ALI as an automatic answer machine. I chose to make the accountant's check an explicit step in the flow. It sets the right expectations and it matches how a regulated firm actually has to work.
Narrow scope over broad scope. ALI is about UK crypto tax. Keeping the scope tight keeps retrieval relevant and makes it clear to users what the tool is for.
Outcome
ALI is live at cryptoaccountants.ai/chat as part of the Crypto Accountants platform, alongside the main site, forum and academy that I also run. It answers UK crypto tax questions with quoted HMRC sections and exact manual links, reads uploaded PDFs, summarises cases and drafts replies. It launched with a demo video series. I have not published usage figures here because I only state numbers I can stand behind.
What I would do next
- Add an evaluation set of known questions with expected sources, so every change to retrieval is tested against it.
- Track which HMRC pages are cited most, to spot gaps in the corpus.
- Connect drafting to the firm's workflow with automation, for example handing a verified reply to the next step without copy and paste.
If you want a grounded assistant like this for your own documents, see my AI automation and n8n service. For the wider platform ALI sits in, read the Crypto Accountants web platform case study, and for how answer engines pick sources, see how to get cited by ChatGPT and AI Overviews. You can also get in touch.
