RAG for Accountants and Tax Advisers: Querying HMRC Guidance With Sources

Accountants and tax advisers spend hours each week hunting through HMRC manuals, Finance Act updates and internal procedure notes to answer a single client question. Retrieval-augmented generation (RAG) lets you ask that question in plain English and get an answer grounded in your own documents, with the exact passage cited, so you can verify it before it reaches a client. Done properly, it also solves the yearly headache of keeping guidance current after each Budget and Finance Act. This guide covers what to index, how to keep it fresh, and the compliance points a UK practice needs to check first.

Why a generic chatbot is risky for tax work

Ask a general-purpose AI model a tax question and it will usually give you a confident, well-written answer. The problem is that it has no way of telling you whether that answer reflects the current tax year's rates, a superseded version of an HMRC manual, or a rule that only applies in a different jurisdiction. For a profession built on citing the right paragraph of the right manual, that is not a minor inconvenience, it is a professional indemnity risk. A RAG system flips the logic: instead of asking the model to recall tax law from memory, you ask it to search a defined set of documents you control and answer only from what it finds there, with the source passage attached.

This distinction matters more in tax than in almost any other domain, because thresholds, rates and reliefs change every year, and because two manuals can appear to say similar things while actually applying to different circumstances. A grounded assistant that quotes the specific paragraph of, say, the Business Income Manual lets a trainee or a client-facing adviser check the source in seconds rather than trusting an unverifiable summary.

What belongs in a tax knowledge base

The most useful practice knowledge bases combine public HMRC guidance with the firm's own internal know-how. Neither on its own is enough: HMRC manuals tell you the official position, but your internal notes tell you how the firm actually applies that position to real engagements.

HMRC guidance is vast and fragmented

HMRC publishes dozens of internal manuals on GOV.UK, covering business income, capital gains, VAT, employment income, corporate intangibles and more, each running to thousands of pages and cross-referencing legislation, case law and other manuals. Finding the right paragraph by browsing is slow, and keyword search on GOV.UK often misses guidance phrased differently from the client's question. Uploading the relevant manual sections as PDF or text files into a RAG knowledge base means a question like "can a sole trader claim mileage for travel between two regular workplaces" retrieves the actual manual passage, not a paraphrase.

Internal know-how is just as valuable

The second layer is everything that lives in your firm's heads and shared drives: technical memos written after a tricky engagement, the practice's house style on how to treat a particular relief, precedent client letters, and notes from CPD sessions. This content rarely gets reused because nobody remembers which folder it is in. Indexed alongside the public guidance, it becomes searchable in the same query, so a junior can ask one question and get both the official HMRC position and "how we normally handle this for clients like X."

• HMRC manuals relevant to your client base (e.g. Business Income Manual, VAT guidance, Capital Gains Manual, Employment Income Manual)
• Finance Act text and explanatory notes for the current and prior tax year
• Professional body guidance from bodies such as ICAEW, ATT or CIOT
• Internal technical memos and precedent letters
• Client-agnostic procedure notes: engagement checklists, sign-off processes, filing deadlines
• Past query logs and FAQ documents compiled from client questions

Keeping the knowledge base current through the tax year

A tax knowledge base that is accurate in April and stale by October is worse than no knowledge base at all, because staff start trusting answers that are quietly out of date. The fix is a refresh cadence tied to the UK's actual tax calendar rather than an arbitrary schedule.

Two habits make this manageable. First, version every document with the tax year it applies to, so a query about a 2023-24 return can be pointed at archived guidance while a query about the current year uses the latest upload. Second, assign one person per tax head (VAT, employment, corporation tax) to own the quarterly check, rather than leaving it to whoever remembers. For teams building chunking and indexing pipelines themselves rather than using a managed service, our guide to chunking strategies for RAG covers how to split long manuals without breaking up a rule and its exceptions across separate chunks.

Why citations matter more in tax than almost anywhere else

An answer with no source is a liability in this profession. If a client query ends up being challenged, the file needs to show which guidance was relied on and when it was current. A RAG system that returns the cited passage alongside the answer gives you that audit trail automatically: the adviser can see the exact wording from the manual or the Finance Act, check it against the question, and attach it to the file note. This is quite different from asking a general chatbot and then trying to reconstruct, after the fact, which rule it might have been thinking of.

If you want to understand how to systematically check whether an AI answer is actually faithful to its sources rather than subtly misquoting them, our piece on evaluating a RAG system walks through retrieval metrics and faithfulness checks you can apply before rolling a tool out firm-wide. For a shorter, practical checklist on reading citations yourself, see checking an AI answer's citations.

A practical workflow for building this

You do not need an in-house engineering team to get a working tax knowledge base. On Kopik, a practice can upload HMRC manual PDFs, internal Word documents and Markdown procedure notes directly, and the platform handles extraction, chunking and hybrid search (combining keyword matching with meaning-based retrieval) so that a question phrased informally still finds the right paragraph of dense legislative text. Hybrid search matters specifically for tax content because queries often include exact terms (a specific relief name, a statutory reference) that pure semantic search can miss; our explainer on hybrid search for RAG covers why combining both approaches beats either alone.

A sensible setup for most practices is a private knowledge base, visible only to the firm and queried through the firm's own API keys, so client-adjacent internal notes never leave the practice's control. You can create one for free from the new base screen, upload your manuals and procedure documents, and start querying straight away, either on the website, through the REST API for integration into practice management software, or through an MCP server so staff can ask questions directly from tools like Cursor or ChatGPT without leaving their usual workflow.

Some sole practitioners and specialist consultants go a step further and make a narrow, well-maintained knowledge base public on the Kopik catalogue, charging other professionals per question for access to deep expertise in, say, R&D tax relief or VAT on property transactions. If that interests you, our articles on selling expertise to AI agents and selling expertise online as a paid knowledge base explain how pricing and revenue share work.

Compliance: UK GDPR and professional conduct

Tax files are full of personal data, National Insurance numbers, salary details, property transactions, so any knowledge base that includes real client correspondence needs the same UK GDPR discipline as the rest of your practice: a lawful basis for processing, a record of what is stored where, and a clear retention policy. Keeping client-identifying material out of the shared knowledge base, and instead storing anonymised technical notes ('a client in this situation asked whether...'), avoids most of the risk while keeping the institutional knowledge. Our UK GDPR checklist for RAG knowledge bases is a useful starting point before you upload anything containing client data.

1. Decide which documents can be anonymised rather than uploaded with client names attached
2. Keep the knowledge base private unless you are deliberately publishing a client-agnostic technical resource
3. Document who in the firm can add, amend or remove sources, and log when each document was last reviewed
4. Check your professional body's position on AI-assisted advice (ICAEW and CIOT have both published guidance on AI use in practice) and keep a human sign-off step for anything client-facing
5. Budget for the ongoing cost of re-indexing updated guidance each tax year rather than treating ingestion as a one-off project; see what RAG actually costs for a breakdown of ingestion, storage and query costs

Getting started without overengineering it

The practices that get the most value tend to start small: one tax head, one well-maintained set of manuals, and a handful of internal procedure documents, rather than trying to index the entire GOV.UK guidance library on day one. Once staff trust the citations on that narrow set, expanding to cover VAT, employment taxes or trusts is a matter of uploading more documents, not rebuilding anything. The discipline of versioning by tax year and assigning an owner per tax head matters far more to long-term accuracy than the choice of underlying technology.