AI that carries a tax question all the way to the memo

Every tax advisor knows the shape of the afternoon. An hour tracking down the right directive or statute, another hour turning what you found into something a client can read. Two tasks, two tools, and the clock running through both.

Closing that gap is the only interesting thing AI does in this category. Everything else is a faster search box.

Why research and drafting are still two steps

The traditional workflow has not moved: the client asks, you open a database, work through the source, take notes, open Word, write it up. The research tool and the drafting tool have nothing to do with each other.

Plenty of AI products have simply reproduced that split — a research assistant on one side, a drafting tool on the other, a manual handoff in between. The handoff is where the time goes, and where errors creep in, especially across several jurisdictions with different VAT rules, corporate tax regimes and documentation requirements.

What to check before you buy

Citation quality and stability. An answer is worth what its source is worth. Look for the specific statute, directive or official guidance rather than a summary of it, and for citations that stay pointing at the same provision when the instrument is amended.

Whether anything checks the answer before you do. Some tools hand you an answer and leave the checking entirely to you. Better ones show what was searched and read, and run their own audit first — for citation coverage, self-contradiction, provisions amended recently enough to matter, and whether the answer actually addresses what you asked.

File and data ingestion. Can it read the client’s financial statements, prior filings or contracts? If not, you are still doing the extraction by hand.

Output format. A research summary is useful. A Word document you can send is more useful. Check what actually comes out.

Jurisdiction coverage. For European work, coverage of directives, country-specific VAT rules and national implementations is not optional, and most tools built for the US market have thin international depth.

azta

azta is built around the research-to-memo motion as one continuous step. You ask in plain language, it researches the relevant statutes, directives and official guidance across 24 European jurisdictions, and it answers with citations down to the paragraph. From there Canvas produces the summary, the advisory memo or the client email as an artifact the agent edits in place, and full Word documents export to Word or PDF. No second tool, no copy-paste.

In 17 of them the research runs on a structured statute tree rather than flat-text search: statutes are held as documents, units, attributes and the references between them, and a researcher walks that structure — listing a provision’s children, reading a unit, following a cross-reference to the unit it points at. Citations then resolve to a stable legal-unit identity rather than a page offset. The remaining seven run the older flow, which is a genuine difference in depth.

Before an answer is published a verifier audits it — citation coverage full, partial or insufficient; internal contradictions; temporal risk where a cited provision was amended recently; scope alignment with the question asked — and can force more research or ask you to clarify instead of publishing. On a question spanning several countries, a coverage gate blocks the answer until each of them has actually been researched, not merely searched.

File upload means the client’s figures come into the draft instead of being retyped: PDF with real OCR, Word, Excel, PowerPoint, XML and Outlook .msg, cited back with the exact passage highlighted inside the original file. For firms on QUiCK the answers can rest on live invoices, cashflow and VAT positions, and entries can be written back with a confirmation preview first. New legislation, directives and guidance are indexed daily, so the research reflects the rules as they stand rather than a database entry from six months ago.

The free tier is five questions a month — enough to judge research quality before paying for anything — and the Statute Finder, a full browser and reader over the statute book, is free to any signed-in user with no plan gate at all. Pro is €27 a month or €270 a year, self-serve, with no procurement process in the way.

TaxGPT

An AI assistant focused on tax research, positioned primarily around the US market. It handles question-and-answer research and topic summaries. On publicly available information the drafting side is lighter than the research side, and the coverage is US federal and state rather than European or cross-border. For domestic US work that is a reasonable fit; for European VAT or international structures the jurisdiction gap is the whole issue.

Blue J

A Canadian-founded platform that predicts how tax authorities and courts are likely to treat a given fact pattern. It is strong on legal reasoning and precedent analysis and well regarded for structured research output. On public information, research and drafting are distinct phases rather than one workflow — an analysis still needs a separate drafting step — and it is sold with sales-assisted onboarding, which is friction for a small firm.

CoCounsel Tax

Thomson Reuters’ AI layer over Checkpoint, one of the most comprehensive tax research libraries available. It answers research questions and assists with drafting, and the Checkpoint integration is a real advantage for a firm already in that ecosystem. It is positioned as an enterprise product, so the entry path for a sole practitioner is less direct, and European coverage depends on the underlying subscription.

Harvey

A general-purpose legal AI platform with significant traction at large law firms. It does legal research and drafting, and some tax practices use it for tax-adjacent legal work. On public information it is aimed at large enterprise clients and sold through a sales process. Research and drafting both exist, but they are not organised around the tax research-to-memo path the way a purpose-built tax tool is.

How they compare on the workflow

• Plain-language research — all five do this.
• Cited to the statute or directive — azta resolves citations to a stable legal-unit identity in the 17 markets on its structured engine; Blue J and CoCounsel Tax cite primary sources; TaxGPT is grounded in US sources; Harvey cites to source documents.
• An audit of the answer before you see it — azta runs a verifier over citation coverage, contradictions, recent-amendment risk and scope; the others leave verification to the reader.
• Client file ingestion — azta on every plan; the enterprise platforms at enterprise tier.
• Word export from the research — azta via Canvas; CoCounsel Tax and Harvey at enterprise tier.
• European and cross-border coverage — azta covers 24 European jurisdictions, indexed daily, 17 of them on the structured engine; the others are US-centred or depend on the subscription.
• Self-serve pricing — azta at €27/month; the others are sales-led or not publicly stated.

Everything said about the other four reflects publicly available positioning, not terms we have verified. Capabilities and pricing move — check with each vendor before deciding.

European coverage is the deciding factor

For advisors working across European jurisdictions, coverage is not a secondary criterion. VAT rules, mandatory disclosure obligations, anti-avoidance provisions and national corporate tax treatments all need source material well beyond one country’s code, and most tools in this category added international coverage after the fact.

That is what decides whether a tool can answer whether a particular arrangement is reportable, how a specific country applies an interest limitation, or what a cross-border restructuring has to document.

The question worth asking

Does this tool close the gap between the research and the memo, or does it make one of those steps faster and leave the other exactly where it was? Everything else in an evaluation follows from the answer.