Blue J

Tax research has a quality that makes it unusually suited to AI and unusually dangerous for it. The questions are precise, the answers exist in a defined body of authority, and there is a right answer — which is exactly the shape of problem a language model handles well.

The danger is that a wrong answer looks identical to a right one — the same reason Harvey AI built its legal product around citations. A confidently stated tax position with no citation behind it is worse than no answer at all, because it invites reliance. And the consequences do not land on the software: they land on the professional who signed the return and on the client who paid the penalty.

Blue J's central design decision addresses this directly. Every answer comes with the authority cited, so the professional can go and read it. That is a smaller promise than answering the question, and a far more useful one.

What Does Blue J Answer?

• Tax research questions, with the authority cited — the product promise is defensible answers complete with verifiable sources, not just fast ones.
• It reasons over licensed professional material: primary authoritative content plus Tax Notes and IBFD, rather than whatever a general model absorbed from the web.
• Three jurisdictions: the United States, Canada and the United Kingdom, each with its own regional site.
• Used by more than 6,000 firms, from sole practitioners through to national accounting practices.
• Four stated use cases — advisory, compliance, tax writing and training — with a 7-day free trial and no card required.

Why Citations Are The Whole Product

Blue J describes delivering defensible answers complete with verifiable sources. Read that as the specification rather than as marketing, because it is what separates a usable professional tool from an unusable one.

A tax professional cannot act on an assertion. They need to know which statute, which regulation, which ruling or which case says so, both because their judgement depends on reading it and because their file needs to show why the position was taken if it is ever questioned.

An AI answer without citations fails that test completely. It might be correct, and there is no way to establish that short of doing the research yourself — at which point the tool has saved nothing.

An AI answer with citations changes the task from research to verification. You are handed a proposed answer and the authority behind it, and your job is to check that the authority says what the answer claims. That is genuinely faster than starting from a blank search box, and it keeps the professional judgement where it belongs.

The failure mode to guard against is subtle. A citation makes an answer feel checked even when you have not checked it, and fluent prose with a section number attached is more persuasive than either alone. Following the citation is the discipline the entire product depends on, and it is the step under time pressure that is easiest to skip.

An uncited AI answer saves nothing, because verifying it means doing the research anyway. A cited one changes the job from research to verification.

Licensed Sources Beat A General Model

Blue J states its answers are powered by primary authoritative content, Tax Notes and IBFD. That sourcing decision matters more than any model choice.

Narrower tools such as TaxGPT compete here too, and general-purpose assistants have absorbed a great deal of tax material, but with three problems a practitioner cannot work around. It is of unknown vintage, which is fatal in a field where a rate or a threshold changed last April. It is of unknown provenance, so a confident answer may be reasoning from a blog post. And it cannot be cited, because the model does not know where it learned something.

Licensing Tax Notes and IBFD means the system is reasoning over material a firm would already subscribe to and already trust — commercial analysis alongside the primary sources, with the currency that comes from a maintained publication.

IBFD's presence signals the international dimension specifically, which is what makes cross-border questions tractable rather than guesswork.

The practical implication for a buyer is to ask about currency rather than coverage. How quickly does new legislation, a new ruling or a budget change appear? In tax, an answer that was right last quarter is simply wrong, and update cadence is the specification that matters most.

How Much Time Does It Actually Save?

The site publishes two figures: three hours saved per user every week, and 75 per cent less time spent on research.

These are plausible in shape and unverifiable in detail, and it is worth being precise about why. Neither carries a stated methodology, sample or baseline. Saved against what — a junior researcher starting cold, or a specialist who already knows where to look? Measured over which mix of questions?

The mechanism, though, is entirely credible. Most tax research time is not spent reasoning; it is spent locating the relevant authority among a great deal of nearly-relevant authority. A tool that puts you within reading distance of the right section removes exactly that, and it is the largest single block of time in the task.

The saving will also be unevenly distributed in a way the averages hide. Routine questions with settled answers should improve dramatically, much as Elicit accelerates literature review without replacing the reading. Genuinely novel or contested positions will improve much less, because the work there is judgement rather than retrieval — and that is the work you would least want automated anyway.

The honest way to evaluate this is to use the free trial on questions you have already researched, where you know both the answer and how long it took you. That comparison is worth more than any published percentage.

• Test on questions you have already answered, so you can grade both accuracy and speed.
• Include one question where the answer is genuinely contested, not only settled ones.
• Follow every citation on at least the first ten answers, without exception.
• Check a question involving a recent change, to test how current the sources are.
• Ask a cross-border question if you handle any, since that is where IBFD should show its value.

Does It Change Who Is Responsible?

No, and this deserves stating plainly rather than as a footnote, because the temptation runs the other way.

Professional responsibility for advice sits with the professional. No software licence transfers it, and no vendor claim about accuracy changes what a regulator, a client or a court expects from the person who signed the work. Using a research tool is like using a treatise: it informs your judgement, it does not replace it.

The practical consequences are concrete. Read the authority the tool cites before relying on it. Record in the file what you relied on and why, as you would with any source. Apply the same review process you would to work produced by a junior — not less, because AI output is fluent, and fluency is not accuracy.

There is a second consideration specific to this technology. Client information entered into a research tool is confidential, and where it goes matters. Blue J publishes a security page and a subscriber agreement; a firm should read both, and should ask specifically whether queries are used to develop the product and where data is held.

None of this argues against the tool. It argues for using it the way the profession already uses every other authority: as an input to judgement that gets checked.

Testing Blue J On Work You Already Did

A 7-day trial with no card required is enough to answer the question properly, provided you test against known outcomes rather than exploring.

The bar is step four. A citation that does not support the answer it was attached to is the one failure that disqualifies a research tool outright.

1. Choose five research questions you completed in the last year, where you have both the conclusion and a rough sense of the hours.
2. Run each through Ask Blue J without looking at your own notes first, so you are testing it rather than confirming yourself.
3. Compare its conclusion with yours. Where they differ, work out which is right — that investigation is the most informative part of the exercise.
4. Open every cited source and confirm it actually supports the proposition it is cited for. This is the check that matters most.
5. Include one question touching a rule that changed in the last twelve months, to test source currency directly.
6. Time one fresh question end to end, including your verification, and compare that with your normal process. That number is your real saving.

Subscribing To Blue J

Access is a subscription with a 7-day free trial and no credit card required, sold to practices ranging from sole practitioners to national firms.

Serving that range usually means tiering by seats and by which content is included, and the second is the variable to examine. Because the value here depends on licensed material — Tax Notes, IBFD, primary sources across three jurisdictions — the question is not what the tool costs but which content your tier actually reaches.

For a firm already subscribing to those publications separately, the calculation is different again and worth doing properly: some of the cost may be replacing spend you already carry rather than adding to it.

The other question to settle is jurisdiction. Coverage spans the US, Canada and the UK, and a firm working across two of those should confirm that both are included rather than assuming.

We do not publish amounts, since terms change and are quoted per firm. Use the trial first — with 6,000 firms as customers, the reference calls are available if you ask for them.

Blue J: Sources And Coverage

Which Practices Blue J Suits

Try the trial when

• You do genuine tax research and the time goes into locating authority rather than reasoning about it.
• You need every answer to arrive with a citation you can put in the file and defend later.
• You already rely on Tax Notes or IBFD, so the tool is reasoning over material you trust.
• You work across the US, Canada or the UK and want one tool covering the jurisdictions you practise in.
• You are training junior staff, where seeing the authority alongside the answer is the teaching.

Not the right tool when

• You practise outside the three covered jurisdictions, where the source material simply does not apply.
• Your bottleneck is producing returns rather than researching positions — that is a different category of product.
• You would rely on an answer without opening the citation, which is the one usage that makes this dangerous.
• Your questions are almost entirely settled and routine, where a good internal precedent file may already serve.
• You need a definitive answer on a genuinely contested position, where the work is judgement and no tool substitutes for it.