AI Tax Research Tools Compared: Citations, Cost and Checks

Pick a tool built on a tax library if research is billable work: Blue J lists $1,498 per user a year and CoCounsel Tax Essentials $1,300 a year for one user. Cheaper options, such as Bizora from $29.99 a user a month or general assistants near $20, need heavier checking. Whichever you choose, open every citation and read the passage.

The price gap between those groups is smaller than it looks once you count checking time. A cheap tool that invents a source costs you an hour of unpicking, and a dear one that cites correctly still needs a quick read of each source. So the useful comparison is not features but the cost per answer you can actually rely on, and the caps matter as much as the prices: CoCounsel Tax Essentials allows 20 Checkpoint-backed answers per user a week, and Bizora's cheapest plan two deep-research questions a day.

Follow me on Instagram@sagnikteaches

The four-point citation check that applies to every tool

An answer with a citation can fail in four different ways. Train staff to ask four questions about every source, in this order, because each catches a different failure.

Connect on LinkedInSagnik Bhattacharya

1. Does it exist? Find the source in your own research library or on the authority's own site, not through the tool's link alone. Invented citations often look perfectly formed.
2. Does it say that? Read the passage. Tools often cite a real source for a point it does not make, or for the opposite point.
3. Is it current? Check the date, and check whether the provision has been amended, replaced or superseded since. A real citation to a rule that no longer applies is still a wrong answer.
4. Is it authoritative? Primary sources (legislation, regulations, rulings, decided cases) outrank commentary, which outranks an unsigned web page. If the only support is a summary, find the primary source behind it.

Here is how each failure looks in practice, using illustrative cases rather than real citations. A tool answers a question about a deduction limit and cites "[Code section] subsection (c)". Check one: your library has the section but no subsection (c). That is a fabricated reference. Check two: the subsection exists but concerns a different type of expense. That is a mismatched source. Check three: the subsection is right but was replaced two years ago. That is a stale source. Check four: the tool cites a firm's newsletter and no primary source. That is a weak source, and the fix is to find the rule the newsletter relies on. The same discipline underlies the checking routine for stopping AI inventing case law, which applies it to legal citations.

Subscribe on YouTube@sagnikteaches

Six research tools: price, trial and citation claims

All figures were read from vendor pages in November 2026 in USD; where a page shows no price the table says so rather than guessing.

The trial column decides your test calendar. Blue J and Bizora both offer seven days without a card, so prepare the ten questions first and run them inside that week. CoCounsel Tax and TaxGPT start with a demo instead, so ask for a trial login of at least a week before you agree to a call that ends in a quote. ChatGPT and Perplexity can be tested on their free tiers, though the paid tiers allow far more use and are the fairer comparison.

Blue J: unlimited questions for one annual fee

Blue J keeps cost simple. Its Individual plan is $1,498 per user a year, which its page compares with roughly one billable hour a week, and its Team plan, for firms with several tax professionals, is by quote. The plan description includes unlimited questions, drafting, file uploads and unlimited in-line citations. It states SOC 2 Type II compliance, and says its sources are curated by tax experts and its primary-source database is updated daily. Those are vendor claims, so test them rather than quote them to clients.

Where it fits: a practice where research is routine and billed, and where a missing or wrong citation costs a client relationship. The seven-day trial gives you time for the ten-question test below, but only if you prepare the questions in advance.

An illustrative question to run: "A client who runs a small business through a company paid for a vehicle used partly for private journeys. How is the benefit treated, and what records should the company keep?" What you would check in the answer: that each rule cited exists in your library, that the tool separates the company's position from the individual's, and that it flags where facts matter, such as mileage records. If it gives a flat answer with no facts-dependent caveats, mark it down. For a focused walk-through of testing this one product against ChatGPT, see how to test Blue J and ChatGPT on your own cases.

CoCounsel Tax: Checkpoint answers with a weekly cap

CoCounsel Tax is Thomson Reuters' AI assistant for tax work. Its product page says research answers are grounded in Checkpoint content, more than 12 million documents, and come with citations you can verify. The entry plan, CoCounsel Tax Essentials, is aimed at small and mid-market CPA firms and is sold online at $1,300 a year for one user, for firms of up to 10 users. It includes unlimited conversations, document upload and templates, but caps Checkpoint-backed answers at 20 per user a week. The plans page says some advanced Checkpoint features, such as deep research tools and state comparison charts, are not included; a Tax Research module adds them at a per-user price the page does not show. No Checkpoint subscription is needed.

The advantage for a practice that already knows Checkpoint is straightforward: the answer links into a library you trust and know how to read. The risk is the cap. Twenty Checkpoint-backed answers a week is about four a working day per user, and a preparer who researches in bursts during the busy season can run out by Wednesday. Count last year's research questions by week before you buy, and ask what the Tax Research module costs for your user count if the cap looks tight.

An illustrative question: "Our client sold a property after six years: four as a rental, then two as a main residence. What rules apply to the gain?" A good answer cites the relevant provisions and flags the facts that decide the outcome, such as the dates of conversion and periods of private use. A weak one states a percentage with no condition. Either way, you open the Checkpoint link and read the section.

TaxGPT: broad scope, quote-only pricing

TaxGPT sells research alongside preparation and review agents. Its pricing page shows no prices: Pro, Pro Plus and Enterprise are tailored to team size and client count, and a demo comes first. Pro includes Research, a Writer tool, Matrix, Agent Andrew and Client Intelligence. Pro Plus adds Cowork, Tax Preparation and a Bookkeeping Agent. The page describes Research as citation-backed and says it covers questions from general rules down to state, local and treaty level, with integrations to common tax software.

For research alone, you may be paying for more than you use. The pricing page says individuals can sign up to use TaxGPT Research on its own, so ask whether a practice can buy just that, since the bundle pushes cost up. An illustrative question that tests the breadth claim: "A client moved from one state to another midyear and has income from a remote employer in a third. Which states can tax what?" A good result names the rules for each state and says where residency tests turn on facts. A weak one picks a state and stops. Run state-level questions on any tool that claims state coverage, because breadth is where citations tend to thin out.

Bizora: the lower-priced specialist with daily caps

Bizora lists three plans per user a month: Essential at $29.99, Pro at $69.99 and Pinnacle at $119.99. On Essential, normal questions are unlimited but deep research, web search, fast search, document uploads and memo writing each have a daily limit of two. A seven-day free trial needs no card. It also offers prepaid API credits, at $0.24 for a fast tax research request, $1.50 for deep research and $0.65 for audit research. The page says answers include citations linked to official materials, which is a claim to verify, not a guarantee.

For a small practice, the daily cap on deep research matters more than the headline price. If you run five complex questions on a busy day, Essential stops you at two. Count your real daily volume before you choose a plan, and the monthly figure of $29.99 can become $69.99 quickly. An illustrative question: "An employer wants to reimburse a remote employee for home internet. What are the tax consequences for both sides?" Check whether the answer addresses both the employer and the employee, and whether its citations lead to official guidance rather than a commentary page.

ChatGPT and Perplexity: cheap, fast and the riskiest on sources

These are the tools many practitioners already use. ChatGPT Plus is $20 a month, and Business is $20 a user a month billed annually or $25 monthly, with a two-seat minimum. Perplexity Pro is roughly $17 to $20 a month depending on billing. Both can search the web and show links. Both can also state a wrong answer fluently, and ChatGPT without web search can invent a citation that looks real.

Perplexity's strength is that each answer comes with links you can click. Its weakness is that the linked pages are whatever the web contains, which is often commentary. ChatGPT with search turned on behaves similarly, and without it the model works from training data that may be stale on recent changes. The practical rule: use them to find angles, vocabulary and candidate sources, then verify in a library you trust. For a wider look at how the two behave on business questions, see Perplexity versus ChatGPT for business research.

An illustrative run shows the pattern. Ask Perplexity whether a client can deduct a home office used two days a week, and a typical answer gives a confident "yes, if conditions are met" with five links: two accounting-firm articles, a software vendor's help page, a forum thread and one page of official guidance. Only the last is authority. Open it, find the passage that sets the conditions, and compare it line by line with the summary; where the official text is narrower, the summary loses. Ask ChatGPT the same question with search switched off and it may cite a ruling by number, which is your cue to look that number up in your library before reading another word.

Privacy needs a separate line. Business plans such as ChatGPT Business do not train on your content by default, whereas consumer plans have settings to review, and Perplexity's data retention option is switched on by default on its Free and Pro plans. Check the settings before any client fact goes near a consumer account.

Does your existing research subscription already include AI?

Before you buy anything, ask the vendor of your current research library. Several have added AI assistants to existing platforms, and a bolt-on that costs a few hundred dollars may beat a new standalone tool. The questions: is the AI included or an add-on, does it search only the library you already subscribe to, and does every answer carry links into the library? An illustrative sum: if your library renews at $2,400 a year and the AI add-on is $600, then $3,000 buys one login, one billing relationship and citations into content you already trust. Compare that with $1,498 for a standalone tool plus the library you would still keep.

Ten past questions run through four kinds of tool

The only test that settles the choice is your own past work. Take ten research questions from the last two years where you wrote a memo or email and still hold the file. For a small practice, a good set includes two simple questions, three that depend on facts, two where the law changed since you answered, two with an unusual fact pattern and one you know contains a trap. Remove client names and exact figures, and run each question, identically worded, through each tool in its trial.

An illustrative result for a three-partner practice, scored against the original memos:

Read it as a pattern, not as a ranking of products. The library-backed tool gets more right and takes the least time to verify, so its higher price may pay back. The no-search assistant invented eight of its 19 citations, which means that most of the time you save on drafting is spent unpicking. The two middle rows are where the decision hides: a tool that is right 8 times in 10 with a 9-minute check might be fine for first-pass research and poor for final advice. Your own ten questions will produce different numbers, and that is the point.

A scoring sheet you can copy and fill in

Use one row per question per tool. This is one filled-in row for an illustrative question about reimbursing an employee's home internet:

Question 4 | Tool: [name] | Date run: [date]
Conclusion matches memo: Yes
Citations given: 3
  1. exists: Yes | supports: Yes | current: Yes | authority: primary
  2. exists: Yes | supports: Partly | current: Yes | authority: secondary
  3. exists: No  | supports: n/a    | current: n/a | authority: n/a
Caveats given on facts that matter: 2 of 3
Minutes to verify: 11
Overall: usable as a lead, not as authority

The key columns are the four checks plus the caveats the tool raised unprompted. A tool that never says "it depends on" is claiming more certainty than tax law ever offers.

Cost per answer you can rely on

Seat prices mislead because they ignore verification. Work out cost per verified answer for each tool using your own volume. Assume 150 research questions a year and a loaded staff cost of $45 an hour, and use the minutes from the example table.

• Library-backed tool at $1,498: $9.99 per question in subscription, plus 6 minutes of checking at $0.75 a minute, which is $4.50. Total about $14.49.
• CoCounsel Tax Essentials at $1,300: $8.67 per question in subscription, plus the same 6 minutes of checking if it performs like the library-backed tool in the table. Total about $13.17, and 150 questions a year, roughly three a week, sits well inside its weekly cap.
• Mid-priced specialist at, say, $69.99 a month ($840 a year): $5.60 per question, plus 9 minutes ($6.75). Total about $12.35.
• General assistant at $240 a year: $1.60 per question, plus 14 minutes ($10.50). Total about $12.10.

The totals sit close together, which surprises people. The cheap tool wins on paper only if its answers survive the check. When a citation fails, you pay again: another 20 to 30 minutes in the library to find the right authority. If a tenth of the cheap tool's answers need that rework, add $2.25 per question and its advantage is gone. The honest conclusion for most practices is that price is a weak tiebreaker and checking time is the real cost, so choose the tool that keeps your checking minutes lowest on your own questions.

A prompt that makes any assistant easier to check

If you use a general assistant, make it show its work in a form you can verify. Here is a prompt and an illustrative output.

Question: [paste, with client details removed].
Answer in this format: 1) short conclusion, 2) the rule in one
sentence, 3) sources as a list with exact name, section and date,
4) facts that would change the answer, 5) what you could not find.
If you are not sure a source exists, write "unverified". Do not
invent section numbers.

A typical response gives a three-line conclusion, then two sources, one of them marked "unverified", and a list of three facts that matter. What you would fix: the assistant often fills "section and date" with plausible numbers, so verify those first. An "unverified" label is useful because it tells you where to start, but the absence of that label proves nothing. Treat the format as a way to speed up your check, not a substitute for it.

Four ways AI research goes wrong in a real practice

A realistic mistake, sketched as an illustration: a preparer asks an assistant about a credit and gets a clean answer citing a ruling. The ruling exists, but a later provision has replaced it. The client claims the credit on the old basis, and the correction costs two amended returns and a difficult conversation. The check that would have caught it takes ninety seconds: look up the provision's amendment history. Make that step part of the routine, not an extra.

From research note to client wording

Research output becomes advice when it reaches a client, and that is where wording matters. A before and after:

Before (tool output pasted into an email)
"Yes, this expense is fully deductible under the applicable rule."

After (checked and rewritten)
"On the facts you gave me, this expense appears to be deductible
to the extent it is used for the business. I have confirmed the
rule in [source] and noted the condition about records. Please
send the usage log for the year so I can confirm the amount."

The "after" names the source, states the condition and asks for the missing fact. That is also what a client file should show if anyone later asks how you reached a view. For drafting longer advice, drafting client advice memos with AI safely sets out a structure. For turning a conclusion into plain words for a client, see how accountants use AI to explain tax in plain English.

What to keep out of the prompt

Research questions can usually be asked without identifying anyone. Remove names, identification numbers, addresses and exact figures, and describe the facts in general terms. Round figures where the rule depends on a threshold, and keep only the numbers the rule needs. Use the business or enterprise terms that exclude training on your content, check the vendor's retention period, and update your written information security plan to list each tool and who may use it. If your engagement letter does not mention AI tools, discuss it with a qualified adviser before you send anything client-related to an outside service. A simple rule for staff: if you would not put it in an email to an outside party, do not put it in a prompt.

Which tool for which practice

• A sole practitioner who researches a few times a week: start with the free or seven-day trials, run the ten-question test, and consider a general assistant for first drafts plus a verified library for the final read.
• A two-to-five person practice with regular research billed to clients: a library-backed tool is more likely to pay back, since checking time is the biggest cost. Blue J and CoCounsel Tax Essentials both publish prices, so both are easy to model; CoCounsel suits a practice that already trusts Checkpoint, as long as 20 Checkpoint-backed answers per user a week covers your busiest weeks.
• A practice with heavy state and multi-jurisdiction questions: test the breadth claims of TaxGPT and Bizora on your hardest cases before you pay.
• A practice with a research subscription already: ask what AI it includes before you add another login.

Whatever you choose, schedule a repeat of the ten-question test every six months, with new questions, because tools and tax rules both change. For how research fits into the rest of a season, see how small tax practices use AI through the busy season. If you would rather design the test with someone, my AI implementation consultation covers it, and the AI guide links the rest of this series.

Questions about AI tax research and citations

Is a citation proof that an answer is right?

No. A citation shows where the tool says its answer came from. It may not exist, may not say what the tool claims, may be out of date or may be a weak source. Open each one and read the passage before you rely on the answer.

Can a general assistant do tax research safely?

It can help you find angles and draft questions, but it is the tool most likely to invent a source. Use it with web search turned on, demand links, and verify each link in your own research library before anything reaches a client.

How many past questions make a fair test?

Ten questions with a memo on file is a workable minimum. Include a few where the law changed, a few with an unusual fact pattern and one you know has a trap. Ten straightforward questions will make every tool look good.

Should client facts go into these tools?

Strip names, identification numbers and exact figures where you can, and use business plans or enterprise terms that exclude training on your content. Check the vendor's retention terms and your engagement letter first. Ask a qualified adviser if your rules on third-party disclosure are unclear.

Further reads

• Tax Software With AI: What a Small Business Should Check — What a small business should check in tax software with AI.
• AI Software for Tax Professionals: Run a Pilot and Measure Cost per Return — Run a pilot and work out cost per return.
• AI Tax Prep Software for CPAs: What to Test Before You Switch — Test preparation tools on ten filed returns.
• AI for Tax Audit Support: Document Requests, Timelines and Draft Replies — Draft audit responses with a checking routine.
• How to Set Up Human Review for AI Work Without Slowing Down — Build review into AI work without slowing down.
• Can AI Do My Business Taxes? What It Can Safely Prepare — What AI can safely prepare and what it cannot.
• Best AI for Accountants in Australia: Tools, Privacy and ATO Checks — Which AI tools suit an Australian accounting practice, what the Privacy Act and TPB guidance expect, and the human checks before anything reaches the ATO.
• AI for Attorney Research: A Citation-Checking Workflow — A repeatable way to turn AI research suggestions into checked authorities before a draft reaches the court.
• AI for UK Tax Advisers: Research Notes, Client Letters and Checks — Set a safe-use line for each kind of output a UK tax adviser produces: research notes, client letters and computations, with the checks each needs.
• How Accountants Can Use ChatGPT Effectively: Review Notes and Memos — Redact first, then use three prompt patterns for review notes, technical memos and client replies, with a source-check routine for every draft.
• Tax Workpaper Prep With AI: Tie the Return Back to Source Documents — How AI can draft a source-to-return tie-out schedule for a tax return, and the reviewer checklist that proves each figure before sign-off.
• Can AI Draft an IRS Notice Response? What to Check Before Sending — A practitioner's routine for AI-drafted IRS notice responses: triage by type and deadline, redact first, verify every code, amount and date, then sign.
• Can AI Summarise IRS Transcripts for a Tax Resolution Case? — A small-practice workflow: redact an IRS transcript, turn its codes into a plain-English case summary, check it, and keep the client updated on a fixed cadence.
• Canopy AI: Smart Tax Prep, Document Renaming and Beta Features — What AI Canopy offers a tax practice, which features carry a beta label on its pricing page, and how to test each before you pay.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: the pricing and product pages of Blue J, Thomson Reuters CoCounsel Tax (product and plans pages), TaxGPT, Bizora, ChatGPT and Perplexity, read in November 2026.