Best AI Tools for US Tax Consultants: Research, Notices, Letters

A US tax consultant can use a general chat assistant (ChatGPT, Claude, Gemini or Copilot, roughly $20 a month) for drafting letters and summarising notices, and a tax-specific research tool (such as Blue J at $1,498 a year per user, or CoCounsel Tax) when answers must cite current authority. Start with the consent and review rules, not the price.

Most tax firms choose the wrong way round: they compare features, then discover that a client's return data cannot go into the tool they liked. Two rules shape the shortlist. Section 7216 governs what you may do with return information, and Circular 230 holds you responsible for the quality of what goes out under your name. The tools below are judged against those two rules first.

Follow me on Instagram@sagnikteaches

Two rules that cut the shortlist before price does

The first rule is Section 7216, the regulation on how a tax return preparer may disclose or use a taxpayer's return information. In the text of 26 CFR 301.7216-3, a preparer may not disclose or use return information without the taxpayer's written consent unless another provision allows it. The consent has to be signed and dated, name the purpose and the recipient where there is a disclosure, and be given before the use. The regulation says that making services conditional on consent makes the consent involuntary, and the default life of a consent is one year unless the document says otherwise. The page does not mention software or artificial intelligence at all, so it does not tell you whether a given AI tool counts as a disclosure. That question turns on the tool's terms, where the data goes and who can see it. Put it to your professional body or a tax lawyer, and keep the answer on file.

Connect on LinkedInSagnik Bhattacharya

The second rule is Circular 230. The IRS Office of Professional Responsibility describes it as setting standards of competency, diligence and ethical conduct for lawyers, CPAs and enrolled agents. It says nothing specific about AI, but the practical reading is plain: if an assistant invents a citation and you send it to the IRS, the problem is yours. That reading is why every tool in this tutorial is placed behind a human review step.

Subscribe on YouTube@sagnikteaches

From those two rules come four screening questions for any tool, in this order:

1. Is client data used to train the vendor's models, and can that be switched off at account level rather than by each user?
2. Where is data stored, for how long, and what happens on cancellation?
3. Is there a published security report (SOC 2 is the usual one) you can read?
4. Can you give the tool anonymised material and still get a useful result?

The last question matters most for general assistants. A research question about a rule needs no client name, no social security number and no dollar amounts. A letter draft needs only the facts you choose to supply. A habit of redacting before pasting turns a risky tool into a safe drafting tool for most jobs, and it does not depend on the vendor's promises.

The field in one table

The prices come from the vendors' own pricing pages, read in November 2026. CoCounsel Tax's calculator did not show a figure when I checked; earlier checks put Tax Essentials at about $1,300 a year for one user, so treat that number as one to confirm with a quote. Annual list prices move between seasons, and TaxDome changed its whole pricing model during 2026, so check anything you will pay for on the vendor's page.

Where a general assistant earns its $20

A general assistant is good at turning messy input into structured output. Four jobs justify the subscription on their own: summarising a notice into plain language, drafting a first-pass client letter, building a document request list, and rehearsing the other side's likely questions before a call.

Reading a notice before you open the return

Paste the notice text with identifying details removed (replace the taxpayer name with "Client A", the identification number with "XXX" and keep the dollar figures only if you must). Then ask for structure, not conclusions.

You are helping a tax consultant. Below is the text of a notice sent to a
client, with identifiers removed. List: (1) what the notice says it has
changed or proposes, with the dollar amounts as written; (2) the response
deadline as written; (3) each document the notice asks for; (4) three
questions I should answer from the return before replying. Do not give a
legal opinion and do not cite any rule you cannot name precisely.

[paste notice text]

An illustrative result: the assistant returns a four-line summary (the notice proposes an extra $3,420 tax and a penalty of $684 for income it says was not reported), a deadline of 30 days from the notice date, a list of two requested documents, and three questions such as "Was the $17,100 on the third-party form already included on the return under another line?". What you fix: it may state the deadline from the wrong date if the notice carries two dates. Always read the original.

Drafting the reply and the cover letter

Once you have checked the figures yourself, ask for a draft in the tone your firm uses. Give it the facts as bullet points and tell it what not to say.

Draft a one-page reply to a tax authority notice in a formal, polite tone.
Facts: the $17,100 reported on the third-party form was included on the
return as part of line 8 (total of two amounts: $12,400 and $4,700). Enclosed:
a one-page reconciliation and a copy of the third-party form. Do not argue
about the penalty yet; reserve the right to request abatement. Leave
placeholders for the notice number, the date and the signature block.

The output is a clean four-paragraph letter with a reconciliation table and bracketed placeholders. What you fix: any sentence that sounds more certain than your evidence ("the amount was clearly included"), and any phrase that creates an admission you did not intend. Letters are where a human read pays back.

Client letters that sound like a person

For client-facing letters (engagement renewals, missing-document reminders, an explanation of a balance due), a general assistant gives the biggest time saving. A before-and-after shows the point. Before: "Please provide the documents requested at your earliest convenience so that we may proceed with the preparation of your return." After, from the assistant with your firm's tone note: "To finish your return we still need three things: your year-end brokerage statement, the mortgage interest form and the childcare provider's receipt. If anything is missing, tell us on Thursday and we'll plan around it." The second version names the items, sets a date and gets answered faster. For more on that, see how accountants explain tax in plain English with AI.

Where a tax-research tool is worth the jump

The gap shows when you ask about a specific rule and need to know where the answer comes from. A tax-research product is built on a curated body of tax material and returns answers with links into that material. A chat assistant works from its training and whatever it finds on the web, and it may produce a convincing citation that does not exist.

CoCounsel Tax

CoCounsel Tax is Thomson Reuters's tax assistant, and Tax Essentials is its entry tier. The plans page says each user gets up to 20 citation-backed tax answers a week, grounded in Thomson Reuters's Checkpoint content, while document uploads, summaries and ordinary AI conversations do not count against that allowance. No Checkpoint subscription is needed, but Essentials does not include the Checkpoint Edge research library; deeper research is a Tax Research module priced per user. The page also cites SOC 2 Type II controls.

The weekly cap matters in practice. A sole practitioner asking eight research questions a week in a quiet month stays well inside it, but in March the same person may reach 20 by Wednesday. Because summaries of uploaded documents sit outside the cap, a sensible habit is to spend the cited answers on questions of law and use uploads for reading notices and client papers. Ask the vendor how extra answers are handled before you rely on it in busy season, and get a written quote, since the online price did not display for me.

Blue J

Blue J's pricing page lists an Individual plan at $1,498 a year per user, aimed at sole practitioners, with full US federal and state-and-local coverage, unlimited questions and drafting, and a SOC 2 Type II report. The Team plan is "talk to sales", adding onboarding, firm-wide usage insights and single sign-on. A seven-day free trial with no credit card is offered, which fits the test described later in this tutorial. The "unlimited questions" line is the contrast with a weekly cap: it favours firms with steady research volume.

What a cited answer looks like, and how you check it

Ask a research tool a question such as "Can a sole proprietor deduct a home office if the space is also used for storage of inventory?" and the useful part is not the sentence. It is the list underneath: the code section, the regulation, the case, each linked. Open each link, find the sentence the answer relies on, and confirm it says what the answer says it says. A cited answer that you have not opened is no better than an uncited one, and Circular 230 holds you to the diligence either way.

Practice management tools with AI added

TaxDome is not a research tool; it runs the work around the return. Its pricing page now charges by active client rather than by seat: one plan starts at 150 active clients for $1,500 a year (about $0.83 a client a month), includes 20 seats and is paid annually upfront, with no monthly option. An active client is anyone with activity in the past 12 months, and one person with a personal return and a partnership return counts as two. The page's own example puts 400 active clients at $4,125 a year. The AI features (document organisation, AI return delivery, agents) are a separate add-on at $200 a month for the whole team, and the page describes the company as SOC 2 Type II certified. Its value for AI purposes is that document requests, reminders and client messages live in one place with an audit trail, which makes your consent forms and engagement letters easier to prove. If you already run a portal and workflow product, check whether the AI features inside it cover your letter drafting before paying for a second subscription.

A worked comparison: one notice, three routes

An illustration with invented numbers. A five-person tax practice receives a notice proposing $3,420 of extra tax on a small-business client. The partner wants to compare three routes to the first draft reply.

The sum behind the last row: $1,498 divided by 12 is about $125. The point of the example is the review column. The assistant saves most of the drafting time and adds checking time, so the net saving is roughly 20 to 30 minutes a notice rather than 40. If the practice handles 30 notices in a season, that is about 10 to 15 hours. Whether that justifies a research subscription depends on whether the same tool also answers the harder questions, which is a different calculation.

Section 7216 in practice: a consent routine for AI use

Your lawyer decides what your consent has to say. What the routine below does is make sure it gets asked, signed and filed.

1. Decide, per tool, whether return information ever enters it. Write the decision in a one-page firm policy: "assistants receive redacted text only" is a clean rule.
2. If a tool will receive identifiable return information, have the lawyer-approved consent text in the engagement letter or a separate form. The regulation expects a signed and dated consent that names the purpose and, for disclosures, the recipients, and gives the client a copy at signing. For individual (Form 1040) clients, a use and a disclosure need separate consent documents; for other taxpayers the regulation allows the consent to sit in an engagement letter, provided it meets the same content rules.
3. Never make the service depend on the client agreeing. The regulation treats conditioned consent as involuntary, with a narrow exception for preparer-to-preparer arrangements.
4. Diary a one-year check, since the default effective period is one year.
5. Record each tool's training and retention settings with the date you checked.

A realistic mistake: a preparer asks an assistant to "check this return for errors" and pastes a full return including names, addresses and numbers. No consent was in place and the account was a personal plan with training switched on. The fix is not clever prompting; it is the policy above plus a redaction template next to the prompt.

Redaction checklist before pasting into a general assistant
1. Replace the client name with "Client A".
2. Delete addresses, phone numbers, dates of birth.
3. Remove all identification numbers and account numbers.
4. Round or keep only the dollar figures the question needs.
5. Re-read the pasted text once before pressing send.

A one-week test before you pay for anything

Run the same five tasks through two tools and score each result. Use real past work with the identifiers removed.

Score each out of five and note the time. A tool that fails the fourth task is telling you how it behaves when the law has moved, which is when you most need to catch it. For a fuller pilot structure, see the pilot plan with cost per return, and for a pre-purchase test of preparation software what to test before you switch.

A research memo, start to finish

Research is where the two tool types differ most, so here is one question run through the same routine in either. The question, illustrative: a client who runs a small landscaping business bought a $46,000 truck in November and asks whether the whole cost can be deducted this year. The consultant wants a short memo, not a chat reply.

Write a research memo outline for a tax consultant. Question: can a
small business deduct the full cost of a vehicle bought late in the year?
Give: (1) the rules I should look up, named by topic only, not by number
unless you are certain; (2) the facts I must confirm with the client
(business-use percentage, vehicle weight, placed-in-service date);
(3) the two or three ways the answer could change; (4) a list of primary
sources I should open before relying on any of this.

A reasonable illustrative output lists topics (first-year expensing, bonus depreciation, limits on passenger vehicles, the effect of business-use percentage), lists the client facts to gather, and ends with a checklist of sources. Notice what you asked for: no conclusion, no numbers from memory. A research tool does the same job and adds links beside each topic. Your next step is the same in both cases: open the sources, read the current text, and write the memo in your own words with the client's facts. The assistant has saved you the blank page, and the research tool has also saved a search.

A second, shorter use: ask the assistant to play sceptic. "Here is my draft memo. List the three weakest points and what a reviewer would ask." It will not know the law better than you, but it will often catch an unstated assumption, such as a use percentage you assumed rather than confirmed.

How out-of-date answers show up, and how to catch them

The most expensive failure is not an invented citation; it is a correct-looking answer about a rule that has since changed. A realistic case: an assistant explains a threshold with confidence, the figure was right two years ago, and you cite it in a client letter. Three habits catch this. First, ask every tool "as of what date is this answer current?" and treat a vague reply as a warning. Second, for any figure that has a dollar limit or percentage, find the number on the tax authority's own page before it goes in writing. Third, keep a short list of the rules that change most often in your client base and re-check those from the source every season, whatever the tool says.

The comparison table's last column ("weakest point") is the guide here. A chat assistant's weakness is citations; a research tool's weakness is coverage, so if a client's issue sits in an unusual corner (a niche credit, an uncommon entity type), ask whether the tool covers it before you rely on a short answer. A research tool that returns a confident answer with thin sources deserves the same suspicion as a chat assistant.

Vendor questions to send before you sign

Send the same short list to every vendor, including those you already use, and keep the replies with your engagement records. Treat a refusal to answer in writing as an answer.

• Is customer content used to train or improve models, on every plan we would buy? Is that set by default or by each user?
• Where is data stored, and for how long after we delete it or cancel?
• Can you share a current security report, and what does it cover?
• Which sub-processors handle our data?
• If a client asks us to delete their information, how do we do that inside your product?
• What happens to our saved prompts, files and history if the price changes and we leave?

For a general assistant on a business plan, the answers are usually on the vendor's trust or privacy pages. The first question has a simple pattern: business plans of the big assistants generally do not train on business content by default, while consumer plans let users switch it off in settings, so the plan you buy matters as much as the vendor. Describe the setting generically in your policy ("training use switched off") rather than quoting a label that may change.

Three small firms, three sensible setups

An illustration to make the choice concrete. A solo preparer with mostly individual returns buys one individual assistant plan at $20 a month, writes a redaction checklist, saves five prompts and trials a research tool during the quiet months; annual cost about $240 until the research tool is judged. A two-partner firm with small-business clients uses a team plan so both partners work from the same saved instructions, adds one research seat shared by whoever has the hard question that week, and reviews usage in spring; this is the firm where the weekly answer cap, if there is one, needs checking against reality. A firm that already uses a client portal with workflow automation asks that vendor which letters and reminders its AI features can draft, keeps a general assistant for the rest, and spends the research budget only if its staff report that they cannot find answers fast enough.

In each case the first month's job is the same: run the test above, write down the settings and the consent decision, and agree who reviews what. The tool matters less than whether the review step actually happens in March.

Picking by firm size

A sole practitioner with 150 clients and mostly individual returns can start with one chat subscription, a redaction habit and a seven-day trial of a research tool. A three-partner firm doing business returns should check team plans, because sharing instructions across staff and keeping business content out of training by default is worth more than a few dollars a seat. A firm already on a practice-management platform should start by asking what its AI features cover. If the busy season is the real pain, how small tax practices use AI through busy season sets out where the hours go.

One more case for completeness: a consultant who mostly advises on planning rather than preparing returns will spend most of their AI time on memos and research, so a research tool earns its place sooner and a client-portal product later.

What none of these tools should do alone

No tool should send anything to a tax authority without your review, state a filing position you have not examined, or contact a client in your name unreviewed. Keep three habits. Save each prompt that works, so the next preparer reuses it. Log which tool drafted which letter, so a later question about a document has an answer. And review the tool settings each quarter, because vendors change privacy defaults. If you would rather map this with someone, that is what my AI implementation consultation covers.

Questions tax consultants ask before choosing a tool

Do I need client consent before pasting a return into an AI tool?

Section 7216 limits how a preparer may disclose or use return information without the taxpayer's signed consent. Whether a given tool counts as a disclosure depends on its terms and your arrangement with it. Ask your professional body or a tax lawyer, and anonymise anything you paste while you find out.

Is a $20 chat assistant enough for tax research?

It is enough for first-pass orientation, drafting and summarising documents you supply. It is weaker where you need to cite current authority, because it may mis-cite or invent a source. Treat every citation as unverified until you have opened the primary source yourself.

Can AI draft the whole reply to an IRS notice?

It can produce a good first draft of the letter and the supporting-document list. You must still check the notice's figures against the return, confirm the response deadline, and sign off the final text yourself, because the representation is yours.

What should I ask a tax-AI vendor before signing?

Ask whether client data is used for training, where it is stored, how long it is kept, whether there is a SOC 2 report, and what happens to your data when you cancel. Get the answers in writing and keep them with your engagement records.

Further reads

• Can AI Do My Business Taxes? What It Can Safely Prepare — What AI can safely prepare for a business client.
• Tax Software With AI: What a Small Business Should Check — What to check in tax software that adds AI.
• How to Check Sources and Citations in AI Research — A routine for verifying any AI citation.
• The AI Tool Stack a One-Person Consultancy Actually Needs — A lean stack if you practise alone.
• AI Receptionist for Accounting Firms: Tax-Season Calls and Cost — Handle the phone surge without adding staff.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check your firm is ready before rolling out any tool.
• 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.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Treasury Regulation 26 CFR 301.7216-3 (Cornell LII), the IRS Office of Professional Responsibility and Circular 230 page, Blue J pricing page, CoCounsel Tax Essentials plans page, TaxDome pricing page, claude.com/pricing, ChatGPT pricing pages and the Microsoft 365 Copilot Business page, all read 4 November 2026.