Claude for Tax Professionals: Projects, Citations and Data Controls

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Claude for Tax Professionals: Projects, Citations and Data Controls.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Claude for Tax Professionals: Projects, Citations and Data Controls.

A tax professional can use Claude for research notes, client letters and workpaper drafts by keeping one project per job, asking for a named source on every claim and opening each source before relying on it. Client data is controlled by plan: Team and Enterprise do not train on your content by default, while consumer plans leave that setting with you.

The order matters. Decide what data may go in, then pick the plan, then build the projects, then write the checking routine. Most firms do it backwards: they start with prompts, discover a client name sitting in a chat history and only then ask what the plan allows. Run it the right way round and each later choice is half made: the data list tells you which plan you need, and the plan tells you whether colleagues can share a project.

Follow me on Instagram@sagnikteaches

Decide what client data may enter Claude at all

Before you open a project, write a three-row list on one page: data that may never go in, data that goes in only in a reduced form, and data that may go in as it is. The list takes twenty minutes and removes most of the argument later, because every future "can I paste this?" gets answered by the page rather than by memory.

Connect on LinkedInSagnik Bhattacharya
TierExamples (illustrative)Rule
NeverGovernment identifiers, bank account numbers, login details, scanned ID documents, full dates of birthStays out of every AI tool, whatever the plan
Reduced formIncome figures, entity structure, prior-year balances, correspondence with an authorityRelabel the client ("Client A, sole trader, two rental properties"), round figures where the question allows, strip addresses
As it isYour own firm procedures, blank letter templates, published guidance you have downloaded, checklistsFree to use, because nothing in it identifies a client

Notice how much useful work sits in the bottom two rows. Research questions rarely need a client's name; they need the facts that drive the rule: residency status, entity type, the year, the amount band. If you can answer a question with "Client A", do.

Subscribe on YouTube@sagnikteaches

Your engagement letter and any professional-body guidance sit above all of this. If your terms say client data stays inside named systems, an AI tool is not one of them until you have updated the terms or the client has agreed. Ask your professional body or an adviser where you are unsure; what follows is a set of practical controls, not a ruling on your obligations.

The Complete Excel Course with AI Integration thumbnail

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Choose the plan, because the plan decides the data defaults

Figures below are list prices from claude.com/pricing, checked on 4 November 2026. Check the page again before you buy.

PlanList priceTraining on your contentAdmin and sharing
Free$0, up to five projectsA setting you control, on consumer termsNone
Pro$17 a month billed annually, $20 monthlySame consumer settingNone
Team Standard$20 a seat annually, $25 monthly, two to 150 seatsNone by defaultCentral billing and administration, single sign-on, shared projects
Enterprise$20 a seat annually plus usage at API rates, 20-seat minimumNone by defaultAdds audit logs, role-based access and a compliance API

For a sole preparer the two-seat Team minimum means about $40 to $50 a month for the seats. That is the price of getting a business plan's default, so weigh it against what a single client file is worth. For a practice of four or five, Team Standard is the realistic choice; audit logs and role-based access only matter once you have a compliance officer or a client who asks for them.

One honest limit: a business plan's "no training by default" is a statement about model training. It does not make a risky paste safe. Retention, who inside your firm can open a chat and what a departing member can still see are separate questions, and your reduced-form rules still apply. If you want to compare the same arrangements in ChatGPT, Claude Projects against ChatGPT Projects for client files sets them side by side.

Build one project per job and write the firm procedure into it

A project is a workspace with its own instructions and its own reference files, so Claude starts each chat already knowing how your firm works. Keep projects narrow. A project called "Tax" collects everything and answers nothing well; four projects with one job each stay sharp.

  1. Research notes. Instructions: how to structure an answer, that every claim needs a named source, and what to say when it cannot find one. Files: your firm's research memo template.
  2. Client letters. Instructions: your house tone, reading level, what a letter must always include (what we did, what we need, by when). Files: three or four of your own anonymised past letters as style examples.
  3. Workpaper drafting. Instructions: your file-note layout, abbreviations, sign-off lines. Files: a blank workpaper and one completed, anonymised example.
  4. Intake and checklists. Instructions: how you onboard a new client. Files: your engagement letter template and document request lists.

Free accounts can hold up to five projects, so the four above fit. Project files are capped at 30 MB each. Sharing a project with colleagues needs Team or Enterprise. Anthropic is also rolling out a new version of projects in stages, starting with Claude Code, so your screen may differ slightly from these steps; the instructions-plus-files idea stays the same. If you are setting this up for staff as well, the shared team assistant set-up covers permissions and naming.

A project instruction written for research notes

Paste something like the block below into the project's instructions box and edit it to your firm. Keep it short enough that you can read it in a minute; long instruction blocks tend to be half-followed.

You assist a small tax practice with research notes. Rules:
1. Never ask for or repeat client names or identifiers. Refer to "the client".
2. For every statement of law or practice, give the source name, the
   section or paragraph, and the tax year it applies to. If you cannot
   name a source, write "NO SOURCE FOUND" instead of guessing.
3. Separate three parts: Facts assumed, Answer, Points to verify.
4. State which tax year and which type of taxpayer the answer covers.
5. If the question depends on facts I have not given, list them as
   questions before answering.
6. Keep the answer under 400 words unless I ask for more.

Rule 2 is the one doing the work. It turns an unchecked paragraph into a list of things you can go and look up, and the "NO SOURCE FOUND" phrase gives you a signal you can search for later.

A research question with numbers, start to finish

Say a four-person practice has a client who is a sole trader with turnover of $84,000 for the year, started using a spare room as an office in March and wants to know how to treat a share of household costs. The preparer anonymises the question and types it into the research project.

Client A, sole trader, turnover about $84,000 this tax year. Uses one
of five rooms at home as an office, part of the year (ten months).
Household costs for the year: utilities $3,100, insurance $900,
mortgage interest $6,400. Which methods exist for claiming a share of
home costs, and what record should the client keep? Tax year: current.

An illustrative answer shape: Claude lists a floor-area method (one room of five is 20 per cent) and, where the rules offer one, a simplified flat-rate alternative; it applies 20 per cent to $10,400 of costs (utilities, insurance and interest together) to get $2,080, adjusts for ten months, giving about $1,733; and it names the guidance it believes applies. Then the preparer does the part Claude cannot: opens each named source, checks that the method exists for this client's entity type and year, and checks whether mortgage interest belongs in the calculation at all, since some rules treat it separately from running costs.

Two realistic failures show up in this exact exercise. First, the answer may blend two regimes: a rule that applies to employees with one that applies to the self-employed. Second, it may quote a threshold from a previous year. Both are caught by the same habit, opening the source and reading the year on it. The arithmetic is a good check on Claude too: $10,400 x 20 per cent = $2,080, and ten twelfths of that is $1,733. If the numbers do not reproduce, the method is wrong.

Ask for citations, then verify them in a fixed order

A citation from Claude is a lead, not proof. It can name a document that exists and misdescribe what is in it, give a correct title with a wrong paragraph number, or cite something that does not exist. Treat each one like a footnote from a stranger.

The verification order below takes about two minutes per citation once you are used to it.

  1. Does the source exist? Search the issuing body's own site for the title. If the search finds nothing, mark it "unverified" and stop using the point.
  2. Is it the right version? Find the date or tax year on the document. A rule from last year is a different rule.
  3. Does the passage say what Claude says it says? Open the paragraph cited and read it. Copy the exact words into your note.
  4. Does it apply to this client? Check entity type, residency, thresholds and any exclusions listed near the passage.
  5. Record it. Save the link or reference, the date you checked and your initials in the workpaper.

A follow-up prompt makes step 3 quicker. After the first answer, ask:

For each source you named, quote the exact sentence you relied on,
word for word, and say how confident you are that the quote is
accurate. If you cannot quote it, say so.

A quote you can paste into a search box is something you can falsify in ten seconds. If the search comes back empty, you have learnt the quote was reconstructed rather than read. Where you have attached the source document itself to the project, ask Claude to quote the passage it relied on from that file; a quote from a document you supplied is quicker to check than a reference recalled from memory, but you still read the passage. The routine in AI contract review for small firms shows the same read-the-clause discipline on a different document type.

Four more jobs to run inside the projects

A client letter from your own style examples

In the letters project, paste the anonymised facts and ask for a draft in your tone. Illustrative input: "Client A, sole trader, needs to send us three items by 15 January: bank statements for the second half of the year, the home-office room measurements, and the van mileage log. Explain why each matters in one sentence, and say what happens if we do not get them." The first draft usually over-explains. Your fix is the same each time: cut the opening paragraph, keep the list, add the deadline in bold. Put that fix into the project instructions so it stops recurring.

Before and after: a file note

Before: "Spoke to client re home office, said 20% ok, check mileage." After, drafted from the same line plus the project's file-note layout: date; who attended; topic; what was agreed (20 per cent floor-area basis, ten months); what is outstanding (mileage log, due 15 January); next step and owner. The second version is what a colleague can pick up in the preparer's absence. Ask Claude to keep a "Points to verify" line at the bottom, then clear it yourself before the note is filed.

A document request list that matches the engagement

Give the intake project a client's reduced-form profile (employed with a side business, two rental properties, one dependent) and ask for a request list grouped by income source with a due date column. The list is a draft of what you would send anyway; compare it with your standing checklist and note the lines Claude missed. Update the checklist file in the project when it misses the same line twice.

A mistake and how it showed up

A preparer asked for the rate that applied to a particular type of income and received a confident figure with no source. It looked plausible, matched what they half remembered and went into a draft email. The reviewer caught it because the email lacked the "Points to verify" block the project instructions demand. The fix was not a better prompt; it was making the block compulsory and making the reviewer's first question "where is the source?". A rule that a draft without sources cannot move to review costs nothing and catches this class of error.

Ask better research questions

Weak: "Is a payment to my client's daughter deductible?" Stronger: "Client A, a sole trader, pays a family member a monthly amount of $600 for bookkeeping. What conditions usually need to be met for the payment to be treated as a business expense, and what evidence would a reviewer expect? Tax year: current. Name your sources." The stronger version states the entity type, the amount, the year and the output you want. It also asks for conditions and evidence, which is where tax answers usually live.

A table of question shapes that work well:

JobGood question shapeWhat to verify
Concept explanation"Explain X for a client who has no tax background, in 150 words."That the simplification has not changed the rule
Rule lookup"What conditions apply to X for entity type Y in year Z? Name the source."Source, year, exclusions
Comparison"Compare two methods for X with a worked example using these figures."The arithmetic and the eligibility of each method
Issue spotting"Given these anonymised facts, what questions should I ask the client?"Nothing; use as a prompt, not an answer
Drafting"Draft a letter in our tone using this template."Facts, dates, amounts, tone

The issue-spotting row is the safest and often the most valuable: Claude generates questions rather than conclusions, so the risk of a wrong rule going into advice is lower. The same discipline of anonymising and of reading sources shows in how small law firms draft letters and routine documents with AI.

Controls to switch on and habits to keep

  • Training setting. On Free and Pro, find the model-improvement option in privacy settings and switch it off. Do this before the first client-related chat, and check it again after any account change.
  • Chat clean-up. Delete chats you no longer need, and name chats by job rather than client. Do not rely on deletion as your only control.
  • Connectors. Do not connect your email, drive or practice-management system until you have decided what Claude may read there. A connector can reach every file the connected account can reach.
  • Browser agent. Anthropic's safety guidance for Claude in Chrome strongly advises against using it to manage or act on sensitive information, recommends against using it on sensitive sites or pages with regulated data, and points out that it screenshots whatever is visible in the tabs it works in. A tax portal or a client's online account is exactly that kind of page. Keep the extension away from it.
  • Sign-in. Claude's login page offers Google, Apple, email and single sign-on, so whoever controls the account you sign in with controls Claude. Protect that Google, Apple or email account with two-step verification, and on Team use single sign-on if your firm already runs it.
  • Leavers. On Team, remove a member's seat as soon as they leave and move shared projects to someone who stays.
  • A review log. A sheet with date, project, question, source checked and checker's initials. Ten minutes a week of looking at it tells you whether the routine is being followed.

If a client asks whether you use AI, answer plainly: which plan, what goes in, what does not, who checks. A one-paragraph policy you can paste into an engagement letter is more reassuring than a long one. Write it once, keep it in the intake project and update it when the plan or the rules change.

How to tell the set-up is working

After four weeks, pull five research notes at random and score them. Does each claim have a source? Was each source opened? Is the tax year stated? Was anything anonymised that should not have needed to be? Five out of five on the first two questions means the routine holds. Less means the instructions need tightening, or that people are skipping the check because it takes too long. If it is the latter, shorten the template rather than drop the check.

Also time yourself. A research note that used to take 40 minutes might take 25 with the drafting done and 10 spent verifying, a saving you can measure and not just feel. If verification is eating the whole saving, the questions are too broad: narrow them. If you would like help choosing which of your tax-season jobs deserves a project first, my AI implementation consultation covers exactly that, but the steps above work without it.

Questions tax preparers ask before putting client work into Claude

Can I use the free Claude plan for client tax work?

You can, but it is the weakest fit. Free and Pro accounts sit on consumer terms where the training setting is yours to manage, and there is no admin console. Team and Enterprise do not train on your content by default and add central administration. Keep identifiable client data out of consumer plans.

Does Claude give me the same answer twice?

Not necessarily. Wording and sometimes the conclusion can change between runs. That is why every research answer needs a source you can open, and why a result you rely on should be saved to the client file with the source attached.

Should I paste a full tax return into a project?

Only if the plan and your engagement terms allow it. Most research and drafting jobs work with an anonymised summary: entity type, figures rounded or relabelled, no names, numbers or identifiers. Start there and widen only with a reason.

Who is responsible if Claude gets a rule wrong?

You are. The tool is an assistant that drafts and organises; the preparer signs. Treat its output the way you would treat a trainee's first draft with no reference list: useful, unfinished and yours to verify.

Further reads

Sources: claude.com/pricing (plans, seat minimums, training defaults, checked 4 November 2026); Anthropic's help centre articles on projects, file uploads and using Claude in Chrome safely.

Want your tax-season AI rules written down?

On a 1:1 call we can map your research, drafting and client-data steps, decide which plan fits your firm and set up projects your team can run.

Book a 1:1 call with me