A client sends a long email, three attachments, and a question that sounds simple. Before you can answer, you need to separate the facts from the assumptions, find the missing document, research the rule, and explain the result.

Much of the effort sits between those steps. You translate the client’s story into a research question, find the governing provision, check its exceptions, and carry the result into a memo or email. Asking AI for an answer can produce text quickly while leaving you to rebuild the research behind it.

Our recommended setup is Claude for organizing and drafting, with TaxMCP for retrieving tax authority. You work through the question in a conversation, bring in source material you can open, and draft from the evidence you have reviewed. The goal is a shorter path to a client answer you can support.

This guide walks through that setup for CPAs, enrolled agents, and tax attorneys. You can start with TaxMCP’s free IRC research and test it on a question you already know.

How Claude, ChatGPT, and TaxMCP work together

Claude can help turn a client message into an issue list, organize supplied material, and draft the explanation. ChatGPT supports similar research, file-analysis, and drafting tasks, with availability depending on your plan and workspace settings. That gives the practitioner one place to develop the work instead of repeatedly starting from a blank document.

The tax-specific question is where the legal support comes from. A request to “include citations” still leaves you to establish whether those citations exist, support the claim, and apply to the client’s year and facts.

TaxMCP brings tax authority into that working conversation. It gives Claude and ChatGPT dedicated tools to search and retrieve the Internal Revenue Code, Treasury Regulations, IRS guidance, supported state tax codes, and U.S. Tax Court opinions, with source links. Coverage depends on your TaxMCP plan.

The practical benefit is continuity. Frame the issue, retrieve the relevant law, open it for review, and ask the assistant to draft from the sources you approved. The research can travel into the client answer and file note with its supporting links attached.

The walkthrough below uses Claude. If your firm uses ChatGPT, the same research-and-drafting sequence works with TaxMCP’s ChatGPT connection. In either assistant, the practitioner reviews the evidence and owns the conclusion.

Give each task a reviewable output

Start with one recurring task rather than asking Claude to handle an entire engagement.

TaskAsk Claude to produceCheck before moving on
Client intakeConfirmed facts, missing facts, follow-up questionsWhether it added assumptions or missed a material date
Document reviewA source-labeled extraction and discrepancy listFigures and references against the original files
Tax research with TaxMCPRelevant authority with exact identifiers and linksSource text, applicability, exceptions, and effective dates
CalculationsA workpaper with inputs and visible formulasInput accuracy, formula logic, and independent totals
Client communicationA draft answer and companion file noteWhether the conclusion matches the approved research

Those outputs give the reviewer something more useful than a polished paragraph. They expose how the answer was built.

Set the data boundary first

Before uploading anything, confirm which Claude account and settings the firm permits for the task. Review retention, sharing, connected services, and the information the workflow actually needs. Start a trial with synthetic examples or de-identified facts.

Anthropic distinguishes its consumer model-improvement settings from its commercial data-use terms. It says commercial chats are not used for model training by default, subject to opt-ins and feedback exceptions. A training setting alone does not establish whether a client-data workflow is appropriate.

Public-law research rarely needs a name, Social Security number, account number, or complete return. TaxMCP does not require client documents or identifiers. Review its security disclosures separately from the Claude account’s terms, and send only the facts needed for the research question.

Use a Project for recurring instructions

Claude’s Projects let you keep reference files and instructions available across chats in a project. That makes them useful for a firm memo format, an intake checklist, or examples of the writing style you want.

A starting instruction might be:

Separate confirmed facts from assumptions. State the tax year, jurisdiction, and taxpayer type. Identify missing facts before reaching a conclusion. Use retrieved or supplied authority for legal claims and preserve the source links. Keep unresolved questions visible. Draft client communications in plain English.

Treat those instructions as a starting point for review. They do not guarantee that Claude will follow every constraint. Check project sharing and memory settings before using engagement material, and keep the approved facts and final workpapers in the firm’s system of record.

1. Turn the client message into an intake checklist

Give Claude a de-identified message and ask it to organize the question before answering it:

Extract the confirmed facts, relevant dates, taxpayer or entity type, tax year, and jurisdictions from this message. List missing facts and contradictions separately. Draft the shortest follow-up request that would let us research the issue. Do not answer the tax question or supply citations yet.

Read the checklist against the original message. If the client says an asset was purchased, Claude should not quietly assume when it was placed in service. If the message omits the entity type, that field should remain open.

The useful output is often a better question to the client. Getting that question right can save more time than drafting an answer early.

2. Review documents with a trail back to the page

For an approved document workflow, ask for extraction before interpretation. A useful schedule identifies the file, PDF page or worksheet, reporting period, amount, and anything that was unclear.

Build an extraction table from these documents. For each item, show the filename, PDF page or worksheet and cell, document type, tax year or reporting period, and relevant amounts. Flag unreadable fields, possible duplicates, and conflicting figures. Do not infer missing values or decide tax treatment.

Claude supports PDF, DOCX, CSV, and other uploads; XLSX uploads require code execution and file creation. Document format matters. Anthropic says non-PDF documents are processed as text, so embedded images may be missed. Its current PDF guidance supports visual analysis for files of 100 pages or fewer; longer supported PDFs receive text-only processing.

Check the extraction against the originals, including the page references. A correct-looking total can conceal a duplicated statement, a missing page, or a number read from the wrong column. An unexplained difference belongs on the follow-up list until someone resolves it.

3. Connect research to authority you can open

Once the facts are stable, move to research. Connect TaxMCP in Claude so you can ask the client question in plain English and retrieve the underlying authority without moving to a separate research interface. You can also supply authority you have already retrieved and reviewed.

Claude supports remote MCP connectors for external tools and data. Its navigation is currently changing, so your account may show Customize or Plugins in the sidebar; Anthropic documents the Plugins rollout here. Workspace policy may control what you can add.

Use a research request that asks for evidence before analysis:

Use TaxMCP to retrieve the authorities relevant to this fact pattern for [YEAR] and [JURISDICTION]. Return exact section identifiers, relevant text, source URLs, effective-date information, and material exceptions or cross-references. Distinguish statutes, regulations, IRS explanations, and cases. Do not draft the client answer yet. If retrieval is unavailable, say so.

Check the tool activity. A response saying it consulted tax law is not evidence that retrieval ran. Then open the sources yourself.

Consider an employer asking whether a thank-you payment to an employee can be treated as a gift. IRC §102 illustrates why the full provision matters:

Finding the general exclusion is only the first step. The researcher must read the employee rule and determine whether any separate provision needs analysis. The client’s word “gift” does not settle the treatment.

This is the job TaxMCP adds to the workflow: retrieve the provision and its source link so Claude can work from the text while you test the analysis. After checking the employee rule, ask Claude to explain its effect using only the approved sources. You still decide whether the source set is complete and how it applies. Our tax research workflow goes further into that verification process.

4. Build a calculation workpaper with visible assumptions

Claude’s code execution and file creation can analyze uploaded data and produce Excel files with working formulas. Availability within a firm also depends on its workspace settings.

Use that capability for a reconciliation, a comparison of approved scenarios, or a supporting schedule. Supply the tax rules and assumptions you have verified instead of asking Claude to choose them while building the workbook.

Create a workpaper using only these approved inputs and assumptions. Keep inputs separate from calculations. Show formulas, units, rounding, and a reconciliation to the source totals. Flag missing inputs. Do not invent tax rates, limits, elections, or carryforwards.

Open the result in the spreadsheet application the firm uses. Recalculate it, inspect the formulas, and independently check the material totals. Correct arithmetic does not rescue an incorrect legal assumption, and a workbook that balances can still omit an input.

5. Draft the client answer from the approved record

After reviewing the facts, sources, and calculations, ask Claude for two outputs from the same record:

Using only the approved facts, authority, and calculations below, draft a concise client email and a separate internal file note. Lead the email with the answer and any condition that could change it. In the file note, preserve the issue, source links, assumptions, unresolved facts, and reasoning. Introduce no new facts or citations.

Compare both drafts with the approved conclusion. Remove any claim the research does not support. Keep the email as a draft until the responsible practitioner has reviewed it.

For a longer analysis, use the same approach in an AI-assisted tax research memo. Claude can help with structure and wording while the practitioner owns the conclusion.

Verify the work before it leaves the file

A second request to “double-check” can help identify issues, but it is still another model response. Verification means returning to the evidence:

  1. Facts — Compare the draft with the client record. Remove invented dates, relationships, and assumptions.
  2. Documents — Tie extracted figures and references to the originals. Resolve duplicates and missing material.
  3. Authority — Open each material citation. Check the text, tax year, jurisdiction, effective dates, and exceptions. Confirm subsequent history when relying on a case.
  4. Calculations — Reperform material arithmetic and inspect the workbook’s formulas and inputs.
  5. Deliverable — Confirm that the client answer and file note reflect the same approved analysis, and preserve the final record with reviewer sign-off.

Start with a familiar question and compare the result with a completed file. Measure time to a reviewable answer and the corrections needed to get there. If Claude saves drafting time but leaves the reviewer rebuilding the research, adjust the task or the requested output.

Try TaxMCP on a question you already know

Start with a completed research question from your own practice. Create a free TaxMCP account, connect it in Claude or ChatGPT, and ask for the relevant Code section, operative text, and source link. Compare the result with your existing file before asking for a draft.

The Free plan includes 20 Internal Revenue Code searches per day. Broader federal, state, and Tax Court coverage is available on paid TaxMCP plans; your Claude or ChatGPT access is separate.

That is a concrete way to evaluate the combination: can your assistant help you get from the client question to the law, then turn the reviewed research into a clear answer with a traceable file note?


Product documentation and the cited statutory text were checked on November 6, 2026. The prompts are suggested workflows, not a report of measured Claude accuracy. This article is general information, not tax or legal advice; verify current authority and the engagement’s facts before relying on any conclusion.