Module 2, Tax Research: A Reliable Plan You Can Use With Real Clients

⚡ The 30-second version. Use it for: researching tax questions reliably, without exposing the client. The method: strip the client's identity first (one click), then make the AI verify every citation against the real source before you rely on it, using free public sources, no subscription required. Start here: the 15-minute starter at the bottom.

Tax research is the most-named accounting-specific use of AI, and the most dangerous, for two reasons this module is built to defeat:

1. Reliability: general AI models confidently fabricate Code sections, regulations, case names, revenue rulings, and dollar thresholds that look perfect and do not exist.

2. Safety with clients: tax research often involves client facts, which can be tax return information (TRI) under §7216, so how you research has data-protection consequences.

The good news: the fix for both is the same discipline. When you abstract the legal question away from your client's identity, you simultaneously get more reliable research and take the client's data out of the §7216 picture entirely.

The one-sentence method: AI finds leads and structures the analysis; primary sources supply the authority; you verify every citation and sign. The AI is never the source of law.

This is a do-it-yourself plan: it runs on free, public sources any member can access, with a documented workflow and prompts you can copy. Paid citation-grounded tools are an optional upgrade, not a requirement.

▶ Want to see it first? Walk through a full worked example, a real (anonymized) question, scrubbed and sent to the AI, where the verification step catches a citation the AI invented. That catch is the entire point.

Why "just ask ChatGPT/Claude" fails, and what to do instead

A general LLM predicts plausible text. "§1.469-2T(c)(7)(iv)" looks like a real regulation whether or not it is. The model has no built-in link to the U.S. Code or the CFR, and it cannot tell you whether a case was later overruled. So the failure modes are:

• Fabricated authority, a cite that doesn't exist.
• Misquoted authority, a real cite that doesn't say what the model claims.
• Stale authority, a real, once-correct cite that's been superseded, repealed, or overruled.
• Wrong-level authority, citing an IRS Publication or FAQ as if it were binding law.

The cure is a retrieve → verify → reason loop: use AI to map the issue and generate leads, then pull and read the primary source yourself (or via a tool that links to it), confirm it exists and says what's claimed, and only then let AI help you synthesize and draft.

The authority stack (free public sources)

Bookmark these. This is the backbone that makes the plan reliable without a subscription.

⚠️ IRS Publications, FAQs, and instructions are NOT substantial authority. They're convenient plain-English summaries, but courts have held taxpayers generally cannot rely on them as binding law. Use them to orient; cite the Code, regs, and rulings they're based on.

Authority hierarchy (highest → lowest):

1. Internal Revenue Code and U.S. Supreme Court
2. Treasury Regulations (final & temporary; proposed regs carry less weight)
3. IRB guidance, Revenue Rulings & Procedures (official IRS position), Notices
4. Case law, Circuit Courts of Appeals, Tax Court, Court of Federal Claims, District Courts
5. Sub-regulatory/secondary, IRS Pubs, FAQs, Private Letter Rulings (no precedential value to non-recipients), treatises, articles, and AI output, which is non-authoritative, period.

Make it applicable without exposing the client, fact abstraction

Here's the real tension in tax research: the answer is only as good as how specific it is to your client, but your client's specifics are exactly what's confidential. Tell people to "anonymize" and they either strip so much the research no longer applies, or leave identifiers in because abstracting feels like work. The resolution is one principle:

Tax outcomes turn on structure, relationships, magnitudes, timing, and elections, NOT on identity. Identity is not a tax input.

A Code section never asks "is this Jane Doe?" It asks "is this an S corporation? a related party? a long-term holding? above the threshold?" So you can keep 100% of what makes research applicable while removing 100% of what's confidential. Abstraction costs nothing in accuracy, done right it sharpens the question.

Keep (abstracted) vs. strip

Data-minimize by issue, let the research tell you what it needs

Don't dump the client file. State the issue first; the research surfaces exactly which facts are determinative; you provide only those, abstracted. This flips the habit that causes leaks (pasting a K-1 with names and EINs) into its opposite. The tax-research skill does this for you, it lists the specific facts the issue turns on and asks you to fill them in scrubbed.

The two-layer method (where applicability and safety both win)

1. Layer 1, Legal research (abstracted facts, any enabled tool): develop the rule and how it applies to this kind of situation. No identifiers, often no exact numbers, just enough structure to get a correct framework.

2. Layer 2, Client application (real numbers, controlled environment): plug the actual figures into the verified framework in a firm-approved tool, a spreadsheet you control, or by hand, never in public AI.

The legal logic is developed once, abstracted; the real numbers only ever touch a place you control. Fully client-specific answer; client data never leaves a safe lane.

Magnitudes: thresholds vs. exact figures

Ask: does the exact number change the analysis, or just whether it crosses a threshold? Usually the question is "are we over the §199A phase-in?" not "is it $384,219?" Use threshold-relative or rounded/scaled figures for the legal research; reserve exact figures for Layer 2. When a computation truly needs the real number (actual QBI, basis, AMT), run it in Layer 2.

Pseudonymize with a local key (keeps relationships intact)

For multi-party fact patterns, replace each party with a consistent token, Taxpayer A, Entity X, Trust T, and keep a private mapping (token → real identity) on your side only. The relationships and structure (what the law cares about) are fully preserved under the tokens, so the research is genuinely applicable; you re-attach real identities locally when you write it up. The tokens go to the AI; the key never does.

Bottom line: you are not choosing between useful and safe. Abstracted-but-complete facts are both, and the Tax Research Fact Sheet makes producing them faster than copy-pasting the client's documents.

Automate the abstraction, the local scrubber

The honest problem with "abstract your facts": busy preparers won't do it by hand, so they paste the raw file instead. The redactor tool removes that friction, drop in a client note and it returns an abstracted fact pattern (safe to share) plus a local key (token→identity, stays on your machine). Crucially, it runs entirely locally with no network calls, because a tool that strips PII must touch the raw PII, and that can't happen in the cloud without defeating the purpose. The safe, abstracted text is then what you bring to the AI.

The round trip: redact out, re-attach locally

Abstraction isn't a one-way street, it's a loop, and the loop closes on your machine:

1. Redact out. Real client facts → an abstracted fact pattern + a local key (token → real identity). Use the redactor tool or do it by hand.

2. Research with the AI using only the abstracted version (the skill, a Project, or any assistant). The answer comes back referring to [Person_1], [Entity_1], [ADDRESS_1], etc.

3. Re-attach locally. Swap the tokens back to the real names using your local key, on your machine, never in the AI. The Redactor tool's "Bring the answer back" step does this for you (paste the AI's answer + your key → real identities restored); or find-and-replace from the key by hand.

The payoff: a fully client-specific work product, while the client's identity never left your desk. The local key is the only thing bridging the two, keep it with the client file and never share it. (Once re-attached, the document contains real client info again, so handle it as confidential, don't paste it back into a general AI tool.)

The reliable research workflow, the plan

Six steps. Steps 1–2 use AI freely (no client data needed). Step 3 is the non-negotiable verification gate. Steps 4–6 produce a defensible, documented answer.

Step 1, Frame the issue without client identifiers

Restate the question as an abstract fact pattern. Strip names, SSNs/EINs, and identifying detail; keep only the tax-relevant facts. "A single-member LLC taxed as an S-corp wants to...", not "My client Jane Doe, EIN..." This both sharpens the legal question and keeps TRI out of the tool (see "Keeping it safe with clients" below).

Step 2, Use AI to build an issue map and a search list, not answers

Ask the model to identify the sub-issues, the likely controlling Code sections and regs, the key terms of art, and what authorities you should go read. Treat everything it returns as a lead to verify, never as the answer.

Step 3, Retrieve & VERIFY every authority against a primary source ⛔ (the reliability gate)

For each authority the AI named, open the free primary source and confirm:

• It exists (the section/case/ruling is real and the number is right).
• It says what's claimed (read the operative text, don't trust the paraphrase).

• It's current/good law (check effective dates; look for later regs, rulings, or decisions that superseded or overruled it).

• It's the right level of authority for the weight you're putting on it.

If you can't verify a cite, it's wrong until proven otherwise, delete it.

Step 4, Synthesize from the verified sources you provide

Now paste the verified authority text back to the AI and have it help organize the analysis, draft the reasoning, and surface counter-arguments. Because it's reasoning over text you supplied, it can't invent cites. Ask it explicitly to flag anything it asserts that isn't in your provided sources.

Step 5, Run a devil's-advocate pass

Ask the model to argue the opposite conclusion and identify the weakest link, the contrary authority, and the facts that would change the answer. This catches the agreeable-but-wrong trap.

Step 6, Document in a research memo (workpaper) and sign as reviewer

Capture issue, facts, verified authorities, analysis, and conclusion in a workpaper, with preparer and reviewer sign-off. Use the Tax Research Memo template. Under SSTS §1.4 you remain fully responsible for the conclusion whether or not you used AI.

Copy-paste prompt library

1. Issue framing (anonymized)

Act as a tax research assistant. Here is an ANONYMIZED fact pattern: [facts, no identifiers].
The question is: [question]. Do NOT give me a final answer yet. Instead: (1) restate the
issue and any sub-issues, (2) list the Internal Revenue Code sections and Treasury Regulations
most likely to control, (3) list the key terms of art and the authorities I should read to
resolve this. Label everything as LEADS TO VERIFY, I will confirm each against primary source.

2. Authority map → my reading list

For the issue above, give me a table of candidate authorities: Code sections, regs, revenue
rulings/procedures, and leading cases. For each, one line on why it might be relevant and what
I should check. Flag which are primary/binding vs secondary. Do not summarize their holdings as
settled, I'll read them.

3. Synthesis from verified sources (the safe-reasoning step)

Here is the VERIFIED text of the authorities I pulled: [paste statute/reg/ruling/case excerpts].
Using ONLY these sources, help me analyze [issue]. Build the reasoning step by step, cite to the
specific provided provision for each point, and explicitly flag any statement you make that is
NOT supported by the text I gave you.

4. Devil's advocate

Argue the OPPOSITE conclusion to [my tentative conclusion], using only the provided authorities.
What's the strongest contrary position, the weakest link in my analysis, and which facts would
change the result?

5. Plain-English client explanation (after the memo is done)

Summarize the verified conclusion for a non-expert client in plain English: what the answer is,
why, and what it means for them. No citations, no jargon. [Paste your verified conclusion only.]

The citation verification gate (print this)

For every authority before it goes in a memo or client answer:

• [ ] Exists, I found it on a primary source (law.cornell.edu / ecfr.gov / irs.gov/irb / ustaxcourt.gov / CourtListener), and the citation number matches.

• [ ] Says it, I read the operative text; it supports the point (not just the AI's paraphrase).

• [ ] Still good, I checked effective dates and looked for later guidance/decisions that superseded or overruled it. (Free citators are limited, this is where paid tools help.)

• [ ] Right level, primary/binding authority for anything load-bearing; Pubs/FAQs/AI are not authority.

• [ ] Documented, the verified cite and where I confirmed it are in the workpaper.

Keeping it safe with clients (the data side)

Tax research is the use case where good method and compliance line up perfectly:

• Default to anonymized fact patterns. If no TRI is entered, §7216 isn't implicated at all, you can use almost any tool for the legal question. This is the single most powerful habit in this module. (See Regulatory Foundation §1: "no TRI entered" is a safe lane.)

• When you must use client-identifiable facts (e.g., running their actual numbers), use a firm-approved tool under the policy, not public/personal AI, and apply the §7216 / AICPA Confidentiality Rule analysis (consent or a confidentiality agreement/DPA). See Guardrails and the AI Acceptable-Use Policy.

• Never paste SSNs/EINs into a research prompt. They are never relevant to the legal question.

Tool picks

• Free primary sources (the authority stack above), non-negotiable backbone.

• General AI (Claude as the Lab default), excellent for Steps 1–2, 4–5 (issue mapping, reasoning over provided text, devil's advocate). Not a source of law.

• AI with live web/retrieval (Claude with web access, or research tools that link to sources), helps Step 3 by pulling real source text, but still verify the link resolves to the real authority, retrieval tools can misattribute.

• Optional paid, citation-grounded tax tools (CCH AXcess Intelligence, Thomson Reuters Checkpoint/CoCounsel, Blue J, TaxGPT, Bloomberg Tax), these are grounded in licensed primary- source databases and include citators (is it still good law?), which free sources do poorly. Worth it for high-volume or high-stakes research; the verification discipline still applies.

Guardrails recap

• AI is never the source of law. Every authority is unverified until checked against primary source. (GUARDRAILS rule 3.)

• You are the reviewer of record. SSTS §1.4, you own the conclusion. (GUARDRAILS rule 4.)

• Anonymize by default; approved tool + consent analysis when TRI is unavoidable. (GUARDRAILS rules 1–2; Foundation §1.)

• Document the verified authorities in a workpaper, research you can't defend is research you didn't do.

Your 15-minute starter

1. Bookmark the four anchor sources: Cornell LII (Code), eCFR Title 26 (regs), irs.gov/irb (rulings), ustaxcourt.gov (cases).

2. Take a real but anonymized question you've researched before (so you know the answer).

3. Run Prompt 1 to map the issue, then Prompt 2 for the authority list.
4. Run the verification gate on each cite the AI gave you. Notice how many were close-but-off, that's the lesson that makes you trust the process, not the model.
5. Paste the verified text back with Prompt 3 and watch the difference: grounded reasoning with real cites instead of confident guesses.

Win condition: you never again copy an AI tax answer without having opened the primary source, and you have a repeatable, documentable workflow you'd be comfortable showing a reviewer or the IRS.

Next module: Writing, Memos & Internal Deliverables.