Turn a tax question into a cited research memo you can trust

Traditional vs AI-native

The traditional way

1. Take the facts in a phone call or email thread, often incomplete and scattered.
2. Frame the issue mentally and start keyword searches in the firm's research database.
3. Read explanations, then chase their citations to the Code, regulations, rulings, and cases.
4. Copy relevant passages into a document and draft the facts, issues, law, analysis, and conclusion.
5. Decide how strong the position is (substantial authority, reasonable basis, and so on), often without saying so explicitly.
6. Partner review; citations are rarely re-checked line by line.

Where it hurts

• Research quality depends heavily on who happened to do it and how much time they had.
• Facts are gathered too late, so the analysis rests on assumptions nobody wrote down.
• Confidence levels are implied rather than stated against the §6662/§6694 standards.
• Memo drafting eats the time that should go to thinking about the answer.

The AI-native way

1. 1AI

   Structured facts intake

   A questionnaire tailored to the issue area (entity type, dates, amounts, ownership, prior filings, documents) is sent to the client or engagement lead. The model reviews the answers, lists missing or contradictory facts, and drafts follow-up questions. The final fact set is frozen as a numbered list, so the memo's assumptions are written down and can be tested against Circular 230 §10.37(a)(2), which requires reasonable factual and legal assumptions.

   Tools: Form builder or portal questionnaire · Enterprise LLM

2. 2AI

   Draft issue statements and a research plan

   The model writes each issue as a precise question tied to specific facts ('Whether the $84,000 roof replacement on Building B is a deductible repair under Treas. Reg. §1.162-4 or a capitalizable improvement under Treas. Reg. §1.263(a)-3'), and produces a plan: Code sections, regulation parts, and the likely ruling and case law areas to check. A reviewer approves or edits the issues before research begins.

   Tools: Enterprise LLM

3. 3Automation

   Retrieve primary authorities

   A retrieval step pulls full text from your licensed research database (through whatever export or integration your license allows) and from official sources: the Code, current Treasury regulations from eCFR, Internal Revenue Bulletins on IRS.gov for rulings, procedures, and notices, and court opinions from the Tax Court and federal courts. Each document is stored with its citation, source URL, and retrieval date, and is split into chunks for search. The model can only cite what's in this retrieved set.

   Tools: Your licensed research database · eCFR · IRS.gov IRB · U.S. Tax Court opinions search · Vector index

4. 4AI

   Read the authorities and extract holdings

   For each retrieved authority, the model extracts the rule, the key facts, the holding or operative language (as exact quotations with pinpoint location), and how it bears on each issue, including adverse authority. Authorities are tagged by type and weight, following the Treas. Reg. §1.6662-4(d)(3)(iii) list of what counts as authority, so the analysis doesn't lean on secondary sources.

   Tools: Enterprise LLM with long context · Authority extraction schema

5. 5AI

   Draft the memo with a stated confidence level

   The memo follows a Facts / Issues / Law / Analysis / Conclusion structure. The analysis applies each authority to the frozen facts, addresses contrary authority, and ends with an explicit confidence level tied to recognized standards: 'more likely than not', 'substantial authority' (relevant to §6662 understatement penalties and §6694 preparer penalties for undisclosed positions), or 'reasonable basis' (generally enough only with adequate disclosure, such as Form 8275). It states which facts, if different, would change the conclusion.

   Tools: Enterprise LLM · Firm memo template

6. 6Automation

   Independent citation and quotation verification

   A separate step, deterministic code backed by a second retrieval pass, parses every citation in the draft and confirms it matches a document in the retrieved set or can be fetched from an official or licensed source. Every quoted passage must appear verbatim (after whitespace normalization) in that source. Unverified citations are removed or highlighted in red with a reason, and the memo can't be marked final while any remain. This step exists because AI-generated citations can look real but be fabricated. In Mata v. Avianca (S.D.N.Y. 2023), the court sanctioned two lawyers and their firm for a filing that cited nonexistent cases generated by ChatGPT.

   Tools: Citation parser (regex for IRC, Treas. Reg., Rev. Rul., Rev. Proc., T.C., F.3d, etc.) · Exact-match text search · Verification report

7. 7You

   Reviewer owns the conclusion

   A credentialed reviewer reads the verification report, opens every authority the conclusion depends on, checks that the characterization is fair (not just that the quote exists), confirms the facts and assumptions, and adjusts the confidence level if warranted. They sign the memo as the firm's advice. Their edits are saved as examples to improve future prompts.

8. 8Automation

   File, tag, and reuse

   The final memo, its frozen fact set, the verification report, and the retrieved authority set are stored together in the client file and in a firm knowledge base tagged by issue and Code section. Future research on the same issue starts from prior memos, with a check for later authority since the prior memo's date.

   Tools: DMS / SharePoint · Knowledge base index

What you need

• A structured facts questionnaire for the issue area
• Access to primary authority (a licensed database where the license allows, plus eCFR and IRS.gov)
• A citation checker that runs separately from the drafting step
• A credentialed reviewer with subject-matter experience

The stack

Named tools are examples, not endorsements. Check any tool against your own data rules first.

Copy the prompt

You are a senior federal tax researcher drafting an internal research memo for review by a CPA or tax attorney. You will receive:
1. FACTS: a numbered list of frozen facts. Treat them as the only facts.
2. ISSUES: the approved issue statements.
3. AUTHORITIES: JSON list of retrieved primary authorities, each with an id, a full citation, the authority type (statute, regulation, revenue ruling, revenue procedure, notice, court opinion), a date, and full text or extracted passages with pinpoint locations.

Draft a memo with these sections: Facts, Issues, Law, Analysis, Conclusion.

Rules for authority and citations:
- Cite only authorities in AUTHORITIES, by their exact citation string, with the authority id in brackets after each citation, for example [A7].
- Quote only text that appears verbatim in AUTHORITIES, and add the pinpoint location. If you paraphrase, don't use quotation marks.
- Never cite an authority from memory. If the analysis needs an authority that isn't provided, write "AUTHORITY NEEDED:" followed by what should be researched.
- Address adverse authority in AUTHORITIES explicitly. Don't omit it.
- Don't rely on secondary sources as authority.

Rules for analysis:
- Apply the law to the numbered facts, referring to facts by number (F3, F7).
- List every assumption you had to make beyond FACTS.
- Don't base any conclusion on the likelihood that the return will or won't be examined.

Conclusion:
- For each issue, state one confidence level: "should", "more likely than not", "substantial authority", "reasonable basis", or "not supportable". Explain in two to three sentences why it clears that standard and not a higher one.
- State which facts, if different, would change the conclusion.

Output JSON:
{ memo_markdown: "", citations: [{ citation, authority_id, quoted_text_or_null, pinpoint }],
  assumptions: [], authority_needed: [], confidence_by_issue: [{ issue, level, reason }] }

Example

Fictional data.

Client: Halloway Properties LLC (partnership), calendar-year taxpayer
F1 Owns a 14-unit residential rental building placed in service in 2011.
F2 In 2025 replaced the entire roof (decking, insulation, and membrane): $84,000.
F3 Roof replaced with comparable materials; no change in building footprint or use.
F4 Building's unadjusted basis: $1.9M; no prior roof replacement.
ISSUE: Is the $84,000 a deductible repair or a capitalizable improvement
  under the tangible property regulations?
AUTHORITIES: 11 retrieved (Code §§162, 263(a); Treas. Reg. §§1.162-4,
  1.263(a)-3; related IRS guidance and cases)

Conclusion: Capitalize as an improvement (restoration). Confidence: should.
  Replacing the entire roof is generally treated as replacing a major
  component of the building structure under Treas. Reg. §1.263(a)-3(k) [A4].
  Alternative repair treatment: not supportable on F2–F3.
  Would change if: only the membrane was replaced, or only a small section of the roof (F2).
Assumptions: roof is part of the building structure unit of property; the
  building doesn't qualify for the small taxpayer safe harbor (F4 basis exceeds the limit).
Verification report: 14 citations · 13 verified · 1 flagged
  FLAGGED: case citation [A9] not found at the cited reporter page. Removed.
  6 quotations · 6 verbatim matches
Status: blocked from final until reviewer clears flagged item

Guardrails

• Circular 230 §10.37 requires written advice to rest on reasonable factual and legal assumptions, consider all relevant facts, relate the law to the facts, and not take audit likelihood into account. The frozen fact set and explicit assumptions list exist to meet that standard.
• No citation reaches a client unless the verifier confirms it exists and every quotation is verbatim, and the reviewer has read the authorities the conclusion depends on. Existence doesn't prove the authority says what the memo claims.
• Stated confidence levels must match the §6662 and §6694 standards as applied by the reviewer. The model's suggested level is a starting point, not a conclusion, and positions below substantial authority need a disclosure discussion.
• Use research database content only as your license permits. Some licenses restrict bulk export or use with external AI tools, so check before building the retrieval step.
• Client facts go only to an enterprise AI tier with no training on inputs and minimal retention. Where the facts identify a taxpayer, apply the same IRC §7216 and WISP controls as for return data.

What can go wrong

How to prove it worked

Where this goes

Research becomes a firm asset instead of a personal skill. Every memo is grounded in retrieved, verified primary authority and stored with its fact set and confidence level, so the firm builds a searchable body of positions that can be re-verified against new authority automatically. Small firms can offer documented, penalty-protection-grade position support that used to require a national tax department.

Demo

References

• Treasury Department Circular No. 230, §10.37 (requirements for written advice)
• IRC §6662 and Treas. Reg. §1.6662-4(d) (substantial authority standard and types of authority)
• IRC §6694 and Treas. Reg. §1.6694-2 (preparer penalty for unreasonable positions)
• IRS Form 8275, Disclosure Statement
• Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. June 22, 2023) (sanctions for filing AI-fabricated case citations)

Summaries for orientation. Verify against the current standard before relying on any of them.

Keep browsing

• Medium impactCAS and bookkeeping

  Analyze a client's monthly financials and flag what deserves your attention

  Review the month's results against prior periods and the client's own context, and decide what is worth investigating, what to ask the client, and what to tell them.

  Saves 30 to 60 minIntermediateWorkflow
• Medium impactTax

  Catch dropped carryforwards and missing schedules before prep starts

  Compare this year's documents and draft return to the last two or three years and surface what vanished, what was dropped, and what swung without an explanation.

  Saves 10 to 30 minIntermediateWorkflow
• Medium impactTax

  Answer an IRS notice with a reconciled, ready-to-sign response

  Identify the notice, calendar the real deadline, reconcile the IRS's change against the filed return, and draft the response and exhibits.

  Saves 30 to 60 minIntermediateWorkflow