Module S3 — Pre-signature anomaly review

The 30-second version. A 15-minute AI-assisted review pass before Form 8879 catches the four anomaly classes that produce most preparer-penalty exposure: cross-year deviations, missing-form omissions, math reconciliation breaks, and untenable position calls. The classes the AI cannot catch — cited-authority validation, client-fact-pattern judgment, professional-skepticism flags — the practitioner still owns. Together this is the §10.22 due-diligence trail running in the background of every signed return. Precondition: §7216 consent + DPA on file per Module S1 for every vendor used.

What this module assumes

You have:

1. Completed Module S1 — §7216-defensible AI setup for the AI vendor you’ll use. The prompts in S3 disclose tax return information to a third-party processor. Running them without the consent + DPA on file is a §7216 disclosure without consent.
2. Read the Circular 230 + AI article — particularly the §10.22 due-diligence section. The article explains why the review-trail discipline matters; this module is the how.
3. A current ChatGPT Team / Claude Team / M365 Copilot / Blue J subscription per the S1 vendor decision tree. The prompts assume Claude Team or ChatGPT Team for the general-purpose pass and Blue J for the substantial-authority verification pass.

If any of these are missing, fix them first. The S3 workflow does not work safely without them.

The Monday-morning sequence (15 minutes per return once habituated)

The pre-signature review is a 7-step pass. Steps 1-2 are one-time per return; Steps 3-6 are the prompt-driven review; Step 7 is the audit-trail log entry.

Step 1 — Sanitize the return for AI input (3 minutes)

The §7216 consent allows the disclosure; it doesn’t license sloppy disclosure. The principle: send the AI the minimum information needed to do the task. Use Artifact 1 — Return sanitization checklist below.

For a typical 1040: replace the taxpayer name with first initial + last initial (“JA” not “James Anderson”); mask the SSN to the last four; replace dependent SSNs identically; blank the bank account numbers. Keep all numbers (income, deductions, withholding, basis, AGI), all form references, all dates, all entity names that are not the taxpayer’s. The AI needs the math + structure, not the identity.

This is faster than it sounds. After 3-4 returns the redaction pattern lives in muscle memory.

Anti-pattern: “I’ll just give the AI the raw return — it’s signed under DPA.” The consent is legal cover, not operational discipline. Sanitization is the second layer that protects against the worst case where the vendor’s DPA fails. Defense in depth.

Step 2 — Open the standard prompt set (1 minute)

Five prompts. They go in order. Run from one tool (Claude Team or ChatGPT Team) — same conversation thread — so each prompt builds on the prior. Use Artifact 2 — Pre-signature prompt set below.

Step 3 — Run the cross-year deviation prompt (2 minutes)

Paste the sanitized current-year return + prior-year return. Run Prompt 1. Expected output: a numbered list of line items where the current-year value differs from prior-year by more than 25%, with a one-line explanation candidate for each.

For each flagged deviation, ask: did the underlying fact change (W-2 increase, new K-1, new dependent), or is it a math/import error? If the former, you already know. If the latter, fix the return. The AI is competent here — false-positive rate is acceptable (you’ll dismiss 60-80% of flags); false-negative rate on cross-year deltas is low.

Step 4 — Run the missing-form prompt (3 minutes)

Same thread. Run Prompt 2. Expected output: forms that should exist based on prior-year forms + current-year inputs but are absent or zero in the current draft. Common catches: Schedule B missing despite 1099-INT income; Form 8938 missing despite foreign-asset disclosure; Schedule D/8949 missing despite a brokerage statement; Form 8606 missing on a non-deductible IRA basis carry; Schedule SE missing on Schedule C net earnings over $400.

For each flag: did you intentionally omit it, or miss it? If missed, add before signing. Document the intentional omissions in the audit-trail log — the §10.22 trail wants to see you considered each flag. AI is competent on standard forms, weaker on niche forms (5471, 8865, 8621, 706, 709).

Step 5 — Run the math reconciliation prompt (3 minutes)

Same thread. Run Prompt 3. Expected output: computed quantities that don’t tie, with the discrepancy magnitude. Common catches: 1099-DIV box 1a not equal to Schedule B; Schedule C net profit not equal to line 31; basis-by-disposition not equal to Form 8949 basis total; §199A QBI aggregation breaks; missing AMT add-backs; §163(j) interest carryforward errors. Each discrepancy is either a math error (fix) or a known difference (document). Competent on form-to-form ties, weaker on multi-form linkages — run Prompt 3.5 (basis carryforward trace) as an extension for complex returns.

Step 6 — Run the deductibility-flag + position-scan prompts (3 minutes)

Same thread. Run Prompt 4 and Prompt 5 (paired).

Prompt 4 (deductibility): line-by-line list of claimed deductions with limit/threshold/eligibility rules, the rule cited, and the test each passes. AI catches §170 charitable AGI limits, §163(h) mortgage-interest acquisition-vs-equity, §469 passive-activity tests, §199A wage/UBIA limits, §174A R&D expensing, SALT cap, §183 hobby-loss exposure.

Prompt 5 (position scan): any return position that depends on a substantial-authority claim under §10.34. Do NOT treat the AI’s “substantial authority exists” call as substantial authority. The AI’s training cutoff and its tendency to hallucinate cites means an AI-asserted substantial-authority claim is a prompt to verify in Blue J (or Checkpoint, CCH AnswerConnect, Bloomberg Tax), not a final answer. The §10.34 + Mata + Thomas v. Commissioner reflex fires here.

This is the most important calibration in the module: the AI is competent at flagging deductibility issues; it is incompetent at validating substantial authority. AI flags, practitioner verifies in primary authority. Never invert this.

Step 7 — Log the review pass (2 minutes)

Per §10.22 due-diligence trail expectations. Use Artifact 3 — §10.22 review-trail log below. The log entry takes 60-90 seconds per return. The discipline is what survives an examination, not the speed.

That’s the workflow. Seven steps. Fifteen minutes per return at steady-state. Over a tax season — 250 returns × 15 min = ~62 hours of structured pre-signature discipline that wasn’t there before. The §10.22 trail goes from “I reviewed the return” (handwave) to “I ran the documented 5-prompt review and these were the flags” (defensible).

Artifact 1 — Return sanitization checklist

Run through it once per return.

Always remove (replace with mask): taxpayer name → first + last initial; SSN → last four only; spouse + dependent names + SSNs → masked; bank/routing numbers → blank; driver’s-license numbers → blank; state preparer numbers → blank; client street address → city + state only; phone + email → blank.

Always KEEP (the AI needs these): all income amounts; all deduction amounts; all adjustments + carryforwards; all entity names that are not the taxpayer (mutual funds, public companies, broker names, partnership names); all form + line numbers; all dates; all percentages; all state tax codes; filing status.

Optional — keep if relevant, mask if sensitive: rental property addresses (city + state usually fine); K-1 issuer names (keep for entity-character reasoning); broker names (keep — helps identify 1099-B form types).

Format for AI ingestion: paste as Markdown-style text or as a CSV of line-item-name + amount + form. Both parse well. Pasting a PDF works but is slower and more error-prone — text input is better.

Artifact 2 — Pre-signature prompt set

Run from one tool in a single conversation thread.

Setup prompt (run once at the start of the thread):

You are assisting a US tax practitioner in a pre-signature review pass on a
1040 return. The practitioner has signed Section 7216 consent + DPA covering
this disclosure. Your role is to surface flags for the practitioner to review.
You do NOT make conclusions, you do NOT recommend tax positions, and you do
NOT validate substantial-authority claims. You flag potential issues with a
short explanation. Each flag includes the form / line / amount + a one-line
read on why it merits review. Numbered list format. The practitioner will
paste a sanitized return + prior-year return in the next message.

Prompt 1 — Cross-year deviation:

The two attached returns are the same taxpayer's prior year (TY-1) and current
year (TY) drafts. Identify every line item in the current year that differs
from the prior year by more than 25% in either direction. For each:
1. Form + line number
2. Prior-year amount, current-year amount, delta as a percentage
3. One-line plausible explanation (W-2 raise, new K-1, dependent change, lost deduction)
4. Flag severity: HIGH if delta over 100% or a typically-stable carry (basis, depreciation), MEDIUM otherwise
Do not include lines where both years are zero or the delta is under 25%.
Output as a numbered list.

Prompt 2 — Missing form scan:

Based on the current-year return draft + the prior-year return, list every form
that you would expect to appear on the current-year return but does not, OR
that appears with zero values when prior-year had non-zero values. For each:
1. Form name + number
2. Reason for expecting it (prior-year inclusion, current-year input data, regulatory requirement)
3. Severity: HIGH if Section 6694 preparer-penalty exposure attaches to the omission, MEDIUM otherwise
Include FBAR/FinCEN 114 reminder if foreign-financial-asset disclosure was made
on Schedule B. Output as a numbered list.

Prompt 3 — Math reconciliation:

For the current-year return, identify every computational quantity that does
not tie to an upstream input, or where two related computations don't
reconcile. For each:
1. Form + line of the unreconciled quantity
2. Upstream source that should tie (the 1099, the K-1, the basis carryforward)
3. Discrepancy magnitude (in dollars and as a percentage of the affected line)
4. Severity: HIGH if discrepancy over $100 or over 5% of line, MEDIUM otherwise
Focus on: 1099-DIV / 1099-INT / 1099-B reconciliations; K-1 allocations to
Schedule E / QBI; basis carryforwards across years; Section 163(j) interest
limits; Section 199A QBI aggregations; AMT add-backs. Output as a numbered list.

Prompt 3.5 — Basis carryforward trace (extension for complex returns):

For each pass-through investment on the current-year return (S-corp,
partnership, MLP, BDC), trace the basis carryforward:
1. Entity name
2. Prior-year ending basis
3. Current-year contributions / distributions / income / loss allocation
4. Computed current-year ending basis
5. Tax-form basis reported (Form 7203 for S-corps where required; partnership basis statement)
6. Match status: MATCH / DISCREPANCY / UNCONFIRMED
Output as a numbered list.

Prompt 4 — Deductibility flags:

For the current-year return, identify every claimed deduction that has a
specific limitation, threshold, or eligibility test. For each:
1. Deduction name + form/line
2. Statutory citation for the limit/test (IRC section)
3. Test the deduction must pass
4. Did the deduction pass? (from the data available) — YES / NO / INSUFFICIENT_DATA
5. If NO or INSUFFICIENT_DATA: severity (HIGH / MEDIUM)
Focus on: Section 170 charitable AGI limits; Section 163(h) mortgage-interest
acquisition vs equity; Section 469 passive-activity (especially 469(c)(7) REP);
Section 199A wage/UBIA + SSTB exclusion; Section 174A R&D expensing post-OBBBA;
SALT cap ($40K post-OBBBA); Section 163(j) interest limit; Section 183
hobby-loss exposure. Do not opine on whether a deduction is "wise" — only on
whether it passes the statutory test. Output as a numbered list.

Prompt 5 — Position scan (the most-important calibration):

For the current-year return, identify any tax position that depends on a
substantial-authority claim under IRC Section 6662 / Circular 230 Section 10.34.
For each:
1. Position taken (form + line + dollar magnitude)
2. The provision the position relies on (IRC section, regulation, ruling, holding)
3. Your best read on whether substantial authority exists (with the caveat below)
4. A short list of primary sources the practitioner should verify before signing

CRITICAL CAVEAT: Do not assert that substantial authority exists. List the
primary sources for the practitioner to verify directly. Your training data may
be stale, incomplete, or contain hallucinated citations. The practitioner is
the Section 10.34 decision-maker. You are the search-pointer, not the
legal-conclusion engine. Output as a numbered list.

Prompt 5 is the most-important calibration in the module: the entire failure pattern of the Mata v. Avianca / Thomas v. Commissioner / Delano Crossing / Dehghani federal-court canon is substantial-authority assertion based on AI-generated content that wasn’t independently verified. The prompt inverts the failure mode — the AI becomes the search-pointer; the practitioner verifies in primary authority before signing.

Calibration table — print it, tape it next to the monitor:

Prompt	AI competence	False-positive rate	False-negative risk	Verification needed?
1 — Cross-year deviation	HIGH	60-80% (dismissible)	LOW	Spot-check only
2 — Missing form scan	HIGH standard; MEDIUM niche	40-50%	LOW standard, MEDIUM niche	Verify niche forms manually
3 — Math reconciliation	HIGH form-to-form; MEDIUM multi-form	20-30%	LOW	Spot-check the linkages
4 — Deductibility flags	HIGH common; MEDIUM edge	30-40%	LOW-MEDIUM	Verify edge cases in primary authority
5 — Position scan	LOW (search-pointer only)	HIGH	HIGH	Mandatory: verify every position in Blue J / Checkpoint / Bloomberg Tax / CCH AnswerConnect

Artifact 3 — §10.22 review-trail log

Extends Module S1’s audit-trail log (Tab 2 — Use log) with four new columns for the pre-signature review pass.

Date	Client	Engagement ID	Vendor	Use type	Prompt summary	Output kept?	Reviewer	Prompts run	Flags surfaced	Flags resolved	Substantial-authority verified?
2026-03-04	Anderson	2026-1040-001	Claude Team	Pre-signature review	5-prompt pass on TY26 1040	Y	DT	1,2,3,4,5	14	14	Y — Blue J verified 469(c)(7) REP claim
2026-03-04	Brown	2026-1120S-007	Claude Team	Pre-signature review	5-prompt pass on TY26 1120-S	Y	DT	1,2,3,4	6	6	N/A — no 10.34 positions

What the four added columns capture: which prompts you ran; the raw flag count; the resolved count (should equal surfaced — if not, you didn’t finish the review); and whether you ran any §10.34 position through primary authority. The defensive use case: if a return is later examined and OPR asks “what AI did you use and how did you check it?”, the log answers directly — “Claude Team. 5-prompt pre-signature pass. 14 flags. 14 resolved. Verified the §469(c)(7) REP claim through Blue J. Here’s the Anderson review folder.” That is the §10.22 trail.

Artifact 4 — Where the AI review pass fits in the standard prep cycle

Per the standard 12-day return-prep cycle, the AI review pass slots in between Step 3 (self-review) and Step 4 (reviewer queue):

Day 0-2   Step 1 — Engagement + intake
Day 2-3   Step 2 — Initial preparation (preparer, 2-4 hrs)
Day 3     Step 3 — Self-review (preparer, missing-doc check)
          v
  >>>     Step 3.5 — AI PRE-SIGNATURE REVIEW PASS (this module, 15 min)
          v
Day 3-5   Step 4 — Sit in review queue (the bottleneck)
Day 5-7   Step 5 — Reviewer first pass (30-60 min)
Day 7-8   Step 6 — Preparer cleanup
Day 8-10  Step 7 — Partner sign-off
Day 10-12 Step 8 — Client review + Form 8879 signature
Day 12-14 Step 9 — E-file submission

The AI pass is upstream of the human reviewer queue, not a replacement. Two reasons: the reviewer queue is the actual bottleneck (returns sit 5-10 days during peak), so AI catching anomalies before the queue means cleaner returns reach the reviewer; and AI is incompetent at professional skepticism, client-relationship judgment, and fact-pattern interpretation — exactly the reviewer’s job.

Anti-pattern (the most dangerous in this module): running the AI review pass and skipping the human review that follows. “The AI flagged nothing, the return must be clean” is the exact failure mode the Mata-Thomas canon punishes. The AI pass augments §10.22 due diligence; it does not replace it.

Artifact 5 — The pre-signature physical checklist

Tape it next to the monitor. Run through it before clicking Submit on Form 8879.

                     PRE-SIGNATURE CHECKLIST
                     Client: _______  Engagement ID: _______  Date: _______

Section 7216 + S1 prereqs
[ ] Section 7216 consent on file for AI vendor used (Vendor: _______)
[ ] DPA on file for the same vendor
[ ] Return sanitized before pasting (Artifact 1 checklist run)

S3 AI review pass
[ ] Setup prompt sent
[ ] Prompt 1 — Cross-year deviation run; flags reviewed
[ ] Prompt 2 — Missing form scan run; flags reviewed
[ ] Prompt 3 — Math reconciliation run; flags reviewed
[ ] Prompt 4 — Deductibility flags run; flags reviewed
[ ] Prompt 5 — Position scan run; ALL substantial-authority claims verified
    in Blue J / Checkpoint / Bloomberg Tax / CCH AnswerConnect (NOT just AI)

Human review pass (Section 10.22 due diligence — not delegable)
[ ] All AI-flagged items resolved in return or documented as intentional
[ ] Client-fact-pattern judgment exercised
[ ] Material new facts (life events, new income, new dependents) verified
[ ] Section 10.22(a) due diligence — preparer signature + Form 8879 ready

Documentation
[ ] Section 10.22 review-trail log (Artifact 3) entry made
[ ] Annual Section 7216 consent expiry checked (S1 Step 6 reminder confirmed)

Signed: _______  Date: _______

FAQ

Do I need to run all 5 prompts on every return?

No. Returns with no §10.34 positions skip Prompt 5. Long-standing clients with minimal year-over-year change may skip Prompt 1 in some cycles. Never skip Prompt 5 if there’s any position claim; never skip Prompt 2 (missing-form scan) on any return — it catches the omission failures that produce §6694 exposure.

What about returns where the AI catches nothing?

Either genuinely clean, or the AI didn’t parse the return. Spot-check: the AI should restate some specific numbers back to you. If it does, the parse worked. A clean AI-pass on a non-trivial return is rare; treat zero flags as a signal to verify the parse before signing.

Doesn’t this slow the prep cycle by 15 min per return?

Yes, and it speeds the reviewer queue, reduces post-signature corrections, and cuts §6694 exposure. Net per-return time tends to drop after 30-50 returns. Realistic savings: 30-60 minutes per return where the AI catches a math issue that would have surfaced at partner or client review.

How does this work for non-1040 returns?

Same five prompts with form-specific adaptations: 1065 (partner-allocation reconciliation replaces cross-year); 1120-S (add §1366 basis check); 1041 (add §663(b) 65-day election check); 706 (skip prompts 1+2, intensify prompt 4 on valuation + transfer-tax positions).

What if the AI flags a position I know is correct but says “no substantial authority”?

The calibration working. The AI’s “no substantial authority” call is unreliable (Prompt 5 has HIGH false-positive). Verify in Blue J / Checkpoint / Bloomberg Tax / CCH AnswerConnect. If primary authority supports it, document the citation and sign. The AI is a search-pointer; you are the practitioner.

Related reading

• Circular 230 and AI: §10.22, §10.35, §10.37 — the why-this layer; read first if you haven’t
• Circular 230 §10.22 due-diligence checklist — the workflow-level defensibility scorecard
• Module S1 — §7216-defensible AI setup — the precondition. Do S1 first.
• Module S5 — OBBBA rename trap — the rename-trap discipline plugs into this module’s Prompt 5
• AI for FBAR Compliance: FinCEN 114 + §10.22 Workflow — the foreign-info-return extension of this review-trail discipline

---

If you’d rather have the pre-signature AI review workflow set up across your firm — vendor evaluation, prompt-library customization, audit-trail log integration into your PM system, staff training on the calibration discipline — Tunderman, the publisher of this site, does AI implementation for tax practices. Typical engagement: 1-3 weeks for a solo / small-firm setup. Reach us at editorial@aitaxpractitioner.com.