Tax notice and response correspondence data for tax-resolution AI

What a notice-to-outcome record must contain

A training-grade record is a full case thread, not a single notice. Notice classification needs only the first document, but response drafting and outcome prediction need the chain that follows: notice, response, IRS or state reply, and closure. Without the closure, you cannot tell a winning response from one that triggered the next notice.

The CP2000 shows why. It is a proposal generated when third-party information returns (W-2, 1099 and similar) do not match the filed return; it is not a bill, the taxpayer may agree or disagree with documentation, and an unresolved case moves to a CP3219A, Statutory Notice of Deficiency, which carries Tax Court rights. A CP2000 record that stops at the response letter hides whether the case closed with no change, a partial adjustment or a deficiency notice.

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "case_id": "c_7f3a91",
  "jurisdiction": "IRS",
  "notice_code": "CP2000",
  "notice_date": "2025-06-16",
  "tax_year": 2023,
  "issue_type": "underreported_income",
  "issue_detail": ["1099-B proceeds without basis", "1099-INT not reported"],
  "proposed_change_usd": 4180,
  "notice_text_path": "notices/c_7f3a91_p1-4.txt",
  "notice_image_path": "notices/c_7f3a91_p1-4.pdf",
  "response": {
    "position": "partial_disagree",
    "letter_path": "responses/c_7f3a91_letter.txt",
    "attachments": ["broker_basis_statement", "form_8949_corrected"],
    "channel": "fax",
    "sent_date": "2025-07-08"
  },
  "agency_replies": [{"type": "revised_proposal", "date": "2025-10-02"}],
  "outcome": "partial_adjustment",
  "final_change_usd": 312,
  "elapsed_days": 108,
  "preparer_role": "resolution_firm_not_original_preparer",
  "deid_method": "tin_name_address_replaced; scanline_masked"
}

The fields that matter most for modeling are issue_detail (multi-label, at the information-return level), position, attachments and the pair outcome plus final_change_usd. Ask whether amounts were preserved, bucketed or perturbed, because response-drafting models learn arithmetic reconciliation from the real deltas.

Notice families to scope by name

Buyers get better matches when they request notice families, not "IRS letters." Each family implies a different task, response pattern and outcome vocabulary, so a corpus heavy in one family will not generalize to another.

• Underreporter (CP2000 and the CP3219A that follows): reconciliation against information returns; outcomes range from no change to a deficiency notice. Best for extraction plus reasoning over amounts.
• Balance due (CP14 and the CP501, CP503 and CP504 sequence): mostly payment, installment and collection-alternative responses; good for routing and escalation prediction.
• Collection and levy (LT11, Letter 1058): time-sensitive, with collection due process rights; responses involve hearing requests and alternatives such as installment agreements.
• Penalty assessments and abatement requests: a well-bounded drafting task, covered below.
• Identity, examination and correspondence audits: document requests and verification letters; often thin on outcome labels.

Notice numbering and wording change over time, so confirm codes against the current IRS notice catalog [6] for the tax years you license and record the notice revision date in the metadata. Practitioners also see agency letters identified by letter numbers rather than CP codes; ask suppliers to map both into one taxonomy.

Penalty abatement letters as a supervised drafting task

Penalty abatement correspondence is the cleanest drafting target in this domain because each request has a defined legal ground and a binary-ish outcome. Requests generally rely on first-time abatement, an administrative waiver for taxpayers with a clean recent compliance history, or on reasonable cause, which turns on facts such as illness, disaster or reliance on advice. A good record captures which ground was claimed, which penalty (failure to file, failure to pay, failure to deposit) and whether relief was granted, partially granted or denied.

Label the request ground separately from the narrative. A model trained on letters without the ground label learns to write plausible prose but not to choose the right argument, which is the actual skill. Denials are valuable: a corpus of only granted requests teaches survivorship, not judgment.

Quantity matters less than curation here. The LIMA results showed that about 1,000 carefully chosen prompt-response pairs can drive strong instruction following [5], so a few thousand well-labeled abatement letters with outcomes can beat a much larger, unlabeled dump of correspondence.

State notices need explicit coverage terms

State tax correspondence varies by state agency, tax type and notice format, so the coverage statement belongs in the license. Specify states, tax types (individual income, sales and use, withholding, franchise), tax years and whether notices are linked to the federal case they derive from, such as a state assessment following an IRS adjustment.

Expect heavy imbalance. A firm's book may be dominated by a few high-population states and sales tax, with a long tail of single-digit notice types. Ask for per-state, per-tax-type counts before you sign, and decide whether the long tail is training data or evaluation data.

Rights and privacy: Section 7216 and the case file

Who holds the file decides the legal path. When a return preparer holds tax return information, Section 7216 and its regulations make knowing or reckless use or disclosure for purposes other than return preparation a criminal matter unless an exception or valid consent applies [1]. A resolution firm that also prepared the return is in a different position from one engaged only for representation, and both must be assessed. Our guide to Section 7216 consent and what CPA firms can license covers that analysis in depth.

Notices are dense with identifiers: SSN or EIN, name control, address, the remittance voucher scan line, and in the representation file, Forms 2848 or 8821 with CAF numbers and preparer PTINs. Names, emails, phones and account numbers should be removed or replaced before delivery, but de-identified data can still be re-identified, as NIST SP 800-188 documents [2]. Rare amounts, small-town addresses and unusual fact patterns in reasonable-cause narratives are the residual risk; request the method record and a sample review. See the buyer's guide to de-identified data for AI training for test patterns.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Document formats and extraction pitfalls

Most notices arrive as scanned or e-faxed PDFs, so plan for OCR noise. Public research sets help pretraining but are generic: RVL-CDIP-style work targets coarse document-class recognition [3] and FUNSD covers 199 noisy scanned forms with entity linking [4]. Neither contains tax notice semantics, which is why licensed case files are needed for code, tax year and amount extraction.

Common failure modes to test for in samples:

• Multi-page notices split into separate files, or response attachments merged into the notice PDF.
• Notice dates confused with response-due dates, which skews elapsed-time labels.
• Redaction applied to images but not to OCR text layers, or the reverse.
• Outcomes inferred from silence ("no further notice") rather than a closing letter or account transcript entry.

If the supplier can attach account transcript transaction codes to the case, outcome labels become verifiable rather than inferred. Agree file layout, encoding and manifest format up front; our dataset delivery formats guide lists the options, and the document AI datasets guide covers annotation layers.

Buyer request checklist for notice correspondence

Write the request as data requirements that describe the records, not a list of target firms. A precise request is the clearest way to learn whether a match exists.

Illustrative example: invented to show structure; it does not describe an available dataset.

Hold out a slice for evaluation before training begins, ideally by firm and by tax year, to avoid leakage of house style and template text. Our guide to agent evaluation task suites shows how to grade multistep resolution agents on end state rather than prose.

Related correspondence datasets

Tax notice response data shares structure with other adjudicated-correspondence corpora. Claim denial appeal letters with outcomes and chargeback representment case data have the same request, evidence and decision chain, and techniques for labeling them transfer. For the research side of tax work, see tax research memos with authority citations, and the industry-specific operational data hub lists other verticals.

On the SourceX side, the accounting buyers page and finance and accounting AI training data describe adjacent categories. SourceX sources operational datasets from US companies, including finance and legal workflows and documents, on request rather than from stock, and a request does not guarantee a match. You can describe the notice data you need as data requirements.

Sourcing IRS notice datasets with SourceX

SourceX looks for US businesses that hold the data you describe, assesses the data and licensing permissions, and agrees pricing and allowed uses in a license; nothing is contracted until a supplier agrees. Each dataset is rights-reviewed, personal details are removed or replaced before delivery with the method recorded, and delivery runs through private, access-controlled workflows after an executed agreement. Start a tax notice correspondence data request.