Medium impactIntermediateWorkflow

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.

Instead of a senior person opening each notice, figuring out what it is, digging through last year's workpapers, and writing a letter from a blank page, a pipeline identifies the notice, calendars the real deadline, reconciles the IRS's proposed change against the filed return and source documents, and drafts the response, exhibits, and any penalty relief request. The representative reviews a complete package and signs.

Time
Saves 30 to 60 min
About 1.3 hr down to 25 min per notice
Who does it today
Partners, senior preparers, or an enrolled agent who handles representation
Practice area
Tax

Illustrative planning estimate, not a measured result. Assumes a typical CP2000 or balance-due notice for a 1040 client whose return and source documents are already in the firm's file. Traditional time covers reading and identifying the notice, pulling the return and workpapers, reconciling the numbers, checking authorization, and writing the letter. AI time is reviewer time: checking the reconciliation, editing the draft, and signing. Complex multi-issue notices or missing client documents take longer either way.

Traditional vs AI-native

Traditional

  1. 1Read the notice cold
  2. 2Pull the return and workpapers
  3. 3Rebuild the IRS math by hand
  4. 4Write a letter from scratch
  5. 5Assemble exhibits
  6. 6Check the 2848

AI-native

  1. 1Notice captured and identified
  2. 2Code sets the deadline
  3. 3Authorization checked
  4. 4AI reconciles to the return
  5. 5AI drafts letter and exhibits
  6. 6You confirm the position and sign

Where you sit

You make the representation decisions: agree or disagree, which exhibits prove it, whether to ask for penalty relief. The reading, reconciling, and first draft are done before the file reaches you.

The traditional way

  1. Client emails a phone photo of a notice (often weeks after its date) or it arrives by mail under the firm's CAF number.
  2. A senior person reads it to work out what it is: a matching notice (CP2000), a balance due (CP14, CP501/503/504), a math error (CP11/CP12), a levy notice (LT11 / Letter 1058), an identity verification letter (5071C), or a state equivalent.
  3. Someone pulls the filed return, the source documents, and last year's correspondence to see whether the IRS is right.
  4. The preparer rebuilds the IRS's math by hand and compares it to Schedule D, Form 8949, or the relevant schedule.
  5. A response letter is written from scratch or adapted from an old one, exhibits are assembled, and Form 2848 status is checked.
  6. Penalty relief is often forgotten, or requested inconsistently depending on who handles the file.

Where it hurts

  • Response deadlines are missed because the notice sat in a client's inbox or nobody calendared it from the notice date.
  • Senior time goes to identifying and reconciling routine notices instead of actual controversy judgment.
  • Clients pay CP2000 proposed amounts that were wrong because the review was rushed.
  • First-Time Abate is left on the table for eligible clients.

The AI-native way

  1. 1Automation

    Capture the notice and link it to the client

    Notices arrive through a portal upload, a scanned mail batch, or an email-to-folder address. Each file is queued with a client match based on name, SSN/EIN last four, and address, and unmatched notices go to an admin queue. The original image is stored as the evidence copy, and nothing downstream edits it.

    Tools: Client portal upload · Scanner / mailroom batch · n8n / Power Automate

  2. 2AI

    Identify the notice and extract its key fields

    A vision-capable model reads the notice number (usually printed in the upper or lower right corner), tax year or period, form, notice date, amounts proposed or due, the response-by date if printed, the IRS unit, phone number, and fax or mailing address. It returns strict JSON with the page and region each value came from. If the notice number can't be read, or the notice isn't in the firm's library of known notices, it goes to a human instead of being guessed.

    Tools: Vision LLM with JSON schema output · Azure AI Document Intelligence / AWS Textract

  3. 3Automation

    Calendar the deadline and open a task

    Deadlines are computed by code, not the model. A rules table maps each notice type to its response window: the printed response date where there is one, 30 days from the notice date for a CP2000, 30 days to request a Collection Due Process hearing after an LT11 / Letter 1058, 60 days to request abatement of a math-error adjustment, and 90 days to petition the Tax Court after a statutory notice of deficiency. The practice management task gets the notice, the client, the deadline, an internal target date a week earlier, and the assigned representative.

    Tools: Rules table (spreadsheet or code) · Practice management API (TaxDome, Karbon, Canopy)

  4. 4Automation

    Check authorization before anyone calls

    The system checks the firm's Form 2848 register (or a Tax Pro Account / CAF export) for a power of attorney covering the tax matter, form, and year on the notice. If there's no valid 2848, a pre-filled draft is queued for client e-signature. A Form 8821 alone lets the firm get information but not represent the client or advocate for them.

    Tools: Authorization register · Form 2848 template · E-signature

  5. 5AI

    Reconcile the proposed change against the filed return

    For a CP2000, the model lines up each information return the IRS lists against where it was (or wasn't) reported on the filed return and in the source documents. Common findings: 1099-B proceeds the IRS treated as zero-basis when the basis was actually reported on Form 8949 with box B or E checked; the same income counted twice (a 1099-NEC also included in W-2 wages, or a 1099-K duplicating 1099-NEC receipts); nominee income belonging to someone else; and 1099-R rollovers the IRS treated as taxable. Code recomputes every dollar figure, and the model only explains the differences.

    Tools: Tax software return export · Document store · Python recalculation

  6. 6AI

    Decide the response position and draft the package

    For each item the model proposes agree, partially agree, or disagree, with the evidence behind it. It drafts the response letter in the IRS's format (notice number, tax year, taxpayer ID last four, a clear statement of position, and an item-by-item explanation), fills in the notice's response form, and lists the exhibits in order (8949 page, broker supplemental basis statement, nominee statement). For balance-due notices it drafts payment options or a hold request instead.

    Tools: LLM drafting · Firm letter templates · PDF assembly

  7. 7Automation

    Screen for penalty relief

    For 2025 and later returns, the rules check first asks whether the IRS's Automatic Exemption from Penalty (which begins in summer 2026) should already have kept the penalty from being assessed. If not, it tests First-Time Abate eligibility under IRM 20.1.1.3.3.2.1: the penalty must be failure-to-file, failure-to-pay, or failure-to-deposit; the same return type must have been timely filed for the prior three years; and no penalty other than the estimated tax penalty may have been assessed in that window unless it was later abated for reasonable cause or IRS error. Prior-year penalty history comes from the account transcript. If the client qualifies, the model drafts a phone script for the practitioner priority line or a Form 843. If not, it drafts a reasonable cause narrative from the facts the client gives, and it never invents circumstances.

    Tools: Account transcript data · Rules engine · Form 843 template

  8. 8You

    Representative review and sign-off

    The representative sees the notice image, the extracted fields, the reconciliation table, the draft letter, the exhibits, and the penalty recommendation side by side. They confirm the position, edit the letter, and approve how it will be sent (fax, mail with certified receipt, or the IRS's online document upload where the notice allows it). Nothing goes to the IRS or a state agency without this step, and every edit is logged.

  9. 9Automation

    Send, track, and follow up

    Once it's sent, the proof of mailing or fax confirmation is attached to the task and a follow-up reminder is set (for example, 60 days out, to check the account transcript for resolution). If a later notice on the same issue arrives, it's linked to the same matter so the history stays in one place.

    Tools: Practice management · Account transcript monitoring

What you need

  • The notice (a scan or phone photo is fine)
  • The filed return and source documents in your file
  • A Form 2848 on file for the tax, form, and year
  • A deadline rules table by notice type
  • An enterprise AI tool under no-training data terms

The stack

Capture

Client portal upload, mailroom scanning, email-to-folder

Document AI

Vision LLM or Azure AI Document Intelligence / AWS Textract for notice text and layout

Reasoning & drafting

Enterprise LLM (ChatGPT Enterprise, Claude for Work, Azure OpenAI) under a no-training, zero- or limited-retention agreement

Rules & math

Notice deadline table, FTA eligibility rules, and recomputation in Python or a controlled spreadsheet

Orchestration

n8n, Power Automate, or a small Python service

System of record

Practice management (TaxDome, Karbon, Canopy) plus the firm's document management system

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

Copy the prompt

CP2000 reconciliation and response-position prompt

You are a senior tax representative at a CPA firm analyzing an IRS CP2000 (proposed adjustment from information-return matching). You will receive:
1. NOTICE: JSON extracted from the CP2000 (tax year, notice date, each information return the IRS lists with payer, form, and amount, and the proposed tax, penalty, and interest).
2. RETURN: JSON export of the filed return lines and schedules (including Form 8949 rows with the box checked, Schedule C gross receipts, and 1099-R treatment).
3. SOURCE_DOCS: JSON list of the source documents in the file (form, payer, amounts, and any supplemental statements).

Tasks:
A. For each item the IRS lists, find where it was reported on RETURN, if anywhere. Classify it as: reported_correctly, reported_elsewhere, basis_omitted_by_irs, duplicate, nominee, nontaxable_rollover, truly_unreported, or needs_client_info.
B. For basis_omitted_by_irs: cite the Form 8949 box (B or E) and the basis amount from RETURN or SOURCE_DOCS.
C. Recommend an overall position: agree, partially_agree, or disagree. Explain it in two sentences.
D. List the exhibits needed for each disputed item, and name the specific document from SOURCE_DOCS.
E. Draft a response letter under 400 words: the notice number, tax year, and taxpayer ID last four only; a clear statement of position; an item-by-item explanation; and the exhibit list. Use a professional, non-argumentative tone.
F. Flag for the reviewer any accuracy-related penalty the notice proposes, and whether the facts suggest a reasonable-cause argument. Do not assert facts that aren't in the input.

Rules:
- Never invent an amount, payer, document, or date. If you need a number that isn't in the input, write "UNKNOWN" and add it to open_questions.
- Do not do final arithmetic. Report the figures as they appear in the input, and set recompute_required to true for any item whose tax effect changes.
- Do not cite any IRC section, regulation, or IRM provision unless it is given in the input.
- If the evidence for an item is ambiguous, classify it as needs_client_info rather than choosing a side.

Output JSON:
{ items: [{ irs_item, classification, return_reference, evidence, recompute_required }],
  position: "", position_reason: "", exhibits: [], penalty_flags: [],
  open_questions: [], draft_letter: "" }

Example

Fictional data.

CP2000 and filed return (abridged)

NOTICE: CP2000, tax year 2024, dated 2026-05-11, taxpayer Priya Raman
  1099-B Harbor Brokerage: proceeds $48,200, IRS basis assumed $0
  1099-NEC Lumen Design LLC: $6,500
  Proposed: additional tax $12,410, §6662 penalty $2,482
RETURN: Form 8949 box E (basis not reported to IRS), Harbor Brokerage,
  proceeds $48,200, basis $44,950, gain $3,250, carried to Schedule D line 9
  Schedule C gross receipts $71,300 (includes Lumen Design per invoices)
SOURCE_DOCS: Harbor supplemental basis statement; Lumen invoices $6,500

Pipeline output (reviewer view)

Deadline: respond by 2026-06-10 (30 days from notice date) · internal target 2026-06-03
Form 2848 on file: yes (Form 1040, tax years 2022–2025)
Items:
  Harbor 1099-B: basis_omitted_by_irs · 8949 box E, basis $44,950 · recompute_required
  Lumen 1099-NEC: reported_elsewhere · included in Schedule C gross receipts
Position: disagree. Both items were fully reported on the original return.
Exhibits: (1) Form 8949 page 2 (2) Harbor supplemental basis statement
  (3) Schedule C and receipts reconciliation showing $6,500 from Lumen
Penalty: §6662 falls away if the adjustment is withdrawn; no FTA needed.
Draft letter: ready for review (312 words)

Guardrails

  • Run notices and returns only through an enterprise AI tier under a data processing agreement with no training on firm data and zero or minimal retention. Record the workflow in the firm's WISP as the FTC Safeguards Rule requires.
  • If your firm prepared the return, using that return information to respond to the IRS for the client is generally permitted without separate consent: Treas. Reg. §301.7216-2(b) exempts disclosures to the IRS, and §301.7216-2(m) covers use in an IRS examination. Confirm the IRC §7216 analysis with counsel before routing data to any new third-party service, and when the return was prepared elsewhere.
  • Check Form 2848 coverage for the specific tax, form, and year before anyone contacts the IRS. A Form 8821 alone doesn't authorize representation.
  • Deadlines come from a rules table keyed to the notice type and the notice date, not from the model's reading of the notice. The earliest plausible deadline wins when there's doubt.
  • Circular 230 §10.22 due diligence applies to every factual statement in the letter. The representative confirms each exhibit supports what the letter says it does.
  • FTA and reasonable cause narratives use only facts the client has confirmed in writing. The model is instructed never to invent hardship, illness, or reliance facts.

What can go wrong

The notice type is misread (for example, a CP504 treated as a CP14), which understates urgency.

The notice number is extracted with a confidence score and checked against a whitelist. Any collection-sequence notice escalates to a senior representative the same day.

The client sends the notice weeks after its date, leaving little time to respond.

Deadlines are always computed from the notice date, never the upload date. If fewer than 10 days remain, the system flags the practitioner line call as the first action (to request more time or a hold).

The model agrees with the IRS on an item the client could have disputed, or disputes an item that is actually unreported.

Classification is evidence-based: every disputed item must point to a return line and a source document. Anything classed as truly_unreported gets a second look for nominee or duplicate treatment before the representative accepts it.

State notices vary widely in format and deadline rules.

Start with the two or three states that make up most of your volume, maintain a per-state rules table, and send unknown state notices to a human.

How to prove it worked

Missed or late notice responses

Compare response-sent dates to computed deadlines in the practice management system for every notice over a full year.

Zero

Reviewer minutes per routine notice

Time-track 30 CP2000 and balance-due notices before and after rollout.

Down 50%+

IRS acceptance rate of disputed positions

Track the outcome of each disagree or partially-agree response (closing letter, revised notice, or CP3219A) against the position taken.

At or above the firm's historical rate

FTA capture rate

For every FTF/FTP/FTD penalty notice, record whether eligibility was tested and relief requested.

100% of eligible penalties requested

Where this goes

Notices stop being fire drills. Combined with monthly transcript monitoring, the firm often knows about an adjustment before the letter arrives, has the reconciliation ready, and calls the client with an answer instead of a question. Notice response becomes a packaged, fixed-fee service that a small team can deliver at volume without burning out its senior people.

Demo

Demo video coming soon

Until then, the steps and the example above are the walkthrough.

References

  • IRM 20.1.1.3.3.2.1, First Time Abate (FTA), and the IRS page on penalty relief by administrative waiver (including Automatic Exemption from Penalty)
  • IRS Form 2848, Power of Attorney and Declaration of Representative; Form 843, Claim for Refund and Request for Abatement
  • IRC §6213(a) (90-day Tax Court petition period), §6213(b) (math error assessments and abatement requests), and §6330 (Collection Due Process hearings)
  • IRC §6651 (failure to file and failure to pay penalties)
  • Treasury Department Circular No. 230, §10.22 (diligence as to accuracy)

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

Want to implement this in your firm?

Bring it to the weekly call. Members work through builds like this one together, with accountants who are already doing it.

Keep browsing