Can AI Draft an IRS Notice Response? What to Check Before Sending

Yes, AI can draft a first response to an IRS notice, but only from a redacted file, and nothing should go out until a practitioner has checked every code, amount and date against the notice and the client's transcript and signed it. The model writes clear explanations. It cannot tell whether the IRS or the client is right.

The useful way to think about it is a division of labour. AI does the typing: sorting the notice into a type, building a document checklist, drafting a cover letter and a plain-English client summary. You do the judgement: agree or disagree, which records support which line, what to concede, and when to call rather than write. That split is the review step below.

Follow me on Instagram@sagnikteaches

Triage first: the notice type decides almost everything

An IRS notice is not one thing. The same model prompt that suits a proposed change to a return is wrong for a balance-due reminder. Start by classifying the notice from the code printed on it, then the deadline, then what the IRS is asking for. The IRS publishes a page for each common notice; the three below are examples of how different the asks are.

Connect on LinkedInSagnik Bhattacharya

Beyond type, capture four things in a one-line intake: the date on the notice, the response deadline, the tax year, and the amount at stake. Put the deadline in your calendar before drafting anything. The most expensive mistakes with notices are missed dates, and AI does nothing to prevent them. If the records cannot be gathered in time, the CP2000 page says you can ask for more time through one of its reply options; ask early rather than send a thin response on the last day.

Subscribe on YouTube@sagnikteaches

Some notices should not be answered with a letter at all. A CP504 about a balance the client can pay is a payment problem. A CP14 on a balance the client believes was already paid is a transcript problem, which a call to the number on the notice may settle faster. Triage decides which of those you are in before a model writes a word.

Redact before you draft

A notice carries a name, address, Social Security number or EIN, a tax year and sometimes account numbers. You do not need any of those to get a good draft. Strip them and use placeholders: [taxpayer], [SSN], [notice date], [tax year]. Do this by hand or with a redaction tool, not by asking the model to redact, because the model has already seen the original by then.

Then check three things before any client material goes into a tool: what your engagement letter says about the use of AI, what the tool's plan does with inputs (training, retention, who at the vendor can see them), and what your own professional rules require for tax return information. If you are unsure, ask your professional body or counsel. The tutorial on what clients must sign before their return data goes into an AI tool covers the consent side, and AI tools for tax consultants, compared looks at tools with those questions in mind.

The brief for the first draft, with a sample output

Give the model the redacted facts and tight rules. Tell it the job is a draft for review and make it flag gaps instead of filling them.

You are drafting a response for a tax practitioner to review.
Use ONLY the facts below. Do not cite any IRS form, code section,
deadline or rule that is not in the facts. If something is needed
and missing, write [VERIFY: what] instead of guessing.

Notice: CP2000, dated [date], tax year [year], response due [date]
Proposed change per notice: additional income of $[amount] reported
by [payer type]; proposed tax and penalty $[amount]
Client position: [agree / partly agree / disagree]
Client explanation: [two or three sentences]
Documents we will enclose: [list]

Produce: (1) a one-paragraph summary for the file; (2) a cover letter
to accompany the signed response form, stating our position and listing
enclosures; (3) a plain-English note to the client of three short
paragraphs; (4) a list of every figure and date in the letter so the
practitioner can verify each against the transcript.

An illustrative output for a client who disagrees will include something like:

Re: Notice CP2000 dated [date], tax year [year]
We disagree with the proposed change of $4,200 in other income.
The amount reported on the information return was a refund of
the client's own deposit, not income. Enclosed are the deposit
receipt dated [date] and the supplier's refund confirmation.

Figures to verify:
- $4,200 (notice)  - matches the proposed adjustment line?
- Deposit receipt date [date] - matches the document?
- Response due date [date]

The useful part is the verification list. It turns the review into a pass through a short table rather than a re-read of prose. What you would fix in that draft: the model gave a reason ("refund of the client's own deposit") that came from the intake note, not a document, so confirm the receipt actually says that. And the draft never mentioned the response form, so add it.

A CP2000 on a contractor's 1099-NEC, checked line by line

Say a sole-proprietor electrical contractor receives a CP2000 proposing changes based on $6,800 that a customer reported paying him on a Form 1099-NEC, with proposed additional tax and, as the IRS page warns, interest and possible penalties. The response is due in about five weeks. Triage: proposed change to a return, deadline known, amount moderate.

The practitioner pulls the account and wage-and-income transcripts from her tax professional access and finds the 1099-NEC for $6,800 from a customer. (The tutorial on summarising IRS transcripts with AI covers reading those transcripts at speed.) The contractor's Schedule C included $5,300 from that customer, because $1,500 of the invoice was a materials reimbursement the contractor says he passed through at cost and recorded as an offset, not as income. The AI-assisted routine:

1. Intake and redaction (10 minutes). Notice type, dates, amounts and the contractor's explanation in his words.
2. First draft (5 minutes). Cover letter stating partial agreement: $5,300 was reported, the $1,500 difference is a reimbursement, enclosed are the invoice and receipts.
3. Verification pass (40 minutes). The practitioner checks the $6,800 against the transcript and the notice, the $5,300 against the filed Schedule C, the invoice total against the receipts, and the proposed tax figure against the notice. She finds the invoice shows $6,300, not $6,800, so the customer's 1099 may itself be $500 high. The draft's assumption that "the $1,500 difference is a reimbursement" would have been wrong by $500.
4. Decision. She calls the contractor, confirms the extra $500 was a deposit later refunded, and rewrites the letter to explain both items with documents: $1,000 of materials passed through at cost, and $500 the 1099 overstates. She also decides how the $1,000 should be presented, as the netted figure the client used or as gross receipts with a matching materials expense. That is a reporting judgement for her, not the model. Without the check, the letter would have asserted a number the evidence contradicted.
5. Sign and send. She signs the response form, attaches the enclosures, sends by the method the notice offers (the IRS page lists upload, fax and mail options), keeps proof of sending, and diarises a follow-up.

The AI saved perhaps half an hour of typing. The check caught a $500 error that would have undermined the whole response.

The verification list: what to check line by line

Codes and form numbers

Any IRS notice code, form number or code section in the draft must appear in the notice or in your own notes. Models mix up notice series and cite the wrong form. Cross-check against the IRS page for that notice type.

Amounts

Compare every dollar figure to the notice and to the transcript. Check totals by hand. If a figure appears in the draft that is in neither, it was invented, even if it looks plausible.

Dates

The notice date, response due date, tax year, and any dates of events described. A model can change a year digit without any visible sign. Pull each date next to its source.

Account details

Make sure the placeholders were replaced with the correct taxpayer and the right tax year for this client. In a busy season the most common real error is a letter with the wrong client's details after a copy and paste.

Position and promises

Does the letter say agree, partly agree, or disagree, matching what the client decided? Does it promise anything, such as a future filing or a payment, that the client has not committed to? Remove it.

Enclosures

Every document named in the letter must be attached, legible and redacted of anything the IRS does not need. Do not let the model describe a document it has not seen.

Authority

Can you speak for the client on this notice? The IRS page for the CP2000 says you can allow someone to contact the IRS about the notice by completing the authorisation section on the response form, or send a Form 2848, Power of Attorney and Declaration of Representative, for broader representation. Check which applies before you send.

Tone

Factual, courteous, short. Delete adjectives, emotional appeals and statements about what the IRS "should" have done.

A CP14 for a balance the client says she paid

A cleaning company owner receives a CP14 for $2,350. She says she paid it in March. The model drafts a letter saying "payment was made on March 14". The practitioner's check: bank statement shows a payment of $2,350 on March 14, but the transcript shows it applied to a different tax year. The response is then not a dispute about whether she owes. It is a request to move a misapplied payment. The IRS page says a CP14 is about unpaid tax, that interest accrues on the unpaid amount after the notice date, and that to dispute it you call the number in the notice's help section with your supporting documents ready. So the practitioner phones with the bank record and the transcript open, asks for the payment to be moved to the right year, and writes a short confirming note to the file afterwards. The AI's draft was correct on the facts it was given and wrong in its theory of the case. Only a person reading the transcript sees that.

A CP504, where the right reply is rarely a letter

A roofing company owner receives a CP504, the IRS's notice of intent to levy. The IRS page says to pay immediately, make a payment plan if you cannot pay in full, or request an appeal under the Collection Appeals Program before collection action by following the instructions on the notice. The page also sets out what follows if nobody responds: a possible Notice of Federal Tax Lien, a levy on a state tax refund, then other property such as wages and bank accounts. It adds that even a client who has already paid or set up a plan should call so the account reflects it. A model asked to "draft a response" will produce a persuasive letter explaining hardship. That is rarely the right first move. Triage says the deadline is urgent and the response is a payment plan application or a call. Use the AI for the client summary, a checklist of financial records a plan application needs, and a diary of deadlines. Keep the strategic decision for yourself.

When the transcript says the IRS is right

An illustration to keep in your pocket: a plumbing firm's owner insists a proposed $3,100 adjustment is wrong, but the transcript shows the 1099 was correct and the income was left off the return. The draft writes itself in the client's favour because that is what the prompt asked for. The check is to read the transcript before the client's story. If the notice is right, the response should say so, ask about the penalty and the payment options, and perhaps include an amended return. Drafting a disagreement the records contradict wastes time and can damage credibility with the IRS. The model will not warn you. Add one line to your brief: "If the facts appear to support the IRS position, say so in a note to the practitioner."

The client note: lowest risk, still checked

This is the part where AI earns its keep with the least risk. After you have decided the position, ask for a three-paragraph note to the client: what the notice says, what we are doing, what we need from you and by when. Sample output, to be edited:

Hi [first name], the IRS has written about your [year] return. A
customer reported paying you $6,800 and the IRS thinks part of that
was left off. We think $1,000 of it was materials you passed on at
cost and $500 was a deposit you refunded, so the proposed change is
too high. We have drafted a reply and need two things from you by
[date]: the supplier receipts for the materials and the bank record
of the refund. Nothing is due to the IRS yet, and we will send the
reply well before the deadline of [date].

Check it against the facts: "nothing is due" is only true if the notice is a proposal, not a bill. Do not leave sentences like that in without checking.

Where models slip on IRS notices, and how it shows up

After a few dozen drafts you will see the same failures. Knowing them makes the verification pass quicker because you read with a short list of suspects in mind.

• Confident wrong rule. The draft says "the IRS has three years to assess" or "you have 30 days to appeal" when the notice says something else. It shows up as a sentence with no source. Delete any rule that is not on the notice.
• Merged documents. A figure from the 1099 and a figure from the Schedule C get blended into one sentence. It shows up as an amount that appears nowhere on either page.
• Silent year change. A prior-year date slips into the timeline. It shows up when your date list runs out of order.
• Over-agreement. If the brief says the client "may owe", the model drifts into conceding the full proposed tax. It shows up as an apologetic tone and a sentence agreeing to pay.
• Invented enclosure. The cover letter lists "bank statements" when you only have one. It shows up when the enclosure list is longer than the file.

A matching table to build before you write

For any notice with figures, build a small table before drafting and paste only its rows into the brief. It takes ten minutes and replaces the guesswork a model would otherwise do. Illustrative rows for a CP2000-type case:

When every row has a position, the draft is mostly a sentence per row, and the model is arranging, not deciding.

When to stop drafting and call

Pick up the phone, or escalate inside the practice, when the notice involves identity theft (the CP2000 page points to Form 14039 for that), a levy or lien already in motion, a balance the client cannot pay, an amount above your review threshold, or a mismatch you cannot explain from the transcript. The CP2000 page says the Taxpayer Advocate Service may be able to help, and the CP14 page points to it for penalty problems you cannot resolve on your own. A model can prepare the call notes, so ask for a list of the three questions you need answered and the documents to have to hand.

A routine for the practice, with time and checks

1. Log the notice on arrival: client, type, date, deadline, amount. Two minutes.
2. Triage into pay, dispute, correct or call. Five minutes with the IRS page for that notice open.
3. Pull the transcript and compare to the notice before anything is drafted.
4. Redact and draft with the brief above. Five minutes.
5. Verify against the list. The time depends on the case; a simple one takes 20 to 40 minutes.
6. Practitioner review and signature. A second person checks any case above a threshold you set, for example a proposed liability over $5,000.
7. Send and record. Keep proof, diarise follow-up, save the final letter and enclosures.

For busy-season volume, the tutorial on how small tax practices use AI through the busy season covers where else the hours go, and building an AI tax organiser that chases missing documents helps you collect the enclosures sooner. If you want help setting up the routine so your team can run it, that is what my AI implementation consultation covers.

One boundary is worth stating plainly. This tutorial describes a drafting and review method for practitioners. It is not advice on any client's case, and IRS pages, forms and deadlines change, so always work from the notice in front of you and the IRS page for its type.

Questions tax practices ask about AI and IRS notices

Can I upload a client's IRS notice to ChatGPT or Claude?

Redact first: names, Social Security numbers, EINs, addresses and account numbers. Then check your engagement letter, the tool's data terms and your confidentiality duties for return information. The model needs notice type, tax year, amounts and dates, not identifiers.

Should AI write the whole response?

No. Let it draft the structure, the explanation and the document list. The practitioner decides whether to agree or disagree with the notice, checks every figure against the transcript, and signs. A model cannot know what the client's records show.

What does the IRS say to do with a CP2000?

Reply by the date on the notice. The IRS says it is not a bill and a response may be required. If you agree and have nothing else to report, follow the notice's instructions; if you disagree, say so on the response form with supporting documents.

How do I stop a model inventing IRS code sections?

Tell it not to cite any code section, form or deadline that is not in the notice or in your own notes, and to write [VERIFY] where it would otherwise guess. Then check every reference it did include against the notice and the IRS page for that notice type.

Further reads

• AI Tax Research Tools Compared: Citations, Cost and Checks — Citation checks for AI tax research.
• Year-End Tax Planning Letters With AI: Personalise Without Errors — Personalised client letters without factual errors.
• AI Software for Tax Professionals: Run a Pilot and Measure Cost per Return — Pilot AI in the practice and measure the result.
• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — AI slips that cost firms client trust.
• Is It Safe for an Accountant to Use ChatGPT With Client Data? — What client data can safely enter an AI tool.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: IRS pages 'Understanding your CP2000 series notice', 'Understanding your CP14 notice' and 'Understanding your CP504 notice' (irs.gov, read 4 November 2026).