The IRS's robo-audit system is firing off CP2000 notices faster than ever this year, and crypto 1099-DA mismatches are the new top trigger. Here is the Claude prompt I'd run the moment one lands on a client's desk, and the three places it will get you in trouble if you trust it.

A CP2000 is not a bill. It is the IRS Automated Underreporter system flagging a mismatch between a return and what a third party reported, and proposing a change. This year the volume is up and the top new driver is Form 1099-DA: a broker reported a Bitcoin or Ethereum sale, the client left it off, and an automated system flagged it (IRS, 2026). The notice gives you 30 days. The work is what it has always been: reconcile the third-party figure, agree or dispute line by line, and write a clean response with exhibits.

That work is structured and repetitive, which is exactly what a well-built AI prompt is good at, and exactly where a lazy one will hurt you. The trap is asking Claude to 'respond to this IRS notice.' It will write a confident, plausible letter that may cite the wrong code section, misstate the deadline, or agree to an adjustment you should be disputing. The fix is not to avoid the tool. It is to give it a structured prompt that forces it to separate fact from recommendation and to show its work.

The prompt in this week's Claude Column does four things. It makes Claude restate the notice's proposed change and deadline in its own words, so you catch a misread immediately. It separates what the IRS is claiming from what you must verify in the client file from draft response language, so you never confuse a draft sentence for a settled position. It flags every figure that needs tie-out to a source document. And it ends with the questions you must answer before anything goes out the door.

Run that against a CP2000 and you get a first-pass response in about five minutes instead of forty: the proposed change summarized, a verification checklist, a draft letter in IRS-instruction format, and a list of exhibits to attach. What you do not get, and must never skip, is the judgment. You still pull the 1099-DA, confirm the basis the broker did or did not report, and decide whether to agree or dispute. The prompt drafts; you sign.

This matters more this year because the IRS put AI tax tools on its 2026 'dirty dozen' list, warning taxpayers not to rely on AI-generated answers to complex tax questions (IRS, 2026). That warning is aimed at taxpayers self-serving with chatbots, not at a CPA using AI as a drafting assistant under review. But the line it draws is the right one for us too: AI on the draft-and-review side, a licensed human on the signature. Use the prompt to kill the blank page, not to replace your read of the file.

Adoption nearly doubling in a year is the tell: AI tax research has crossed from early-adopter curiosity to standard practice. The edge is no longer using it. It is using it with a verification discipline your less careful competitors skip, the same discipline a CP2000 response demands.

What it is: An AI tax-research assistant built for practitioners. You ask a complex question in plain language and it returns a structured answer grounded in the IRC, Treasury Regulations, IRS guidance, and case law, with clickable citations you can verify.

What it does well: Triangulating a thorny question fast and showing its sources. For a CP2000 dispute or a client's odd 1099-DA basis question, it reaches the relevant authority in minutes instead of an hour in Checkpoint, and the citations let you confirm before you rely.

What it doesn't do well: Replacing your judgment or your malpractice exposure. It can misframe a question or miss a state wrinkle, and the IRS just flagged over-reliance on AI tax answers. Treat every output as a research memo to verify, not a conclusion to quote to a client.

Pricing: Gated by firm size and usage rather than one public price; competitors Blue J and Accordance price similarly. Get a current quote against your seat count and question volume, and trial it on questions you already know the answer to before you rely on it for ones you don't.

Here is the CP2000 prompt. Paste it into Claude, then paste the notice text and the relevant lines from the client's return beneath it. Redact the SSN; keep everything else.

Why it works: the RESTATE step catches a misread in the first ten seconds, before it poisons the draft. Splitting VERIFY from DRAFT stops you from mistaking a fluent sentence for a confirmed fact. The bracketed placeholders make the draft's gaps visible instead of letting Claude paper over them with a confident guess.

Where it fails: it will still occasionally cite a wrong or outdated code section, and it cannot see the client's actual basis unless you give it the document. It has no idea whether disputing is the right call. The prompt is a faster first draft and a built-in checklist, not a decision. You verify every figure, and you make every judgment.

Take the last CP2000 you handled, strip the SSN, and run it through the prompt above this week. Compare the AI's first-pass response to what you actually filed. Where did it save you time, and where would it have gotten you in trouble? You will end with a reusable prompt tuned to how your practice writes notice responses, and a clear-eyed sense of which steps you can never hand off.

P.S. Tax folks: reply with the IRS notice type you get most, and I'll build and share a tuned prompt for it in a future Prompt of the Week.