An AI-Assisted CP2000 Notice Response Workflow

CP2000 is the automated under-reporter notice. The IRS thinks the income on your client's return doesn't match what payors reported. Resolving one cleanly takes hours of comparison work, and most of that work is mechanical.

This workflow hands the mechanical part to software and keeps the judgment with the CPA.

What CP2000 is, in one paragraph

The IRS got 1099s and W-2s reporting income they think your client didn't include on the return. Or your client agreed in writing to something the IRS thinks is taxable that the return didn't reflect. The notice proposes a tax assessment. Your job: agree, disagree with explanation, or partially disagree.

The work is forensic. You compare what the IRS claims against the return, the source documents, and the client's actual situation. The agreement/disagreement letter is mostly templated.

The workflow

Step 1: OCR the notice. Most CP2000s arrive as PDFs (client scans them or the practice receives a copy). OCR with a quality engine (AWS Textract and Google Document AI both work). Extract: tax year, proposed assessment amount, line items with reported amounts vs. return amounts.

Step 2: Claude classifies the line items. A prompt that takes the OCR'd notice text and returns structured JSON:

For each line item the IRS is flagging:
- payor_name
- payor_tax_id (if visible)
- form_type (1099-MISC, 1099-INT, W-2, etc.)
- amount_reported_to_irs
- amount_on_return (if you can determine from cross-reference)
- variance
- confidence

Step 3: Cross-reference against the return. Pull the prior-year return data from Drake Tax. For each flagged line item, find the corresponding entry on the return (if any). The system flags three buckets:

• Match found — the return reported this income, IRS may have a different categorization
• Missing — the return doesn't show this income
• Partial — the return shows a related amount but it's different

Step 4: Draft the response letter. Different template per bucket. The system generates a draft letter referencing each line item with the specific argument:

• For matches: "The income reported on Form 1099-MISC from [Payor] is reflected on Schedule C, Line X of the 2023 return as part of [aggregated amount]. The reporting categorization differs but the amount is included."
• For missing: a request for the source document and a placeholder line
• For partial: an explanation of the reconciliation

Step 5: Human review and signature. The CPA reviews the draft, verifies every line item, signs.

How to measure it

Time your current process on the next few notices before you build anything, with research, drafting and review timed separately. Then time the same steps once the workflow runs. The comparison tells you whether the build paid off, and which step to improve next.

What tends to break

OCR confidence on scanned-from-phone notices. Phone camera scans break OCR worse than scanner PDFs. Add a "rescan required" flag that asks the client to use the practice's secure portal upload, which produces a clean PDF.

Hallucinated payor matches. A cross-reference can match a 1099 to an unrelated Schedule C line. Set a minimum confidence, and send anything below it to a human.

The "client doesn't have the document" path. Some CP2000s reference 1099s the client never received. Have the system generate a Form 4506-T request for the IRS transcript along with the response letter.

What this isn't

This isn't replacing the CPA's judgment. Every response is reviewed and signed by a human. The system is doing the mechanical work (data extraction, cross-referencing, drafting) so the CPA can spend time on the analysis.

It also isn't meant for other notice types. CP2000 comes from the IRS's automated under-reporter program, which is why the response is templatable. Balance-due and levy notices such as CP504 and CP90, and real examinations, are different work and need different handling.

What to build first

If you're a tax practice considering this:

One, get the OCR pipeline working on 10 sample notices. Measure accuracy. Fix it before adding Claude.

Two, build the cross-reference. This is where the real time savings come from. The drafting is window dressing on a working cross-reference.

Three, build the drafting last. By the time you have the cross-reference, the drafts almost write themselves with a 50-line template.

Scope it as a small project for a competent developer, and judge the payback against the times you recorded before the build.