AI doesn't replace the CPA's judgment. It eliminates the steps that don't require it.

Here's the seven-step workflow.

Why Now: The IRS Is Running AI on Your Clients' Returns

Before the workflow, the context: the IRS finished the 2026 filing season with approximately 27% fewer staff than a year ago — down from roughly 102,000 to 74,000 employees. TIGTA's analysis of IRS challenges for 2026 flagged the gap explicitly: fewer humans, more algorithmic scrutiny.

Simultaneously, the IRS is using AI to cross-reference returns against bank interest records, side-gig income, and payment app transactions. Discrepancies that previously required a human reviewer to flag are now flagged automatically. More notices, fewer staff to resolve them.

For accounting firms, this is a structural demand signal. Clients who've handled their own IRS correspondence are about to find the process slower and more confusing than it's ever been. The firm that can resolve notices efficiently — and offer IRS representation as a service line — has a structural advantage that compounds as the agency continues to shrink.

The Seven-Step Workflow

This breakdown is based on CPA Pilot's documented notice response framework, adapted for small and mid-size accounting firms.

Step 1: Triage and Intake

What AI does: When a notice arrives (scanned or forwarded), AI classifies it automatically: CP2000 (income underreporting), CP14 (balance due), CP504 (urgent), Letter 12C (additional information needed), CP5071C (identity verification). It extracts the key data points — tax year, amount at issue, response deadline — and routes it to the appropriate workflow.

Why this matters: Misclassifying a notice type leads to the wrong response. Manual classification is fast but error-prone under volume. Automated triage eliminates the mistake and creates a timestamped intake record.

Step 2: Research

What AI does: Based on the notice type and the specific issue flagged, AI pulls the relevant IRS code sections, Revenue Procedures, and prior-year audit statistics. For a CP2000, it identifies what the IRS is claiming wasn't reported and what documentation is typically required to resolve it.

Why this matters: Most notices require the same research every time — the same code sections, the same procedural rules, the same documentation checklist. AI completes this in seconds. A human doing it manually takes 15-20 minutes per notice.

Step 3: Draft Response Letter

What AI does: Using the notice classification and research output, AI drafts a response letter from your firm's template library. The draft follows IRS correspondence standards: identifies the taxpayer, states the firm's position, references the applicable authority, and requests the specific relief sought.

The human step: The CPA reviews and edits. The AI draft is a starting point, not the final letter. For simple CP14 balance-due notices, the draft may need minimal editing. For complex CP2000 disputes, more substantive review is required.

Step 4: Exhibit Building

What AI does: AI assembles the supporting documentation package — W-2s, 1099s, bank statements, prior-year returns — based on the checklist generated in the research step. It organizes exhibits in the order IRS correspondence centers expect them.

Why this matters: Incomplete or disorganized exhibit packages are one of the most common causes of delayed notice resolution. An organized, complete package gets reviewed faster.

Step 5: Fact Confirmation

What AI does: Before the response goes out, AI cross-references the draft letter against the source documents to confirm consistency — amounts, dates, taxpayer identification numbers. This catches the transposition errors and copy-paste mistakes that can occur when working under volume pressure.

Why this matters: A letter with an inconsistent number goes back to square one. This step takes less than two minutes and prevents that outcome.

Step 6: Client Communication

What AI does: AI generates a plain-English client update: what the notice says, what your firm is doing about it, what the client needs to do (if anything), and what the expected timeline is. The language is adjusted for a non-tax professional audience.

Why this matters: Client anxiety during an IRS notice is real. A clear, prompt communication — drafted and sent the same day the notice is received — reduces client calls and demonstrates responsiveness. This step, done manually, takes 10-15 minutes per notice. AI drafts it in under a minute.

Step 7: Case Tracking

What AI does: The notice, response, deadline, exhibit package, and client communication are logged in your practice management system with automated follow-up reminders. If the IRS response deadline is 60 days out, the system flags a review at 30 days and again at 14 days.

Why this matters: Missed deadlines on IRS notices have consequences — additional penalties, loss of appeal rights, escalation to collections. Automated tracking eliminates the manual calendar entries and reduces missed deadlines to near zero.

What This Actually Saves

CPA Pilot's analysis of firms implementing this workflow shows:

• Average resolution time: 35 days → 14 days
• Time per notice: 45 minutes → approximately 15-20 minutes (for CPA review and sign-off on AI-generated work)
• Monthly time savings (30 notices): approximately 22.5 hours
• Annual capacity recovered (at $200/hr): $54,000

The time savings compound as you build a template library. The first five CP2000 responses require more CPA judgment to refine the templates. By notice 20, the templates are calibrated to your firm's voice and the IRS's preferences, and review time drops further.

How to Start

You do not need a dedicated notice response platform to implement this. Here's the minimum viable starting stack:

1. ChatGPT Business or Claude for Business — for research, drafting, and exhibit checklists. Build a custom prompt for each notice type (CP2000, CP14, CP504) using your existing response templates as the base.

2. Your practice management tool — for tracking and deadline management. If you're using Tax Dome, Karbon, or Canopy, the workflow integration is straightforward. If you're on a basic system, a shared spreadsheet works for tracking until volume justifies upgrading.

3. A template library — three to five letters per notice type, refined over 30 days of use. This is the institutional knowledge layer. Once built, it's the most valuable artifact in the workflow.

This week's action: Pull your last 10 IRS notices. Note how long each took to resolve and how much time your team spent on each step. That's your baseline. Then build one prompt for your most common notice type — CP2000 or CP14 — using your existing response letter as the template. Run your next notice through it and measure the difference.

Post-tax-season is the right time to build this. The notices are coming. The question is whether you're resolving them in 14 days or 35.