Drake Tax + AI Document Parsing: A Working Pipeline

Drake Tax is the most common tax-prep software at small and mid-sized CPA practices. It also doesn't have a real API. The integration pattern most firms want — "auto-populate the return from these PDFs" — is not officially supported.

Here's the workable pattern.

The two paths

Path A: CSV import. Drake supports CSV import for many schedules. You produce a clean CSV; Drake ingests it. This is the path that works without a real API.

Path B: Drake's web service (limited). Drake's API offers some functionality but it's read-mostly and not exposed for full return manipulation. Useful for pulling prior-year data, less useful for writing entries.

We built around Path A. Here's the flow.

The pipeline

Step 1: PDF intake. Client uploads tax docs to a secure portal (we used the firm's existing Canopy portal; SharePoint or Box also works). Webhook fires when a new doc arrives.

Step 2: OCR. Each PDF goes through OCR. We used AWS Textract for forms because it understands tax-form structure (named fields, key-value pairs). For documents Textract doesn't natively understand, we fall back to Google Document AI's Form Parser.

Cost note: Textract is roughly $1.50 per 1000 pages. Document AI's Form Parser is about $30 per 1000 pages. Use the cheap one when it works.

Step 3: AI classification. Claude classifies each document:

``` What kind of tax document is this? Options: W-2, 1099-INT, 1099-DIV, 1099-MISC, 1099-NEC, 1099-R, 1099-B, 1098, K-1 (1065), K-1 (1120-S), K-1 (1041), Schedule K-1, SSA-1099, 1098-T, 1098-E, other.

Confidence 0-1. If confidence <0.85, return "needs_human_review". ```

Step 4: Field extraction. For each document type, a tailored extraction prompt. Example for W-2:

``` Extract from this W-2: - Employer name and EIN - Box 1 (wages) - Box 2 (federal tax withheld) - Box 3 (Social Security wages) - Box 4 (Social Security tax) - Box 5 (Medicare wages) - Box 6 (Medicare tax) - Box 12 codes and amounts (each separately) - Box 14 entries - State wages and tax (each state separately)

For each field, return value and confidence. ```

Step 5: Variance check. Compare against prior year for the same client. If a number is materially different (>20% delta on the same payor), flag for human review.

Step 6: CSV generation. Map the extracted fields to Drake's CSV import format. Each return type (1040, 1120S, 1065) has different CSV templates. Build the right one.

Step 7: Human import. The CPA reviews the generated CSV (we ship it as an attachment in their daily digest), then imports it to Drake manually. Drake's import UI is the human checkpoint.

What the numbers look like

Per W-2: extraction takes 3-5 seconds. Variance check is instant. Total processing per document: under 8 seconds.

Per return with average 6 source documents: about a minute of automated processing replacing 35-45 minutes of manual entry.

Accuracy on clean PDF W-2s: 96-98% per field. On phone-camera scans: 75-85% per field (we route these to manual entry).

What broke

Box 12 codes are tricky. Different employers use different codes. Some 401(k) deferrals show as code D, some show as code DD, some show as both. Claude got confused on edge cases. We added a normalization step that maps known code variations.

State withholding for multi-state clients. When a W-2 has multiple state entries, the layout varies wildly. Some employers stack them, some put them side-by-side. We added a "multi-state W-2" detection prompt that flags these for the CPA to verify the parsing.

K-1s are not solved. K-1s have so much variation in how partnerships and S-corps report items that auto-extraction is unreliable. We extract the header info (entity, EIN, percentages) and route the rest to manual entry.

What's not covered

Schedule C income. Client provides a P&L summary; we don't try to parse it from raw bank records.

Rental property. Schedule E requires too much judgment about what to depreciate, capital vs. expense, etc. Manual.

Cryptocurrency transactions. This is a separate beast. Use a dedicated tool (CoinTracker, Koinly) that exports to Drake-compatible CSVs.

What this isn't

Not a replacement for the CPA. Every CSV is reviewed before import. Every variance flag gets human attention.

Not a fully automated tax return. The CPA still does the analysis, the planning, and the strategy. We're just removing data entry from their day.

What to build first

If you're a tax practice considering this:

One, the OCR layer on the top 3 document types (W-2, 1099-INT, 1099-NEC). These cover 70-80% of intake volume for most practices.

Two, the CSV generation for 1040 specifically. Business returns are more complex.

Three, the variance check. Catches errors in both the AI output and the CPA's prior year.

Total build: 3-4 weeks of dev work. Annual time savings: 80-120 hours per preparer at average practice volumes.