How do you automate data entry with AI?

Data entry means reading something and typing part of it somewhere else. That is what language models do well when the source is messy text and the destination is a structured table.

Good candidates

• Invoices and receipts into an expense sheet.
• Customer emails into a CRM-ready list (name, company, request, urgency).
• Application or intake forms into a tracker.
• Business cards or contact details into a contact list.
• Meeting notes into a task list.

Poor candidates: handwritten scans of poor quality, or anything where a single wrong digit has serious consequences and cannot be verified.

Step 1: Define the schema

List each field, its format and what to do if it is missing:

Fields: invoice_number (text), vendor (text), invoice_date (YYYY-MM-DD), total (number, no currency symbol), currency (3 letters), due_date (YYYY-MM-DD or blank). If a field is not clearly present, leave it blank; do not guess.

The “do not guess” line is the most important sentence. Anthropic guidance on reducing hallucinations in Claude says to explicitly give Claude permission to admit uncertainty. A 2025 paper argues that language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty.

Step 2: Test on a sample

Take 20 documents you have already entered by hand. Run the extraction. Compare field by field. Note error types: dates swapped, totals including tax, vendor names varying.

Step 3: Refine the instructions

Add rules for each error: “Total means the amount due including tax.” “Dates are U.S. month/day/year.” Re-test on the same sample, then a new one.

Step 4: Build the pipeline

For a batch, an agent tool like Claude Code can loop through a folder, extract fields from each file into one spreadsheet, and add a “needs review” column for uncertain entries. See what AI agents are and Claude Code.

Step 5: Validate automatically

Add checks that do not rely on the AI:

• Totals recalculated from line items.
• Date within expected range.
• Vendor exists in your vendor list.
• No duplicate invoice numbers.

Structured output is not the same as correct output

If you use an API rather than a chat window, both major vendors offer structured output. OpenAI says Structured Outputs ensures the model will always generate responses that adhere to your supplied JSON Schema. Anthropic says that without structured outputs, Claude can generate malformed JSON responses or invalid tool inputs that break your applications. Both guarantees are about the format. A field can match the schema and still hold the wrong value, which is why Step 5’s checks matter. Anthropic guidance for Claude says critical information must always be validated, especially for high-stakes decisions.

Step 6: Keep a human in the loop

Review flagged rows fully, plus a random sample of the rest. Keep the original file linked to each row so anyone can verify.

Risks

• Silent errors: a plausible but wrong value looks fine.
• Privacy: documents often contain personal or financial data; confirm the tool is approved. See whether it is safe to put company data in ChatGPT.
• Drift: when vendors change their invoice layout, accuracy can drop without notice. Re-test periodically.

Spreadsheet side

Once you have a table, formulas and cleanup steps in Excel or Google Sheets finish the job. In Learn AI Fast, attendees practice cleaning and analyzing a provided spreadsheet with AI and verify the results.

Sources

• Anthropic, Reduce hallucinations (Claude Platform Docs): “Explicitly give Claude permission to admit uncertainty”
• Kalai, Nachum, Vempala and Zhang, Why Language Models Hallucinate (arXiv 2509.04664): “sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty”
• OpenAI, Structured Outputs guide: “ensures the model will always generate responses that adhere to your supplied JSON Schema”
• Anthropic, Structured outputs (Claude Platform Docs): “Without structured outputs, Claude can generate malformed JSON responses or invalid tool inputs that break your applications”
• Anthropic, Reduce hallucinations (Claude Platform Docs): “Always validate critical information, especially for high-stakes decisions”