How to Build Business Automation Prompts in 2026
Identify the Right Use Cases for Automation
Not all processes benefit from the same prompting technique. Focus on tasks that meet three key criteria:
- High repetitiveness: Tasks performed multiple times with structured data.
- Clear input and output points: Incoming data is consistent, and the output has a defined format.
- High impact on productivity: Automation saves time or reduces errors.
Classic examples include extracting data from invoices, generating sales reports, triaging support tickets, and composing follow-up emails.
Leveraging New Multimodal Models
The AI landscape in 2026 offers tools that go beyond text. Incorporate multimodal capabilities into your prompts whenever possible:
- Tencent’s WeVisOne-VL: Unifies detection, OCR, segmentation, and matting in a single 2B/4B parameter model. Ideal for extracting text and structures from images of invoices, receipts, or scanned forms.
- Soniox STT v6: Adds text input for voice agents and echo filtering. Use this model when you need to transcribe calls and then extract structured data from transcripts.
- OpenAI Pando: A Git-compatible backend that reduces push latency. Integrate Pando into your CI/CD workflows to ensure the latest prompts and datasets are instantly available.
When designing a prompt, explicitly consider these capabilities. For example: “Analyze the image of this invoice using WeVisOne-VL’s OCR to detect text and table borders. Extract the following fields: invoice number, due date, total amount, and vendor name. Return a single JSON object with these fields. If a field is missing, set its value to null.”
Practical Example: Extracting Data from Invoices
Goal: Extract amounts, invoice numbers, due dates, and vendors from scanned invoice images.
Prompt (with multimodal capabilities):
Analyze the provided image using WeVisOne-VL to detect text and table borders.
Extract the following fields:
- Invoice number
- Due date
- Total amount
- Vendor name
Return a single JSON object with these fields. If a field is missing, set its value to null.Result: The model returns a parseable JSON that can be directly inserted into an ERP system or accounting database.
Practical Example: Automated Sales Report Generation
Goal: Receive a daily sales summary in PDF format, including charts, based on raw data stored in a Google Sheets spreadsheet.
Prompt (leveraging data and Pando):
Load the latest spreadsheet from Google Sheets (use Pando for faster data retrieval).
Filter data for date = today.
Calculate:
- Total revenue
- Quantity of best-selling products
- Growth rate compared to weekly average
Generate a Markdown report with headers, a table, and a brief analysis block.
Convert Markdown to PDF and attach it to the sales managers' email.Result: Every morning, an email with an updated, ready-to-read sales report is sent with a single click.
Practical Example: Email and Support Ticket Management
Goal: Classify customer emails as “billing,” “technical,” or “general” and automatically create a ticket in the support system.
Prompt (using Soniox STT v6 for voice input and classification):
Receive the email message text:
"[email body]"
Analyze the content and assign one of the following labels:
- Billing
- Technical
- General
Return a JSON with a "label" field and optional "sub-category."
If the email contains a complaint, add the "urgency" field: "high," otherwise "normal."Result: The system creates a ticket with the appropriate label, streamlining the support workflow.
Tips for Writing Effective Prompts for Business Automation
- Be specific, not vague: Instead of “analyze the data,” specify “extract sales metrics by region and quarter.”
- Define the output format: JSON, Markdown, or CSV makes results machine-parseable.
- Include constraints: “If a field is missing, set to null” prevents incomplete responses.
- Use contextual cues: “Assume the data is in European layout (date format DD/MM/YYYY).”
- Test with edge cases: Send an empty PDF, a noisy call, or a corrupted spreadsheet to see how the model performs.
Key Takeaways and Next Steps
Key points:
- Identify high-volume, low-value processes that can be managed by AI.
- Leverage multimodal models (WeVisOne-VL, Soniox STT v6, Pando) to create richer, more functional prompts.
- Write prompts with clear instructions on format, constraints, and context.
- Implement a feedback loop: monitor prompt results and refine prompts based on errors.
Next steps: Start with a proof-of-concept that extracts data from a single invoice, then scale the process enterprise-wide. Track metrics (time saved, error reduction) to quantify the ROI of prompt-based automation.
By creating prompts that leverage the latest AI capabilities, businesses can transform manual work into automated, future-ready processes that increase productivity and reduce errors.