How new foundation models are transforming business automation
Building an effective prompt: key principles
A good business automation prompt follows a few simple but powerful rules:
- Clear objectives: Define exactly what action you want the AI to perform (e.g., classification, generation, retrieval).
- Relevant context: Provide background, industry terminology, and regulatory constraints.
- Structured constraints: Use labels, separators, and format instructions to guide the output.
- Feedback-driven iteration: Include steps for review and continuous improvement.
- Adapt to model capabilities: Leverage the specific model’s strengths (e.g., multimodality for GEN-1.5, mobility for ME-POIs).
Example 1: Employee onboarding
A prompt that converts a new hire into a personalized learning plan could look like this:
Role: HR AI Assistant
Goal: Create a personalized onboarding plan for a new employee.
Input:
- Name: [Name]
- Role: [Role]
- Reference departments: [List]
- Preferred learning tools: [Video, Text, Practice]
Instructions:
1. Generate a 5-day timeline with activities, links, and estimated time.
2. Include a short hands-on exercise using GEN-1.5 to simulate a real task.
3. Provide a resource checklist and a business context mind map.
Output: Structured JSON with sections: timeline, practice, checklist, mind map.This prompt leverages GEN-1.5’s ability to learn from short demos, giving the new hire a hands-on experience from day one.
Example 2: Data extraction from documents
To extract key fields from PDF contracts, try this prompt for a modern LLM:
Task: Structured data extraction from contract.
Instructions:
- Analyze the provided PDF.
- Identify: Party name, Expiration date, Amount, Penalty clauses.
- Return a JSON array with validated objects.
- If a section is missing, return null for that field and note the reason.
- Apply the ME-POIs framework to consider actual usage (e.g., place of agreement execution) as an additional signal.Including ME-POIs helps the model consider operational context, improving accuracy in detecting critical locations.
Example 3: Customer support with contextual awareness
When a customer requests a refund, a contextual prompt can automatically link the issue to the support system:
Scenario: Customer refund request.
Input:
- Ticket ID: [ID]
- Customer message: [Text]
- Interaction history: [Last 5 messages]
Goal: Determine eligibility and propose next action.
Instructions:
1. Evaluate compliance with refund policy (based on ME-POIs for place of purchase).
2. If eligible, generate an automatic approval with a reference code.
3. If not eligible, propose an alternative plan (e.g., product exchange, warranty extension).
4. Record the outcome in the CRM with justification.This flow reduces resolution time and maintains consistent support tone across channels.
Integrating the latest innovations: GEN-1.5, ME-POIs, XPENG IRON
The latest models offer unique capabilities that can be leveraged directly through prompts:
- GEN-1.5
- ME-POIsadds mobility information to places. Ideal for space optimization, resource planning, and product offering localization.
- XPENG IRONbrings physical AI into business processes (e.g., warehouse, quality inspection). Prompts can send commands to humanoid robots for repetitive tasks.
Prompts that leverage GEN-1.5 for physical tasks
Create a step-by-step tutorial that teaches an XPENG IRON robot to place a package on a shelf:
Goal: Generate a physical robot demo using GEN-1.5.
Input:
- Robot type: XPENG IRON
- Task: Place a package on a shelf
- Demo duration: 8 seconds
Instructions:
1. Write a short movement description: "Grasp the package, walk to the shelf, place the package on shelf A, label the item."
2. Provide an angled view of the robot with reference points.
3. Include a safety check: verify height, weight, and stability.This prompt transforms a short observation into a concrete action plan for robotic hardware.
Prompts that incorporate ME-POIs for space optimization
Map the best office layout based on actual usage:
Task: Optimize office layout using ME-POIs.
Input:
- Current floor plan: [SVG]
- Movement data: [Entry/exit trajectories]
Goal: Propose a new layout that improves flow.
Instructions:
1. Calculate high-traffic routes using movement data.
2. Identify hotspots (places with intensive usage).
3. Suggest relocation of desks, meeting rooms, and break areas.
4. Visualize the result as a 3D floor plan with labels.This data-driven approach reduces congestion and increases employee well-being.
Error handling and continuous improvement
Even the most accurate prompts can fail. Implement a feedback loop:
- Capture failed outputsin a structured log.
- Identify error patterns(format, content, context).
- Update the promptwith corrective examples or additional constraints.
- Test iterativelyon a subset of use cases before full deployment.