Introduction: Why autonomous AI workflows are crucial for SMBs today
To create autonomous AI workflows for your small or medium-sized business, you need to combine a suitable large language model, effective prompt engineering, and local inference tools like Lily, while also considering explainable AI for trust. In a market where every minute counts, a workflow that learns, decides, and acts without human intervention can transform operations, sales, and customer support.
Core components of an autonomous workflow
An autonomous workflow rests on three pillars:
- Selected LLM:Models such as Qwen3.6-35B-A3B strike a good balance between performance and size, making them ideal for local inference.
- Prompt engineering:Well-structured prompts steer the model toward specific actions, reducing generic responses.
- Execution tools:Engines like Lily (built with Rust and Metal for Apple Silicon) or cloud APIs such as OneRail’s delivery-optimization service.
How to choose the right LLM tool for your business
When evaluating a model for your SMB, keep the following factors in mind:
- Model size versus available hardware resources.
- Explainability capabilities (similar to the Motional-MIT system used in self-driving AI).
- Integration with existing stacks (CRM, ERP, analytics tools).
For many SMBs, an open-source model running on Lily on a Mac/PC with Apple Silicon chips offers the best trade-off between cost, privacy, and performance.
Practical example: Building a sales assistant with a local LLM
Suppose you want to automate the creation of sales proposals. Here’s a step-by-step autonomous workflow:
- Data acquisition:CRM data is exported to a JSON file.
- Processing with the LLM:A prompt asks the model to synthesize a customized proposal.
- Action generation:The model writes the proposal and saves it to the ERP system.
- Monitoring:A dashboard reports conversions and feedback.
Here’s a prompt you can copy and adapt:
You are an expert sales assistant. Given the following customer information: {customer_data}, generate a concise sales proposal (max 150 words) highlighting the value proposition, price, and next steps.Paste this prompt into an interface that supports the{customer_data}variable, and the LLM will produce a proposal ready to send.
Integrating explainable AI and monitoring
- Attention visualizations from models for key touchpoints.
- Real-time alerts when an action exceeds defined risk thresholds.
Tools like MLflow or Lily’s native dashboards can track decision outcomes and generate reports that are understandable even for non-technical staff.
Prompt engineering best practices for autonomous workflows
Effective prompts for workflows are structured, concrete, and action-oriented:
- Define the role:
- Specify the output format:
- Include constraints:
- Add context:Reference relevant business data within the prompt.
Use placeholders (e.g.,{var_name}) to enable prompt reuse across multiple situations.
Current trends: from Lily to Nvidia, what's changing for SMBs
Recent developments are making autonomous workflows more accessible:
- Perplexity’s Lily:A Rust + Metal inference engine that runs on Apple Silicon, allowing SMBs to execute large models locally without cloud costs.
- OneRail’s delivery platform:Leverages Nvidia AI to optimize the final steps, demonstrating how AI can be woven into existing logistics processes.
- Motional and MIT’s explainability:Provides transparent reasoning for critical decisions, a valuable asset for regulated industries.
These trends lower the barrier to entry: you no longer need a high-powered cluster to launch an autonomous workflow.
Concrete steps to get started: an implementation checklist
Use this checklist to launch your first autonomous workflow:
- [ ] Identify a repetitive process (e.g., proposal generation, customer email monitoring).
- [ ] Choose a suitable LLM and install Lily or an equivalent local engine.
- [ ] Write a base prompt using the schema above.
- [ ] Connect the LLM to your business data (API, CSV files, webhooks).
- [ ] Add an explanation module for high-risk decisions.
- [ ] Set up monitoring and alerts.
- [ ] Test the end-to-end cycle with real data.
- [ ] Optimize the prompt and performance metrics.
By following these steps, you can have a functional autonomous workflow in just a few weeks.