Introduction: Why SMEs Need Autonomous AI Workflows Today
Which tool should you use to get started? GLiNER2.5-Decide
Fastino Labs has released GLiNER2.5-Decide, a 340 million parameter decision-making model that runs on standard CPUs. It’s ideal for extracting data structures and making rule-based decisions without sending data to cloud services.
Practical example: structured lead extraction
Imagine your SME wants to transform website contacts into organized CRM records. With GLiNER2.5-Decide, you can define a schema (Name, Email, Phone, State) and have the model return validated JSON directly on your server.
import torch
from gliner import GLiNER2_5_Decide
schema = {
"type": "object",
"properties": {
"name": {"type": "string"},
"email": {"type": "string"},
"phone": {"type": "string"},
"state": {"type": "string"}
},
"required": ["name", "email"]
}
model = GLiNER2_5_Decide.from_pretrained("GLiNER2.5-Decide")
text = "Marco Rossi, marco@rossi.it, +393331234567, Lombardia"
result = model.extract(text, schema)
print(result)The result is a JSON ready to be inserted into your CRM database without further cleaning.
Why integrate security? The role of Altar-1
Secure workflow with Altar-1
- Step 1: Run the raw data through Altar-1 for a risk assessment.
- Step 2: If the risk is low, send the data to GLiNER2.5-Decide for extraction.
- Step 3: After extraction, use FLUX 3 Action to decide the next action (email, call, archive).
This three-step approach ensures that only secure content is processed, minimizing data breaches.
Adding action capabilities: FLUX 3 Action for automation
Black Forest Labs has introduced FLUX 3 Action, a 7 B open-weight action model capable of interpreting intentions and generating commands for robots or software systems. For SMEs, this means being able to automate repetitive tasks such as sending follow-up emails or updating records.
Example prompt for FLUX 3 Action
prompt = """
Given the following extracted information: {"name": "Luca Bianchi", "email": "luca@bianchi.it", "state": "Expired"}
Generate a follow-up email to thank the customer and propose a renewal.
Use a courteous tone and include a 'Renew now' button.
"""
response = flux3_action.generate(prompt)
print(response)The response can be sent directly to an email service, completing the automation cycle.
Complete end-to-end workflow: customer management for an SME
We combine the three models into an autonomous pipeline:
- Ingestion: Contacts are saved in an S3 bucket.
- Security check: Altar-1 analyzes each file for malware or sensitive data.
- Extraction: GLiNER2.5-Decide extracts name, email, phone, and state.
- Action: FLUX 3 Action generates a follow-up email or task note.
- Storage: The result is stored in the CRM and a webhook notifies the team.
The entire flow can be orchestrated with a Python script or a low-code tool like Prefect, running 24/7 without human intervention.
Implementation tips and takeaways
- Start with a single use case.Use GLiNER2.5-Decide for a small dataset before expanding its use.
- Integrate security from the start.Run Altar-1 on every input to prevent risks.
- Use open-weight models.They save API costs and give you full control.
- Document prompts.Models like FLUX 3 Action depend on well-defined prompts.
- Monitor performance.Track latency and accuracy; adjust the schema or security filters accordingly.