Introduction: Automate business processes with just a few clicks
Why autonomous AI workflows are crucial for SMEs today
An autonomous workflow delivers immediate benefits:
- Reduces repetitive tasks by up to 80%.
- Controls costs with models optimized for edge devices.
- Enables instant scalability through containers and Kubernetes.
- Delivers real-time, data-driven decision-making.
- Ensures continuous adaptation through automated feedback loops.
Key AI tools available in 2026
Google Cloud Gemini Agent: The universal enterprise agent
Google Cloud has launched Gemini Agent, a cloud-hosted agent that unifies bot creation, data extraction, and conversational interfaces. It’s ideal for replacing custom scripts with a single, managed service.
from google.cloud import aiplatform
from google.ai.generativelanguage import GenerateContentRequest
aiplatform.init(project="your-project", location="europe-west1")
agent = aiplatform.gemini.Agent(
model="gemini-agent",
description="Analyze invoices and approve payments"
)
response = agent.query("Upload the invoice PDF and extract amount, date, and supplier.")
print(response)This code snippet initializes the agent, connects it to a specific model, and queries it using natural language. The result can be sent directly to an accounting system via a webhook.
JetBrains Mellum2.1: A 12B MoE for code agents
JetBrains’ Mellum2.1 is a Mixture-of-Experts model optimized for code completion and test generation. With its 2.5B active parameters, it delivers high performance while consuming minimal resources.
from jetbrains.mellum import MellumClient
client = MellumClient(api_key="YOUR_KEY")
# Generate a function that validates an email address
code = client.generate_code(prompt="Create a Python function that validates an email address and returns True/False")
print(code)Insert the generated code into a Git repository and enable pre-commit checks for automatic feedback.
Saluki 27B: A compressed model for edge inference
// Example of inference with llama.cpp
gguf_model model("saluki-27b.gguf");
auto output = model.generate("Generate a short quote for a customer: ");
std::cout << output << std::endl;You can run the same model on a Raspberry Pi or a serverless service, keeping operational costs low.
How to build an autonomous workflow step by step
- Define the business objective.Identify the process you want to automate (e.g., invoicing, customer support).
- Choose the right AI model.Use Gemini Agent for data extraction, Mellum2.1 for code generation, and Saluki 27B for edge inference.
- Design effective prompts.Follow the format: *Goal* + *Context* + *Style* + *Constraints* + *Example*. Ready-made templates are available in Google Cloud’s open-source repositories.
- Integrate with existing systems.Expose REST endpoints or use webhooks to connect to ERP, CRM, or data storage platforms.
- Implement monitoring and feedback.Record metrics (latency, cost, accuracy) and feed the data back into the model for continuous improvement.
- Scale with containers.Dockerize each component and deploy it on Kubernetes; use Knative for serverless functions with automatic scaling.
Complete example: Automated invoicing workflow
Imagine you want to convert an invoice PDF into an approved accounting record. The process is as follows:
- Ingestion:A webhook from an accounting system receives a PDF uploaded by the user.
- Parsing:Gemini Agent reads the document and extracts amount, date, supplier code, and invoice number.
- Validation:A Python microservice (using Mellum2.1 for business logic) verifies that the amount is positive and that the supplier exists in the catalog.
- Approval:The system sends a notification to the manager via Google Chat; Gemini Agent handles the "approve/reject" response.
- Storage:The validated data is written to the company database via a REST API.
Here’s an example FastAPI that combines these steps:
from fastapi import FastAPI, UploadFile, File
from google.cloud import aiplatform
import json
app = FastAPI()
aiplatform.init(project="your-project", location="europe-west1")
@app.post("/invoice-process")
async def process_invoice(file: UploadFile = File(...)):
# 1. Read the PDF
pdf_content = await file.read()
# 2. Extract fields with Gemini Agent
agent = aiplatform.gemini.Agent(model="gemini-agent")
extraction = agent.query(f"Analyze this PDF and return JSON with amount, date, supplier, invoice_number: {pdf_content}")
data = json.loads(extraction)
# 3. Validation (simplified example)
if data["amount"]This API can be deployed with Docker and connected to a messaging service for the approval workflow.
Best practices and security
- Use limited API keys and rotate credentials regularly.
- Encrypt sensitive data both in transit (TLS) and at rest (AES-256).
- Version prompts and maintain a change log for audit purposes.
- Monitor latencies and costs via Cloud Monitoring; set alerts for anomalous spikes.
- Perform bias and privacy checks on models before deploying them to production.
Actionable takeaways
- Start a proof-of-concept with Gemini Agent to extract data from common documents.
- Integrate Mellum2.1 into your CI/CD pipelines to automatically generate unit tests.
- Deploy Saluki 27B on an edge device to reduce latency in customer interactions.
- Document each workflow with Mermaid diagrams for future maintenance.
- Establish a quarterly feedback loop to measure accuracy and costs, and adjust models accordingly.