The 13 pipeline step types that power every Qbiqal AI employee
In the early days of building AI applications, developers relied on a single design pattern: take user input, wrap it in a prompt template, send it to an LLM, and display the response. While this works well for simple chatbots, it fails catastrophically when applied to complex, multi-stage business workflows. A single model call cannot reliably search database entries, compose emails, handle API exceptions, and ensure policy compliance in one go. To solve this, QBIQAL developed a pipeline execution system. Every task performed by our AI employees is broken down into a structured sequence of discrete pipeline steps. In this technical deep-dive, we analyze the 13 pipeline step types that make our AI workforce reliable, scalable, and safe.
Why Workflows Need Pipelines
When a QBIQAL AI Employee like "Aagam" is assigned to research a business, he doesn't just ask an LLM to "find info." Instead, the Cortex engine compiles a workflow graph composed of separate nodes. Each node represents a specific functional primitive.
This modular execution model guarantees predictability. If a step fails (e.g., a third-party website is down), the pipeline doesn't crash; it handles the error, retries the step, or routes the task to a human for intervention.
Let's break down the core pipeline steps that build our operational foundation.
The 13 Pipeline Step Definitions
Each step has a specific role, input schema, and validation criteria. Here are the key steps configured in the QBIQAL workflow builder:
1. llm_call: Generates text, code, or structured reasoning. Backed by our multi-model routing protocol (switching between Claude, GPT, and custom local models).
2. human_review: A mandatory verification gate. The pipeline pauses, notifies a manager via the Command Center, and waits for a physical click before resuming.
3. api_execution: Outbound call to external systems (e.g., Salesforce CRM, Zenoti, Gmail, Stripe) with secure credentials.
4. vector_search: Semantic search across the business's private vector databases to retrieve contextual documents or customer logs.
5. conditional_branch: Evaluation logic that routes the flow down path A or path B based on specific criteria (e.g., if ticket_value > ₹50,000).
6. data_extraction: Converts messy, unstructured HTML, PDF, or image files into clean, structured JSON schemas.
7. loop_iterator: Iterates over lists (e.g., looping through 50 leads found during outreach).
8. db_query: Read/write records from the tenant's isolated SQL database.
9. code_sandbox: Runs custom Python scripts in a secure, isolated container to verify calculations or generate charts.
10. message_send: Dispatches transactional alerts via WhatsApp, SMS, or Slack.
11. wait_delay: Pauses execution for a specified duration (e.g., waiting 24 hours before sending an outreach follow-up).
12. agent_handshake: Passes execution from one specialized AI employee to another (e.g., Aagam passes verified lead data to Karan for copywriting).
13. audit_log: Commits the execution state, token cost, and latency metrics to the immutable compliance ledger.
A Workflow Pipeline Configuration Example
These steps are combined into simple, readable YAML configurations. Here is a snippet of a lead qualification workflow pipeline:
pipeline:
id: lead_qualification_flow
steps:
- id: extract_profile
type: data_extraction
source: input.website_url
- id: evaluate_fit
type: llm_call
prompt_template: templates.eval_fit
input: steps.extract_profile.output
- id: check_risk
type: conditional_branch
condition: steps.evaluate_fit.output.risk_score > 7
routes:
true: steps.flag_for_review
false: steps.auto_approve_outreachConcluding Thoughts
By dividing complex business activities into structured pipeline steps, QBIQAL provides enterprises with a clear, auditable trail of AI decisions. Business owners can inspect exactly which step occurred, how much it cost, and who verified the output, ensuring compliance at every stage.
Deploy AI Workforces in Your Operations
QBIQAL provides robust, compliant, and ready-to-work AI specialists that interface with your booking, sales, and analytics systems seamlessly.