Workflows & Reasoning Pipelines
Single prompts aren't enough for complex tasks. Learn how to design tool-augmented flows, multistep reasoning pipelines, and orchestrated workflows that chain AI capabilities together.
Beyond single prompts
A single prompt to an AI model is like asking one question. A workflow is like giving the AI a project plan with multiple steps.
Real-world tasks often need multiple steps: search for information, analyse it, make a decision, take an action, verify the result. Workflows chain these steps together — each step can use a different tool, model, or data source.
Three workflow patterns
| Feature | Simple Chain | Tool-Augmented Flow | Multistep Reasoning |
|---|---|---|---|
| Pattern | Step A output becomes Step B input | Model calls external tools mid-generation | Model breaks complex problem into sub-tasks |
| Control | Deterministic — you define each step | Model-driven — model decides when to call tools | Model-driven — model plans its own steps |
| Example | Translate text, then summarise | Search database, then generate report | Analyse contract, identify risks, suggest changes |
| Complexity | Low | Medium | High |
| Best for | Predictable, sequential tasks | Tasks needing real-time data | Complex analysis requiring reasoning |
Tool-augmented flows
In a tool-augmented flow, the model can call tools (functions) during generation:
| Step | What Happens |
|---|---|
| 1. User sends request | ”What’s the delivery status for order #12345?“ |
| 2. Model reasons | Determines it needs to look up the order |
| 3. Tool call | Model requests: call get_order_status(order_id="12345") |
| 4. Your code executes | Your function queries the database, returns status |
| 5. Model receives result | Gets the status data back |
| 6. Model generates | Produces a natural-language response with the order status |
Common tools in workflows
| Tool Type | What It Does | Example |
|---|---|---|
| Code interpreter | Runs Python code for data analysis | Calculate statistics from CSV data |
| Web search | Retrieves current information from the web | Get latest shipping rates |
| Custom function | Calls your own APIs | Check inventory, create ticket, send email |
| Knowledge retrieval | Searches indexed documents | Look up company policies |
Multistep reasoning pipelines
For complex problems, models can plan and execute multiple reasoning steps:
| Component | Role | Example |
|---|---|---|
| Planner | Breaks the task into sub-tasks | ”To analyse this contract: 1) Extract parties 2) Identify obligations 3) Flag risks” |
| Executor | Carries out each sub-task | Extracts party names from Section 1 |
| Evaluator | Checks quality of each step’s output | Verifies extracted names match document |
| Aggregator | Combines sub-task results into final output | Produces complete contract analysis report |
Real-world example: Atlas Financial's loan review pipeline
Atlas Financial automates loan application review with a multistep pipeline:
- Extract — Content Understanding pulls applicant data from the PDF application
- Retrieve — Azure AI Search finds relevant lending regulations
- Analyse — GPT-4o compares application against regulations
- Score — Model assigns a compliance score with justification
- Route — Score above threshold: auto-approve. Below: flag for human review
Each step produces structured output that feeds the next. The entire pipeline runs in under 30 seconds per application.
Exam tip: Workflow vs agent
The exam distinguishes between workflows and agents:
- Workflow = pre-defined steps, deterministic control flow, you decide the sequence
- Agent = model decides the sequence, plans dynamically, may retry or adapt
If the scenario needs predictable, repeatable execution → workflow. If it needs adaptive, flexible reasoning → agent.
Key terms
Knowledge check
MediaForge needs to process client briefs that arrive as PDFs. The workflow is: extract text from PDF, identify the target audience, generate 5 headline options, then translate headlines into 3 languages. What type of workflow is this?
A user asks Kai's logistics chatbot: 'How much would it cost to ship 500 packages from Auckland to Sydney next week?' The bot needs to check inventory, look up shipping rates, calculate volume discounts, and factor in the calendar for next week's availability. What's the best pattern?