What is an AI Workflow?
An AI workflow is a repeatable process made from connected steps.
Each step has a clear purpose. One step may receive information, another may ask an AI model to analyse it, and another may return or save the result.
In Feluda, you build workflows visually in Studio by placing blocks on a canvas and connecting them in the order they should run.
A simple workflow might look like:
Input → Summarise with AI → Output
A more detailed workflow might classify information, follow different paths, use a tool, and prepare several results.
Why use an AI workflow?
A normal AI conversation is useful for a one-time task.
A workflow is useful when the same process needs to be repeated.
Without a workflow, you may need to:
- copy information into an AI chat;
- repeat the same instructions;
- move the answer into another application;
- complete another AI step;
- check which version of the instruction you used; and
- repeat the process again for the next item.
A workflow brings these steps together in one organised process.
You build the process once, test it, and use it again with new information.
The main parts of a workflow
Most workflows contain some combination of the following parts.
| Part | Purpose |
|---|---|
| Input | Receives the information the workflow needs |
| AI task | Uses a selected model to complete a defined task |
| Decision | Chooses which path to follow |
| Tool | Retrieves information or performs an approved action |
| Transformation | Organises, combines, or changes information |
| Output | Returns or saves the final result |
A workflow does not need every part.
The simplest useful workflow may contain only an input, one AI task, and an output.
How information moves through a workflow
Connections show where information goes next.
For example:
Customer message
→ Identify the main issue
→ Prepare a draft reply
→ Return the result
The customer message enters the workflow.
The first AI step identifies the issue.
The next step uses that result to prepare a draft.
The output returns the completed response for review.
Each step depends on the information passed from the previous step.
A workflow is more than one prompt
A prompt gives an AI model one instruction.
A workflow can contain several prompts, decisions, tools, and outputs.
For example, one prompt might classify a message. Another prompt might prepare a response based on the category.
This makes it possible to separate a large task into smaller, clearer actions.
Smaller steps are easier to:
- understand;
- test;
- review;
- improve;
- replace; and
- troubleshoot.
A simple document workflow
Imagine that you regularly receive project updates.
A workflow could:
- receive the update;
- create a short summary;
- extract blockers;
- identify deadlines;
- prepare a list of next actions; and
- return the result.
The flow may look like:
Project Update
→ Summarise
→ Extract Blockers
→ Identify Deadlines
→ Prepare Actions
→ Final Output
Each time a new update arrives, you can run the same process again.
A simple classification workflow
A workflow can also sort information into categories.
For example:
Incoming message
→ Classify the message
→ Question
→ Complaint
→ Request
→ Other
Each category can lead to a different next step.
A complaint might be sent to a response-drafting step.
A question might be sent to an information lookup.
An unclear message might be sent for human review.
A workflow can use AI and non-AI steps
Not every block needs an AI model.
A workflow may also include steps that:
- receive input;
- combine information;
- check a condition;
- format a result;
- use a tool;
- write to the Journal; or
- return an output.
Use AI where interpretation, language, classification, or generation is needed.
Use simpler steps where the task follows a fixed rule.
A workflow can use more than one model
Different AI steps can use different models.
For example:
- one model classifies the input;
- another prepares a detailed analysis; and
- a faster model formats the final response.
This can be useful when different parts of the task have different needs.
Begin with one model when possible.
Add another model only when testing shows a clear benefit.
More models can make the workflow harder to review and maintain.
A workflow can use tools
Tools let a workflow retrieve information or perform supported actions.
A tool may allow the workflow to:
- search an approved source;
- read or write a supported file;
- create a Journal entry;
- retrieve current information;
- use a connected service; or
- perform another action provided by a Gene.
Review each tool before adding it to a workflow.
Check what information it receives and whether it performs a read or write action.
A workflow can make decisions
A decision lets the workflow follow different paths.
For example:
Document
→ Identify document type
→ Invoice: extract payment details
→ Contract: extract dates and obligations
→ Other: send for review
Decisions are useful when:
- the possible categories are clear;
- each category needs different handling; and
- every path can be tested.
Avoid adding branches when one clear path is enough.
A workflow can return more than one result
A workflow may produce several outputs.
For example, a meeting workflow could return:
- a short summary;
- a list of decisions;
- a table of action items; and
- a list of unanswered questions.
These results may be combined into one final output or sent to different destinations.
Decide what the user needs before building the workflow.
Workflows and Workbench are different
Workbench is designed for interactive conversations.
It is useful when:
- the task is still being explored;
- you expect follow-up questions;
- you are testing an instruction;
- you want to compare models; or
- you only need the result once.
Studio is designed for repeatable workflows.
It is useful when:
- the steps are clear;
- the process should be reused;
- the result needs a consistent format;
- tools or decisions are involved; or
- the flow should be run through RunFlows.
A common path is to test the task in Workbench before building it in Studio.
Workflows and RunFlows are different
Studio is where you design and test the workflow.
RunFlows is where you use a saved workflow.
A normal journey is:
- understand the task;
- test the instruction in Workbench;
- build the process in Studio;
- test the workflow;
- save it; and
- run it with new information through RunFlows.
Studio is the design area.
RunFlows is the execution area.
Local and cloud workflows
A workflow can use cloud models, local models, or a combination of both.
A local workflow can continue without an internet connection when every required model, tool, source, and destination is available locally.
A workflow needs internet access when it uses:
- a cloud AI provider;
- an online tool;
- a remote data source;
- an external service; or
- a Gene that connects to the internet.
Review every step before deciding whether the complete workflow is local.
Benefits of AI workflows
A well-designed workflow can help you:
- repeat the same process consistently;
- reduce manual copying between applications;
- reuse successful instructions;
- make each step visible;
- choose different models for different tasks;
- include tools in a controlled way;
- review where an error occurred; and
- save time on repeated work.
A workflow does not guarantee a correct result.
It gives you a clearer and more repeatable process to test and review.
When a task is suitable for a workflow
A task may be suitable when:
- you already perform it regularly;
- the starting information follows a recognisable pattern;
- the steps can be explained clearly;
- the expected result is known;
- the result can be reviewed;
- missing information can be handled; and
- errors can be made visible.
Begin with a task you already understand.
It is difficult to automate a process that is unclear even when performed manually.
When not to build a workflow yet
Keep the task in Workbench when:
- the goal changes often;
- every case needs a different process;
- you do not yet know what a good result looks like;
- the instruction is still being developed;
- the source information is unpredictable; or
- the task needs constant human judgement.
Explore and test the task first.
Build the workflow after the process becomes clear.
Start with a small process
A first workflow should be easy to understand.
Begin with:
Input → AI Task → Output
For example:
Notes → Create Summary → Return Summary
Test this simple version before adding:
- another AI step;
- a decision;
- a tool;
- a second model;
- a saved destination; or
- a schedule.
Adding one change at a time makes problems easier to find.
Give each step one clear purpose
A step should do one understandable job.
Instead of asking one block to summarise, classify, extract, compare, and prepare a report, divide the work into focused steps.
For example:
Input
→ Classify
→ Extract Details
→ Create Summary
→ Format Output
Focused steps make it easier to identify where a result became incorrect.
Name the workflow clearly
A useful workflow name describes the result.
Examples include:
- Summarise Meeting Notes;
- Classify Customer Requests;
- Extract Invoice Details;
- Prepare Weekly Project Update; or
- Review Support Handover.
Avoid unclear names such as:
- Test;
- New Flow;
- Workflow 1; or
- Final Version.
A clear name makes the workflow easier to find in RunFlows.
Plan the input
Decide what information the workflow expects.
The input may be:
- text;
- a document;
- a message;
- a set of fields;
- a category;
- a date; or
- another supported value.
Explain the expected input clearly.
A workflow designed for meeting notes may not handle a full contract correctly without changes.
Plan the output
Decide what the workflow should return.
The output may be:
- a summary;
- a table;
- a list;
- a draft;
- structured fields;
- a Journal entry; or
- another supported result.
Choose a format that is easy to review.
When the result will be used by another step, keep the format consistent.
Handle missing information
Decide what should happen when a required detail is missing.
An AI instruction may say:
If the source does not include the information, write "Not provided."
Do not guess.
A decision may send incomplete input to a review path.
A workflow should not hide missing information behind a confident result.
Handle unexpected input
Test the workflow with input that does not match the normal example.
For instance:
- an empty message;
- a very short note;
- an unrelated document;
- missing fields;
- several languages;
- unusual formatting; or
- conflicting information.
Decide whether the workflow should stop, ask for better input, or return a clear warning.
Make errors visible
A workflow can fail because:
- a provider is unavailable;
- a model cannot complete the request;
- a tool is not configured;
- a connection is missing;
- the input is incomplete; or
- a step returns an unexpected format.
A useful workflow makes the problem visible.
Do not allow an important failure to appear as a normal final result.
Test every path
Test:
- the normal path;
- each decision path;
- missing information;
- unexpected input;
- tool errors;
- provider errors; and
- the final output.
Review intermediate results when available.
One successful run is not enough to show that the workflow is ready.
Keep human review where it matters
A workflow can prepare information without making the final decision.
Human review is especially important when the result affects:
- customers;
- employees;
- money;
- contracts;
- legal rights;
- health;
- safety;
- security; or
- access to important services.
Use workflows to organise, summarise, draft, or recommend.
Keep important approvals and final decisions with a person.
Improve the workflow over time
A workflow may need changes when:
- the task changes;
- a provider changes its available models;
- a tool is updated;
- the source format changes;
- users need a different output; or
- testing reveals a repeated problem.
Review saved workflows regularly.
Keep names, instructions, models, tools, and expected inputs up to date.
A practical workflow checklist
Before using a workflow regularly, confirm that:
- the purpose is clear;
- every block has one job;
- all connections are correct;
- the expected input is defined;
- the final output is reviewable;
- missing information is handled;
- every decision path has been tested;
- tool actions have been confirmed;
- errors are visible; and
- human review is included where needed.
The next step
Once you understand the task you want to repeat, open Studio and build the simplest useful version.
Start with:
Input → AI Task → Output
Test it with several examples.
Then add decisions, tools, extra models, or saved destinations only when they solve a clear need.