What is an AI workflow? A plain definition, its anatomy (trigger, steps, decision points, actions), and how it differs from a pipeline, an agent and ML.
What is an AI workflow? It is a sequence of steps, started by a trigger, in which at least one step requires judgement rather than a fixed rule. A model reads, classifies, drafts or decides at that step; everything around it can still be deterministic. The term confuses readers because it is used for automation products, ML pipelines and agent platforms alike.
What is an AI workflow?
An AI workflow is an ordinary workflow with one change: somewhere in the sequence, a step is handled by a model that weighs its inputs instead of matching them. The rest of the definition is the same as for any workflow. A trigger starts it, steps run in order, actions change something in a tool, and the run ends.
The distinction that matters is deterministic versus probabilistic. A fixed rule returns the same output for the same input every time: if the form field says “Enterprise”, route the lead to the enterprise queue. A model’s output is a judgement: read the whole form, the company website and the free-text message, then decide whether this looks like an enterprise buyer. Run it twice on a borderline input and the answer can differ.
So the test is whether the result of any step depends on interpretation, not whether a model is present. A workflow that calls a model only to reformat a date is still rule-based in practice. A workflow in which a model decides which invoice lines look wrong is an AI workflow, even if every other step is a database write.
Automation vs AI makes the full rules-versus-judgement argument. This page takes that distinction as given and builds the vocabulary around it.
The anatomy of an AI workflow
Every AI workflow has four parts, and judgement belongs in only some of them.
Part | What it is | Rule or judgement |
|---|---|---|
Trigger | The event that starts a run: a new CRM record, an inbound message, a schedule, a file landing in a folder | Rule. Triggers are matched, not interpreted |
Steps | The ordered work between trigger and outcome: fetch, transform, draft, check | Mostly rule. A step becomes a judgement step when a model has to read and decide |
Decision points | Where the run branches: approve or reject, route A or route B, escalate or close | Judgement, when the branch depends on meaning rather than a field value |
Actions | What changes in the world: an email sent, a record updated, a ticket opened, a report posted | Rule. The action itself is fixed; only the choice of action was judged |
A worked example. Trigger: a support ticket arrives. Step one, a rule: fetch the customer’s plan and recent tickets. Step two, judgement: a model reads the ticket and the history, then classifies urgency and topic. Decision point: if urgent and about billing, escalate; otherwise, draft a reply. Step three, judgement: the model drafts the reply from the help center. Action, a rule: post the draft to the support rep’s queue. Nothing reaches the customer until a person approves it. The judgement steps are the ones that need logging, review and a human gate.
The same workflow can be described instead of drawn. Viktor, an AI employee who works in Slack or Microsoft Teams, takes it as a plain-language message in the channel: “Every weekday at 8am, read new support tickets, post anything about billing from an Enterprise account in #support-escalations, and draft replies for the rest.” His product page puts it as “No workflows to build. Just delegate.” and states that “Viktor figures out the steps, the tools, and the schedule.” A run like that one, repeated on a schedule, is a Task. The gate stays where it was: his security page states that “Money moves, code pushes, and customer emails wait for your explicit approval in Slack.”
AI workflow vs ML workflow: what’s the difference?
Both carry the word “workflow” and both involve a model, but they are different objects. An ML workflow produces a model. An AI workflow uses one.
| ML workflow | AI workflow |
|---|---|---|
Purpose | Build, evaluate and maintain a model | Produce a business outcome, using a model at one or more steps |
Starts with | Raw data | A business event: a record, a message, a schedule |
Core steps | Collect, clean, label, train, evaluate, deploy, monitor | Trigger, fetch, judge, decide, act |
Ends with | A model artifact and its evaluation metrics | A changed record, a sent message, a posted report |
Typical owner | Data scientists and ML engineers | Operations, finance, sales, support, with or without engineering |
How often it runs | Per training or retraining cycle | Per event, or on a schedule |
Where the model comes from | Built inside the workflow | Called from outside, usually a hosted model through an API |
The two meet at exactly one point: the model an ML workflow produces can be the one an AI workflow calls. The model at the judgement step can be a machine learning (ML) classifier trained in such a workflow, or a large language model (LLM), a generative AI model that reads and writes text, which supplies the natural language processing (NLP) a ticket, email or form message needs. An ML workflow is concerned with data quality, training runs and drift. An AI workflow is concerned with triggers, tools, approvals and whether the judgement step gave a defensible answer.
A quick test: if a page describing “AI workflows” talks about feature stores, training data and model registries, it is describing ML workflows. If it talks about triggers, tickets, CRMs and approvals, it is describing AI workflows.
Workflow, pipeline, orchestration, agent: the vocabulary readers keep colliding with
Four terms get used interchangeably and should not be.
Workflow. A predefined path from trigger to outcome. The order of steps is fixed in advance; what varies is the content flowing through them and, in an AI workflow, the judgement at particular steps.
Pipeline. A workflow that moves data rather than work. Pipelines take input at one end and produce transformed output at the other, usually in batches or streams, with no decisions about what to do next. Data pipelines and ML pipelines are both pipelines; a support-ticket workflow is not, because it branches and acts.
Orchestration. The layer that runs workflows: it fires triggers, sequences steps, retries failures, passes outputs between steps and records what happened. Orchestration is plumbing; it decides nothing itself.
Agent. A loop rather than a path. An agent is given a goal and chooses its own next step at runtime, calls tools, inspects the result and chooses again, until it judges the goal met. Where a workflow fixes the path and leaves judgement to specific steps, an agent leaves the path itself to judgement. AI agent vs AI assistant separates agents from tools that only answer, and What is agentic AI covers the loop in detail.
When is an AI workflow the right fit?
Three questions settle it.
- Can every step be written as a rule? If yes, use plain rule-based automation. A model adds cost and variability without adding anything a rule could not do.
- Is the path known but at least one step needs interpretation? If yes, an AI workflow fits. Fix the path, put the model at the judgement step, gate the actions that follow it.
- Is the path itself unknown until the work starts? If yes, the job is agent-shaped, and oversight has to cover the path, not only the output.
Routing, classification, drafting, reconciliation and summarizing against a known set of tools are workflow-shaped. Open-ended research that decides which source to read next is agent-shaped. Form-to-row copying is rule-shaped.
Building the second kind, choosing tools, writing the brief and scheduling the runs are covered in the AI workflow automation guide. This page stops at the definition.
Benefits and limitations of AI workflows
A refund request written in a customer’s own words slips past a keyword rule. An AI workflow reads it, classifies it and routes it anyway.
Benefits
- It accepts free-text messages and emails that rule-based automation rejects.
- The path stays predictable, because the trigger, fetches and actions remain deterministic.
- Routing, classification, drafting, reconciliation and summarizing become jobs an operations lead can describe.
Limitations
- The judgement step can return different answers on the same borderline input.
- The judgement step adds model cost that a rule step doesn’t.
- It needs logging, sampling and a human approval gate before any external action.
Put a model only at the step a rule can’t handle, and gate the action that follows it.
Common misconceptions about AI workflows
“An AI workflow is just automation with a model bolted on.” The model changes what the workflow can accept. Rule-based automation fails the moment the input stops matching the pattern; an AI workflow accepts the messy input and makes a call. That is a different capability, with a different failure mode and different oversight.
“An AI workflow is an ML pipeline.” An ML pipeline produces a model. An AI workflow consumes one. Confusing them means staffing a data science project for what an operations lead could describe in a paragraph.
“AI workflows are autonomous by default.” They are not. A workflow runs a fixed path, and the judgement steps inside it can and should be gated. A model classifying a ticket is not autonomy; a model sending a refund without anyone looking is a design choice someone made. Judgement steps whose output triggers an external action, such as an email, a payment or a deployment, need a human approval gate.
“The judgement step makes the whole workflow unreliable.” Only the judgement step is probabilistic. The trigger, the fetches and the actions stay deterministic, and the judgement step can be logged, sampled and reviewed. A workflow with one gated judgement step can go wrong in one place; an agent can go wrong in its path, its tool choice and its stopping condition.
Related concepts
RPA. Robotic process automation drives existing software interfaces, clicking and typing where a human would, following scripted steps. RPA vs AI agents covers where scripted interface automation ends and model judgement begins.
AI automation. The umbrella term for software that uses an AI model to do work a fixed rule could not handle, or that a person would otherwise do by hand. An AI workflow is one shape of AI automation: a fixed path with judgement at specific steps. Agents, which choose their own path, are another.
Agentic AI. Systems built around the agent loop: goal in, self-chosen steps, tools, result. An AI workflow can call an agent at one step; the two remain distinct shapes.
AI assistants. Tools that answer, draft or summarize in response to a prompt, without triggers, steps or actions of their own. An assistant can sit inside a judgement step; it is not a workflow.
AI employee. A product that takes the workflow as a plain-language description, runs the steps in the connected tools, and reports back, with sensitive actions held for approval. Viktor is one; the anatomy section above shows the support-ticket workflow in that form.
Judgement steps read company data, so where that data lives and who else can see it is part of the definition in practice.
See how Viktor keeps data isolated →
See usage-based pricing: up to $100 in free credits, no credit card →
FAQ
How is an AI workflow different from RPA?
RPA replays scripted actions on a software interface: open this screen, copy this field, click this button. Every step is a rule, and it breaks when the screen or the input changes. An AI workflow has at least one step where a model interprets the input and decides. RPA can be a rule step inside an AI workflow, but it cannot supply the judgement step.
What is the difference between an AI workflow and an AI agent?
An AI workflow follows a path fixed in advance and uses judgement at specific steps. An AI agent chooses its own path at runtime, deciding which tool to call next until it judges the goal met. The workflow is more predictable and easier to gate; the agent handles jobs where the path cannot be known before the work starts.
What is an example of an AI workflow?
A support ticket arrives (trigger), the customer’s plan and history are fetched (rule), a model classifies urgency and topic (judgement), billing issues are escalated while the rest get a drafted reply (decision point), and the draft waits in a rep’s queue for approval (action). Lead routing, invoice reconciliation and weekly reporting share the same shape.
How do I build my first AI workflow, and which tools can I use?
Pick one recurring job whose path is already known, such as ticket triage or lead routing. Write down the trigger, the steps, the one step that needs judgement and the action that follows it, then put an approval gate on that action.
A rule-based automation product fits when no step needs judgement. An AI employee such as Viktor fits when one does: he takes the workflow as a plain-language message in Slack or Microsoft Teams and connects to 9,800+ tools.
The AI workflow automation guide linked above covers tool choice, the brief and scheduling.
Do AI workflows need human oversight?
Yes, at the judgement steps whose output triggers an external action. Classifying and drafting can run unattended; sending, paying and deploying should wait for approval until the error rate on that step is known. Oversight belongs at the gate, not on every step.
How much does it cost to run an AI workflow?
The cost depends on the product running it, and each vendor publishes its own pricing unit. The judgement steps drive it, because a model adds cost a rule step doesn’t, so count how many judgement steps each run makes. Viktor’s pricing is usage-based with no per-seat fee, and new users start with up to $100 in free credits, no card required.