AI workflow guide
When should you use AI in a workflow — and when not?
AI is useful when a workflow needs to interpret language, documents or ambiguous information. It is usually unnecessary when the next step can be determined with a clear rule. Reliable AI workflows combine both!
At a glance
Rules, AI or a person?
Most workflows are not “an AI workflow” or “a rules workflow”. They are a sequence of steps, and each step has its own best fit.
| The step | Usually best handled by |
|---|---|
| The same input should always produce the same action | Rules |
| An exact calculation or validation | Rules |
| Moving data between systems | Rules |
| Routing based on known fields | Rules |
| Understanding an email or document | AI |
| Extracting meaning from unstructured text | AI |
| Classification where the wording varies | AI |
| Drafting or summarising text | AI |
| An ambiguous judgement | AI or a person |
| An irreversible or high-impact decision | Human control |
Start with rules
If you can write it as if → then, you probably do not need AI.
Many workflow steps already have a clear answer. When the condition and the action can be described completely, a deterministic rule is the better tool: it behaves the same way every time, so it is easier to test, inspect and change.
| If | Then |
|---|---|
| Country = Belgium | Assign the Belgian sales team |
| Form = demo request | Create a follow-up task |
| Invoice total > €10,000 | Request approval |
| API returned 429 | Wait and retry |
Rules can still fail — a field can be empty, an API can change. But when they fail, the failure is usually easier to see and trace. If the system already knows the answer, AI should not have to guess it.
Where AI helps
Use AI when the input needs interpretation.
AI becomes useful when the workflow has to understand information that does not arrive in a fixed format: an email written in someone’s own words, a CV with its own layout, a request that could mean several things.
The AI step interprets that input and returns a structured result. The rest of the workflow can then work with fields and values again. AI can interpret information without controlling the entire workflow.
Typical AI tasks in a workflow
- Classifying incoming emails
- Extracting structured information from documents
- Understanding the intent of a request
- Summarising free text
- Categorising information when the wording varies
- Drafting communication
- Proposing a classification based on several inputs
From unstructured input to structured data
- 01 · InputAn email or document arrives
- 02 · AIThe content is interpreted
- 03 · OutputA structured result the workflow can check and use
Hybrid workflows
The strongest workflows combine AI and rules.
In a reliable workflow, AI handles the interpretation and explicit workflow logic controls everything around it: which input is processed, whether the output is complete and valid, where it goes next and what is written to other systems.
This keeps the uncertain part small and well defined. When the AI output is incomplete or unexpected, the surrounding checks can stop it before it reaches a CRM, an inbox or a customer.
AI interprets. The workflow decides what happens with that interpretation.
A hybrid workflow, step by step
- 01 · TriggerAn email arrives in a monitored mailbox.
- 02 · RulesRelevance checks decide whether this workflow should process it.
- 03 · AIThe email and attached document are interpreted.
- 04 · RulesThe structured output is validated against required fields and formats.
- 05 · AIA classification or assessment is proposed.
- 06 · RulesBusiness rules decide the route, based on that result and known data.
- 07 · HumanUncertain or high-impact cases go to a person for review.
- 08 · RulesThe CRM or business system is updated through controlled integration.
In production
How Logi-technic’s recruitment workflow divides the work.
At Logi-technic, staffing agencies send candidate proposals by email, with the details spread across the message, CV attachments and later replies. Flo, the workflow I built for them, uses AI for the parts that need interpretation and workflow logic for everything that can be defined.
AI supports the assessment; it does not make the hiring decision. Predefined workflow logic decides how its output is used, and recruiters work with the result.
Workflow logic handles
- When the workflow runs and which messages it picks up
- Relevance checks on incoming messages
- Lookups for existing Salesforce records
- Validation of identity, contact details, agency, role and CV
- Routing of exceptions and incomplete cases
- The Salesforce/Connexys integration
- Error handling and processing statuses
AI assists with
- Structuring information from candidate emails and documents
- Interpreting work experience and education
- Proposing a technical core profile
- Weighing that information against detailed technical role profiles
- Preparing the assessment information used in the follow-up
Where the AI steps sit in Flo
- 01 · WorkflowCollect and route proposals
- 02 · AIStructure the candidate information
- 03 · AIAssess against role profiles
- 04 · WorkflowCheck records, prepare follow-up
- 05 · WorkflowValidate, then update Salesforce
Human in the loop
Keep people where the consequences justify it.
Not every AI step needs a person to approve it. Asking for approval on every routine action mostly moves the manual work somewhere else. Human review is most valuable at the points where a mistake would be costly or hard to undo.
Human review matters more when an action is
- Difficult to reverse
- Customer-facing
- Financially significant
- Based on uncertain information
- Legally or compliance-sensitive
- Able to change important records or permissions
Selective control works better than blanket approval: let the workflow run where the risk is low, and route the cases that matter to a person.
Agents and workflows
Do you need an AI agent or a workflow?
Short answer
Use a workflow when you know the steps. Use an agent when the system needs to decide which steps or tools are required. Many business automations need a workflow with a small amount of AI, not a fully autonomous agent.
An agent is useful when the path cannot be planned in advance, for example when it has to investigate a question across several tools. Most operational processes, such as intake, routing, enrichment and registration, have a known sequence. A workflow makes that sequence explicit, which makes it easier to test, explain and maintain.
More autonomy is not automatically better automation.
How a workflow runs
Workflow, steps 1 to 4
- 01 · TriggerSomething happens
- 02 · StepsKnown steps run in order
- 03 · ConditionsRules choose the branch
- 04 · ActionA defined result
How an agent runs
Agent, steps 1 to 4
- 01 · GoalA task to accomplish
- 02 · AIAI decides the next action
- 03 · ToolsIt uses one of the tools it may use
- 04 · LoopIt repeats until the goal is met
Approach
How I design AI-assisted workflows.
This is the approach behind my AI workflow and business process automation projects. It applies whether the workflow runs in n8n or another platform, and whether AI ends up handling one step or several.
01Map
Map the process as it works today, including the exceptions.
02Separate
Separate the deterministic decisions from the steps that need interpretation.
03Structure
Give AI structured inputs and ask for structured outputs where practical.
04Validate
Validate the output before it triggers consequential downstream actions.
05Control
Add human control where the consequences justify it.
06Monitor
Monitor failures and edge cases, and adjust the rules or prompts.
Quick check
A quick decision framework for each step.
Go through these questions for every step you are considering for AI.
Can the decision be expressed reliably as a rule?
YesUse a rule.
NoContinue with question 2.
Does it require interpreting language or ambiguous information?
YesConsider AI.
NoContinue with question 3.
Does the action have significant consequences?
YesAdd appropriate human control.
NoContinue with question 4.
Can the output be validated before acting on it?
YesAutomate.
NoRedesign the step.
Questions about AI and rules
AI vs rule-based automation, answered briefly
- What is the difference between AI automation and rule-based automation?
- Rule-based automation follows explicit conditions: the same input always leads to the same action. AI automation uses a model to interpret input such as text or documents and returns a likely result. Rules execute what is known; AI helps where the input first has to be understood.
- When should I use AI in a business workflow?
- When a step needs to interpret something a rule cannot capture reliably: free-text emails, documents with varying layouts, intent, summaries or classifications where the wording varies.
- When should I avoid using AI?
- When the step can be described completely as a rule, when the answer already exists in a system, or when an exact result is required, such as calculations, mappings, validation and routing on known fields.
- Is AI automation reliable enough for business processes?
- It can be, when AI is limited to the steps that need interpretation and its output is validated before it is used. AI output is a likely answer, not a guaranteed one, so the workflow around it should expect occasional mistakes and handle them.
- Should AI be allowed to update a CRM automatically?
- AI can prepare the data, but the lookup and the write are usually better handled by controlled integration logic with validation first. Whether an update needs review depends on how important the record is and how reliable the input is. More on CRM automation.
- Does every AI workflow need human approval?
- No. Human review is most useful where actions are hard to reverse, customer-facing, financially significant, sensitive or based on uncertain information. Low-risk steps with validated output can run automatically.
- What is the difference between an AI agent and an automated workflow?
- A workflow follows steps that are defined in advance. An agent receives a goal and decides which steps or tools to use. Many business processes have known steps and are better served by a workflow with AI in specific places.
- Can n8n combine AI with normal automation?
- Yes. n8n can run model calls as steps inside a workflow that also handles triggers, conditions, validation, API calls and system updates.
- How do you prevent AI errors from affecting other systems?
- Errors cannot be ruled out, but their impact can be limited: ask for structured output, validate it against required fields and known values, keep writes in controlled integration steps, route unclear cases to review and keep the source data available.
- Is AI workflow automation GDPR compliant?
- That depends on the project, not on the tool or the architecture. Compliance depends on which personal data is processed and why, the legal basis, the processors and where they process data, the configuration, access, retention and other project-specific requirements. A workflow can be designed to support those requirements, for example by minimising what is sent to a model. This is practical guidance, not legal advice.
AI workflow project
Not sure which steps need AI?
Bring the process as it works today. We can go through it step by step and decide what should be a rule, what needs AI and where a person should stay involved.