AI workflows
Use AI where rules stop being enough.
I build workflows that use AI to interpret messy input — emails, documents, free text and other unstructured data — while deterministic logic controls validation, integrations and what happens next.
AI in context
The useful part is not the model. It is what surrounds it.
Calling an AI model is relatively easy. The work is in the process around it: AI handles the interpretation, and workflow logic controls what happens next.
What a production workflow has to know
- When AI should run
- What context it receives
- What output format is expected
- Whether required information is present
- How the output is validated
- What happens if the model fails
- Which system receives the result
- Which actions are allowed automatically
- When a person should take over
A model produces an answer. A workflow turns that answer into a controlled business process.
Where AI fits
Three AI tasks, and the workflow that continues.
AI helps with understanding, classifying and preparing. The fourth group is not AI: it is the workflow logic and integrations that check the output and carry the process forward.
Understand
- Emails
- Free text
- Documents
- Notes
- Extract information
- Identify intent
- Summarise content
- Turn unstructured input into structured fields
Classify
- Requests
- Profiles
- Messages
- Documents
- Categorise input
- Propose a type or profile
- Identify relevant characteristics
- Support routing or assessment
Prepare
- Replies
- Summaries
- Internal notes
- Structured output
- Draft a response
- Prepare information for review
- Generate a structured summary
- Transform data into the required format
Continue (workflow logic)
- CRM
- APIs
- Internal systems
Not AI: the deterministic part of the workflow that takes the AI output and decides what happens next.
- Validate the AI output
- Apply deterministic rules
- Look up existing records
- Create or update data
- Route exceptions
In production
AI interprets the candidate. The workflow controls the process.
At Logi-technic, Flo processes targeted candidate proposals from staffing agencies: the agency’s email, supported CV or documents and the candidate information in them. AI helps turn that unstructured input into structured information, such as candidate details, experience, education, certificates and relevant profile information.
Logi-technic’s detailed technical role profiles provide structured input for the assessment. AI proposes a core profile, supports the assessment and prepares information for the next steps. AI supports interpretation and assessment. Predefined workflow logic controls how that output is used.
The AI step is surrounded by deterministic checks, system lookups and integration logic.
Where the AI steps sit in Flo
- 01 · ReceiveA staffing agency sends a candidate proposal and supported documents.
- 02 · ExtractAI turns unstructured candidate information into structured data.
- 03 · AssessThe information is considered against the detailed technical role profiles.
- 04 · PrepareThe workflow prepares the assessment and the follow-up information.
- 05 · ValidateDefined checks and workflow conditions decide whether the process can continue.
- 06 · RegisterThe workflow can create or supplement Salesforce/Connexys records and transfer the CV.
Control
Use probability for interpretation. Use rules for rules.
A model gives a likely answer; a rule gives the same answer every time. Each belongs in a different part of the workflow, and some decisions belong with a person.
The guide when to use AI in a workflow, and when not works through this split with examples.
Good candidates for AI
- Understanding free text
- Extracting information from varied documents
- Classification
- Summarisation
- Preparing draft responses
- Interpreting context
- Supporting an assessment
Better suited to deterministic logic
- Required-field checks
- Exact mappings
- Record lookup
- Ownership rules
- Status changes
- API requests
- CRM writes
- Fixed routing
- Retry behaviour
- Known business rules
Keep human judgement where needed
- Ambiguous cases
- Sensitive decisions
- Final approval
- Commercial judgement
- Exceptions with insufficient context
Do not ask AI to guess something the system already knows.
From text to data
AI becomes more useful when its output has somewhere specific to go.
A production workflow rarely asks AI for a long paragraph and stops there. It asks for structured information that can be validated, mapped, stored, compared, routed or sent to another system.
Two examples of what that can look like; the actual fields depend on the process.
Candidate CV + agency email → for example
- Name
- Phone
- Location
- Role
- Experience
- Education
- Certificates
- Availability
Customer request → for example
- Request type
- Product or topic
- Urgency indication
- Relevant identifiers
- Summary
Structure gives the rest of the workflow something predictable to work with.
Validation
AI output should be treated as input, not truth.
Models make mistakes, and no setup removes that completely. The workflow manages the risk by checking what comes back before anything depends on it.
Practical controls
- Check that required fields are present
- Validate formats
- Compare against known values
- Reject impossible values
- Use confidence or conditions where appropriate
- Route uncertain cases
- Keep the original source data available
- Log errors
- Avoid destructive actions without controls
The workflow should be designed for the possibility that the model is wrong.
Reliability
The workflow needs a plan for when AI does not answer.
AI calls fail in ordinary ways. In Flo, model and API capacity limits are a real failure mode, handled with retries and error handling. No workflow can make failure impossible, but it can make it visible and recoverable.
Possible failure modes
- API timeout
- Rate limit
- Unavailable model
- Malformed output
- Missing fields
- Unexpected document
- Insufficient context
Possible controls
- Retries
- Waiting and backoff
- Error outputs
- Logging
- Fallback paths
- Manual review
- Stopping before downstream writes
- Alerts where useful
A failed AI call should become a known workflow state, not a mystery.
Orchestration
AI is strongest when it can work with the systems around it.
n8n is not required for every AI workflow, but it is useful when several systems and controlled steps are involved. It can run the whole sequence around the model call.
What the orchestration layer does
- Trigger the workflow
- Gather context
- Call AI models
- Structure requests
- Validate responses
- Call APIs
- Query the CRM
- Write data
- Send email
- Route errors
- Continue the process
The model interprets. The orchestration layer coordinates.
CRM
AI can prepare CRM data without owning the CRM decision.
A common pattern is unstructured input → AI extraction → validation → CRM lookup → create or update → next action. The AI prepares the data; the lookup and the write stay with controlled integration logic.
Logi-technic is the production example, with Salesforce/Connexys. The same pattern applies to other CRMs; see CRM automation and the Salesforce and HubSpot pages.
Inputs this pattern can handle
- Candidate information
- Incoming requests
- Meeting notes
- Service messages
- Document data
Marketing
Use AI for messy inputs, not as an excuse to hide marketing rules.
In marketing automation, AI is useful where the input is unstructured. The rules that decide who receives what should stay explicit where appropriate.
AI may assist with
- Classifying inbound responses
- Summarising information
- Preparing content
- Interpreting free-text form fields
- Extracting attributes from unstructured data
Keep explicit where appropriate
- Consent
- Suppression
- Audience rules
- Lifecycle conditions
- Ownership
- Send and stop conditions
AI can enrich the decision context. The workflow should still know why someone enters, branches or exits.
AI agents
Not every automation needs an autonomous agent.
Some AI systems are designed to choose their own sequence of actions. That can be useful in some contexts. Many business processes, though, are already well understood, and they are better served by a controlled workflow. The right architecture depends on the process.
What a controlled workflow makes explicit
- Which actions are allowed
- Which systems it can access
- The order of the steps
- Validation before writes
- How exceptions are handled
- Who is responsible
If the process is already known, a controlled workflow is often easier to test, explain and maintain.
Data
Know what you are sending to the model.
Before business data goes to an AI provider, it helps to decide what actually needs to be sent. Where personal or confidential data is involved, the model and hosting setup should be selected with the project’s privacy and processing requirements in mind. This is practical guidance, not legal advice.
Questions to answer first
- Which fields are necessary?
- Is personal data involved?
- Where does processing happen?
- What are the retention settings?
- What do the provider’s terms say?
- Who has access?
- Can the data be minimised?
- Does the task need AI at all?
Questions about AI workflows
AI workflows, answered briefly
- What is an AI workflow?
- An automated process where an AI model performs one or more interpretation tasks — such as extraction, classification or summarisation — while workflow logic controls the surrounding systems, validation and next actions.
- What business tasks are suitable for AI automation?
- Tasks with unstructured email, documents or free text, classification, summarisation, information extraction and preparing responses. Tasks with exact, known rules often do not need AI.
- What is the difference between AI automation and normal automation?
- Normal automation works best when inputs and rules are predictable. AI helps when the input varies and needs interpretation. A production workflow can combine both.
- Do you build AI agents?
- I build AI-assisted workflows and can use agent-like patterns where they make sense. For many business processes, I prefer controlled workflows with explicit actions, validation and system boundaries.
- Can AI read emails and documents?
- Yes, depending on the file format, the model’s capabilities and the available access. It can help extract, classify and summarise information, though extraction is never guaranteed to be perfect.
- Can AI write directly into Salesforce or HubSpot?
- AI can prepare structured information, but the CRM lookup and write should normally be handled by controlled integration logic. More on CRM automation and the Salesforce and HubSpot pages.
- How do you prevent AI mistakes?
- They cannot be prevented completely, but they can be reduced and managed: with structured output, validation, deterministic rules, checks against known values, exception paths, human review where appropriate and by keeping the source data available.
- What happens if the AI model is unavailable?
- Depending on how critical the process is, the workflow can use retry logic, error handling, fallback paths, alerts or manual review.
- Can AI workflows use n8n?
- Yes. n8n can orchestrate the model calls and the surrounding CRM, email, API and business system steps.
- Do you build AI workflows in Belgium?
- VNCK works from Willebroek, Belgium and builds AI-assisted workflows that connect language models with CRM, email, documents, APIs and operational systems.
AI automation project
Start with the part of the process that still needs interpretation.
If a workflow currently stops because someone has to read an email, interpret a document, classify a request or turn free text into structured data, AI may be useful there. We can map the process and decide which steps should use AI, fixed rules or human judgement.