Trigger
What starts the process?
Form submitted · order created · email received · CRM stage changed · scheduled job · webhook
BUSINESS AUTOMATION · N8N · API INTEGRATION · AI AGENTS
Automate repeatable work without hiding the decisions that still need people.
OJC Labs designs business automation systems that connect your existing tools, move data between them and handle repetitive actions without turning every business decision into an AI experiment.
We use deterministic workflows where the rules are clear, AI where interpretation is genuinely useful, and human approval wherever the consequence deserves one.
n8n · APIs · CRM · Databases · AI · Human approval
AUTOMATION STARTS WITH THE PROCESS
Start with the process, not the connector.
A useful automation has a repeatable starting point, known inputs, clear decisions, defined actions and an owner when something goes wrong.
Choosing n8n, Make, Zapier or an AI-agent framework before mapping that process usually creates a more sophisticated version of the same confusion.
Input
What information does the process need?
Customer record · order · document · CRM fields · product data · message
Decision
Which rules or interpretations determine what happens next?
Status · amount · owner · category · qualification · confidence
Action
What should the system actually do?
Create · update · notify · send · enrich · assign · publish · synchronise
Exception
What happens when the expected path breaks?
Retry · queue · request approval · notify owner · stop safely
If these five pieces are unclear, another automation tool will not fix the process.
CHOOSE THE RIGHT EXECUTION MODEL
Not every automation needs an AI agent.
The architecture should match the uncertainty of the work.
Workflow
A predefined sequence where triggers, rules and actions can be described in advance.
Use when- The process is predictable
- Inputs are structured
- Decisions can be expressed with rules
- APIs or webhooks are available
- The same action should produce the same result
A new lead enters the CRM → source is recorded → owner is assigned → notification is sent → follow-up task is created.
Best-fit toolsn8n · Make · Zapier · Power Automate
Use determinism when determinism is enough.
Workflow + AI
The workflow remains deterministic, but one bounded step uses an AI model for interpretation or generation.
Use when- Classify an enquiry
- Extract data from unstructured text
- Summarise a document
- Generate a draft
- Categorise content
- Enrich metadata
- Compare text meaning
Email arrives → workflow extracts the text → AI classifies the request → deterministic rules route it to the correct team.
AI solves the ambiguous step. It does not need to own the whole process.
AI agent
A model receives instructions and a set of permitted tools, then decides which tool or sequence of tools to use for the current task.
Use when- The path cannot be fully predetermined
- Several tools may be relevant
- Interpretation affects the next action
- The task benefits from planning or delegation
- The agent can operate inside controlled permissions
A support agent reads a request, checks the order system, searches the knowledge base and decides whether to answer or escalate.
The agent chooses among allowed actions. It should not invent new authority.
OpenAI Agents SDK — Agents and tools →RPA / computer automation
Automation interacts with software through its user interface when a reliable API or integration is unavailable.
Use when- Legacy software has no API
- The required operation exists only in the interface
- Manual clicking is repetitive
- Replacing the system is not currently practical
Open legacy software → find customer → export report → save file → continue workflow.
Interface automation is usually more fragile than an API integration because layouts, authentication and UI states can change.
API first. RPA when the interface is genuinely the only practical route.
Use the least autonomous architecture that can reliably solve the problem.
AI AGENTS · WITHOUT THE THEATRE
An AI agent is not an employee, a business model or a money printer.
It is software operating with instructions, context and permitted tools.
If someone says “my agent made £50,000”, the missing question is what actually caused the revenue.
A Stripe screenshot plus an n8n canvas is not a causal model.
Agents can create leverage. They do not suspend attribution.
TO ATTRIBUTE THAT CLAIM, YOU WOULD STILL NEED TO KNOW
- Where the customers came from
- What offer was sold
- Whether advertising generated the demand
- Whether humans closed the sale
- What the existing customer base contributed
- How much the system cost to operate
- How failures were handled
- Whether the revenue was booked or actually collected
- What would have happened without the agent
Instructions
What is the agent supposed to achieve and what must it never do?
Context
What information may it use for the current decision?
Tools
Which APIs, databases, workflows or external systems may it call?
State
What does the system need to remember between actions?
Guardrails
Which inputs, outputs or tool calls must be checked or blocked?
Human control
Which decisions require approval, escalation or review?
OpenAI Agents SDK — Agents, tools, guardrails and human approval →
SAME BUSINESS PROBLEM · DIFFERENT ARCHITECTURE
The difference becomes obvious when we make it concrete.
Example: an incoming sales enquiry.
Deterministic workflow
TriggerWebsite form submitted
Then- Validate fields
- Create CRM contact
- Assign by territory
- Send confirmation
- Create task
- Notify salesperson
No agent required.
Agent-assisted version
TriggerComplex free-text enquiry submitted
Then- Workflow validates contact data
- AI interprets the unstructured request
- AI can query approved product or service information
- Result is returned in structured fields
- Workflow applies qualification rules
- CRM contact is created
- Human owner receives the summary and source material
The AI handles interpretation. The workflow still owns the record and the side effects.
AGENT TOOLS
An agent becomes useful when it can interact with real systems safely.
The model itself does not contain your CRM, inventory, inbox or database. Those capabilities have to be exposed as tools.
Search / knowledge
Search API · vector search · knowledge base · documentation
Finding information before deciding.
CRM
HubSpot · Salesforce · Pipedrive · Zoho · Odoo
Reading or updating contacts, deals, stages and tasks.
Communication
Email · Slack · Teams · WhatsApp Business
Sending messages, requesting approval or escalating work.
Database
Supabase · PostgreSQL
State, records, audit history, queues and application data.
Business systems
Sage · Odoo · Shopify · WooCommerce · custom ERP
Orders, customers, stock, invoices or operational records.
Custom APIs
Internal API · vendor API · webhook · MCP tool
Capabilities that do not exist inside a standard connector.
An agent does not “know your business tools”. Someone has to design the interfaces, permissions and failure paths that make those tools safe to call.
TOOLS & MCP
MCP can standardise access to tools. It does not decide whether the tool should be trusted.
Model Context Protocol can expose tools and resources through a standard interface so compatible agent systems can discover and call them.
That solves part of the connection problem. You still need authentication, permissions, validation, logging and control over side effects.
MCP gives the agent
- Tool discovery
- Structured tool inputs
- A common integration interface
MCP does not automatically give you
- Good permissions
- Safe business logic
- Correct data
- Reliable approval rules
- A useful automation strategy
Protocol is not governance.
TOOLS WE CAN BUILD WITH
Choose the tool around the architecture, not the logo people are posting on LinkedIn this week.
OJC Labs is not tied to one automation platform. We choose based on integrations, control, hosting, complexity, maintainability and the systems already used by the business.
n8n
OJC DEFAULT FOR MANY CONNECTED WORKFLOWSA strong fit when the system needs webhooks, APIs, branching, transformations, code, AI steps and several connected services inside one workflow.
Good when- Custom API work matters
- Workflows are more than simple A → B actions
- Data needs transformation
- AI steps sit inside deterministic flows
- Execution needs to remain inspectable
- Deployment / hosting control matters
CRM sync · content pipelines · server-side events · enrichment · notifications · multi-system orchestration
n8n documentation →Make
Useful for visual SaaS integrations where the required applications already have solid modules and the workflow can be expressed cleanly as scenarios.
Good when- Many supported SaaS apps are involved
- Visual scenario design is useful
- Deployment speed matters
- Custom code is limited
Current Make AI Agents can use modules, scenarios and MCP tools as agent tools.
Make Developer Hub →Zapier
Useful for straightforward application-to-application workflows where speed and connector coverage matter more than deep custom orchestration.
Good when- Automation is relatively simple
- Supported integrations already cover the task
- Operational ownership should remain accessible to a non-development team
Power Automate
Often a logical choice when Microsoft 365, Teams, SharePoint, Outlook, Dynamics or other Microsoft systems already dominate the organisation.
Good when- Microsoft identity and permissions already matter
- SharePoint / Dynamics / Office data participates in the workflow
- The organisation already operates inside Microsoft's administration model
Custom agent runtime
Useful when the agent itself is part of a custom product or when tool use, handoffs, approvals, state and business logic need tighter control than a visual workflow platform provides.
Good when- The application owns the user experience
- Custom permissions are required
- Tool calls need application-level guardrails
- Several specialist agents may collaborate
- Tracing and evaluation are part of production operations
OpenAI Agents SDK · application code · custom APIs
OpenAI Agents SDK →Supabase / PostgreSQL
Not an automation builder. Often the useful data layer underneath one.
Good when- Source-of-truth records
- Workflow state
- Audit logs
- Queues
- Structured content
- Vector search with pgvector
- Storing enrichment results
- Tracking what has already been processed
A workflow engine moves work. A database remembers what happened.
Supabase documentation →
DATA ARCHITECTURE
Automation becomes dangerous when nobody knows which system owns the record.
Before synchronising data, define which system is authoritative for each object.
Customer
Possible owner
CRM
Order
Possible owner
E-commerce platform / ERP
Stock
Possible owner
ERP / inventory system
Content
Possible owner
CMS
Automation state
Possible owner
Database / workflow platform
Analytics event
Possible owner
Measurement pipeline
Synchronisation should move the information other systems need without creating several competing versions of the same truth.
FAILURE IS PART OF THE WORKFLOW
The happy path is not the automation.
Production automation needs to define what happens when an API is unavailable, a record is incomplete, a model gives an uncertain answer or the destination rejects the update.
Retry
Can the action safely run again?
Idempotency
Will retrying create a duplicate order, contact or message?
Queue
Can the work wait safely until the dependency recovers?
Timeout
How long should an external service be allowed to block the workflow?
Escalation
Who needs to know when automation cannot continue?
Audit
Can we reconstruct what the system attempted and why?
If nobody designed the failure path, the workflow is still a prototype.
HUMAN IN THE LOOP
Automation should remove repetitive effort, not remove accountability.
Safe to automate directly
- Synchronise approved records
- Resize / transform data
- Generate internal summaries
- Assign routine tasks
- Update deterministic fields
- Send agreed transactional notifications
- Create drafts
Approval may be required
- Unusual refunds
- Destructive database actions
- Financial commitments
- Publishing sensitive content
- Changing contractual information
- High-impact customer communication
- Exceptions with low model confidence
The right question is not “Can AI do this?” It is “What happens if it is wrong?”
AUTOMATION BY INDUSTRY
The useful automation depends on how the business actually operates.
Do not sell a restaurant, an industrial company, an e-commerce store and a professional-services firm the same “AI agent”.
B2B sales & industry
Leads arrive from ads, forms, email, WhatsApp or events and need qualification before sales should spend time on them.
ArchitectureLead source → CRM → enrichment → qualification → owner → task → follow-up → opportunity → reporting
AI can help with- Classify free-text requirements
- Summarise enquiries
- Enrich company context
- Identify likely category / service
- Draft internal qualification summaries
- Contact creation
- Duplicate detection
- Territory rules
- Ownership
- CRM stage
- Notifications
- SLA timers
Meta / Google · website forms · HubSpot / Pipedrive / Salesforce / Odoo · n8n · email · WhatsApp
E-commerce & retail
Orders, stock, customer communication and acquisition data often live in separate systems.
ArchitectureStore → order → payment → inventory → fulfilment → customer → CRM / email → analytics
AI can help with- Classify support enquiries
- Product enrichment
- Categorisation
- Summarise reviews
- Generate draft product metadata
- Identify anomalies for human review
- Payment status
- Inventory sync
- Order creation
- Shipment updates
- Purchase events
- Transactional communication
Shopify · WooCommerce · Odoo · Sage · Supabase · n8n · Brevo · WhatsApp
Professional services
Enquiries need context, documents, qualification, scheduling and follow-up before work can begin.
ArchitectureForm / email → contact → qualification → document collection → booking → CRM → proposal → follow-up
AI can help with- Extract information from briefs
- Categorise requests
- Summarise documents
- Prepare first-draft scopes
- Identify missing information
- Client records
- File storage
- Reminders
- Meeting creation
- Owner assignment
- Status updates
Google Workspace / Microsoft 365 · CRM · Drive / SharePoint · Calendars · n8n / Power Automate
Marketing & content
Research, briefs, drafting, review, publishing and internal linking often become disconnected manual steps.
ArchitectureIdea → research → structured brief → draft → review → metadata → links → approval → CMS → measurement
AI can help with- Research synthesis
- Classification
- Metadata suggestions
- Content enrichment
- Semantic matching
- Draft generation
- Publishing status
- Approvals
- CMS updates
- Schema
- Canonical fields
- Logging
- URL creation
WordPress · Next.js · Supabase · OpenAI / other model provider · n8n · APIs
Customer support
Teams waste time identifying intent, looking up account context and routing routine enquiries.
ArchitectureEmail / chat / WhatsApp → identify customer → retrieve context → classify → answer or escalate → update ticket / CRM
AI can help with- Intent classification
- Retrieval from approved knowledge
- Summarisation
- Draft replies
- Multilingual assistance
- Permissions
- Customer identification
- Ticket state
- Escalation
- Refunds / account actions
- Audit history
Helpdesk · CRM · WhatsApp · email · knowledge base · agent runtime
The agent may draft or route. High-impact account actions should retain appropriate validation and permissions.
Finance & administration
Documents arrive through email or portals and need extraction, validation, approval and entry into business systems.
ArchitectureDocument → extract → validate → match → approval → accounting / ERP → archive
AI can help with- Extract unstructured information
- Classify document type
- Compare descriptions
- Flag unusual records
- Totals
- Identifiers
- Approval thresholds
- Duplicate checking
- Accounting entry
- Audit logging
Sage · Odoo · ERP · email · document storage · n8n / Power Automate
Do not let a language model become the accounting source of truth.
NEXT LAYER
These are systems, not six paragraphs we have to keep squeezing into one page forever.
As OJC Labs expands these implementations, each industry can become its own automation architecture page with workflows, integrations, examples and implementation requirements specific to that sector.
MARKETING AUTOMATION
Marketing automation should react to customer state, not blast everyone through the same sequence.
Capture
Form, purchase, event, campaign or other useful signal.
State
Prospect, qualified lead, customer, inactive customer or another relevant lifecycle state.
Rule
What communication or operational action is appropriate now?
Action
Email, CRM task, audience change, notification or another approved step.
Automation becomes useful when the lifecycle state is more important than the mailing-list membership.
SALES AUTOMATION
Automate the administration around selling, not the relationship that makes the sale happen.
Automation can handle
- Create CRM contact
- Enrich company record
- Assign territory
- Create task
- Send acknowledgement
- Schedule reminder
- Update lifecycle state
- Prepare meeting context
Keep human ownership for
- Qualification exceptions
- Negotiation
- Pricing decisions
- Relationship management
- Strategic account decisions
The goal is to give salespeople better context and fewer administrative steps, not to make prospects negotiate with a workflow.
AI AGENT ORCHESTRATION
More agents do not automatically create a better system.
Multi-agent architectures are useful when different specialists genuinely need separate instructions, tools, permissions or evaluation.
Manager + specialists
Manager agent can call: Research specialist · Data specialist · Drafting specialist · Validation specialist
Useful when one runtime should own the final result while specialists handle bounded tasks.
Handoff
Triage routes to: Sales · Support · Billing
Useful when the specialist should take responsibility for the next part of the interaction.
Do not create five agents when five functions would do.
NO MODEL REQUIRED
If an IF statement can make the decision, use the IF statement.
Do not use AI to decide
Is order total > £500?
Userule
Do not use AI to decide
Has invoice already been processed?
Usedatabase lookup
Do not use AI to decide
Which salesperson owns Tunisia?
Userouting table
AI may help
What type of commercial request is described in this free-text message?
Useclassification
AI may help
Which existing knowledge article is semantically closest to this support question?
Useretrieval / embeddings
Models are useful for ambiguity. Code is better for facts the system already knows.
REAL OJC SYSTEM
A controlled enrichment workflow built on Supabase.
This system enriches structured content while keeping the database, processing state and resulting records inspectable.
Record created / changed → queue → worker / workflow → AI enrichment → validation → structured result → database update → downstream publishing / search use.
The architecture discusses queueing, retry, structured outputs, pgvector / semantic similarity where used, validation, processing status and auditability.
Supabase — automatic embeddings with pgvector, queues, Edge Functions and scheduled processing →
PERMISSIONS
Give the automation the smallest amount of authority it actually needs.
Read vs write
A workflow that only needs to read customer data should not automatically receive permission to delete it.
Tool-specific access
Expose only the actions required by the workflow or agent.
Environment
Separate testing from production where the consequences justify it.
Secrets
API keys and credentials belong in secure connection or secret management, not inside prompts.
Audit
Important side effects should be traceable to the workflow, tool call or user approval that caused them.
The problem is not that an agent can use tools. The problem is giving it tools without designing the boundaries.
AUTOMATION COST
Automation cost is more than the subscription price of n8n or an AI model.
Build
Mapping, integration and implementation.
Execution
Workflow runs, API calls and infrastructure.
Model
Tokens, embeddings, image / audio processing or other AI consumption.
Data
Database, storage, queues and retrieval.
Maintenance
APIs change. Business rules change. Workflows need ownership.
Failure
The cost of duplicates, missed actions or incorrect automation.
The cheapest workflow is not cheap if nobody can safely maintain it.
OJC LABS PROCESS
We automate the process in the order it actually exists.
Observe
Map how the work happens today.
Define
Identify triggers, inputs, rules, actions and exceptions.
Choose the source of truth
Decide which system owns each important record.
Choose the execution model
Workflow, AI-assisted workflow, agent or RPA.
Connect
Implement APIs, webhooks, databases and required tools.
Control
Add validation, permissions, approvals and failure handling.
Test
Run normal, duplicate, missing-data and failure scenarios.
Operate
Monitor the system, document ownership and update it as the business changes.
PROJECT INPUTS
You do not need to arrive with an n8n diagram.
Start with the work your team is doing repeatedly.
Process
What happens today?
Tools
Which applications contain the data or actions involved?
People
Who owns the decisions and exceptions?
Problem
What is slow, duplicated, inconsistent or difficult to track?
We translate the operating reality into the automation architecture.
CONNECTED SYSTEMS
Automation is usually the layer between systems, not a system in isolation.
Email & CRM
Connect lead records, lifecycle stages, tasks and customer communication.
Explore Email & CRM →Messaging
Connect conversations to qualification, support and follow-up.
Explore Messaging →Content Operating Systems
Connect research, preparation, review, publishing and updating.
Explore Content Operating Systems →Ads & Tracking
Connect conversions and downstream business outcomes to acquisition.
Explore Ads & Tracking →Web & CMS
Use forms, content and application events as workflow inputs.
Explore Web & CMS →E-commerce
Connect catalogue, orders, payments, customers and fulfilment.
Explore E-commerce →FIELD NOTES
Read the decisions behind the automation.
AI Agents vs Automation Workflows: What Actually Works
Choose model reasoning or deterministic workflow logic according to the task.
Read the guide →Best n8n Workflows for Business Automation
Identify repeatable processes with useful triggers, actions and exception paths.
Read the guide →How to Build an AI Automation System
Structure inputs, outputs, validation and human review around an AI step.
Read the guide →How to Build an AI Content Pipeline
Plan the stages between content intake, AI processing, review and publishing, including recovery when a step fails.
Read the guide →Automating Blog Publishing with APIs
Connect approved content to your CMS through supported interfaces while preserving metadata and editorial checks.
Read the guide →Internal Linking SEO Strategies That Actually Work
Connect enrichment and publishing automation to the editorial and SEO layer around it.
Read the guide →OFFICIAL DOCUMENTATION
What the tools actually support.
FAQ
Questions before automating a business process.
What is business automation?+
Business automation uses software to move information, apply rules and perform repeatable actions between systems with less manual work. It can include deterministic workflows, AI-assisted steps, agents or interface automation depending on the process.
What is workflow automation?+
Workflow automation follows a defined process from a trigger through rules and actions. It works best when the path is predictable enough to describe in advance.
What is an AI agent?+
An AI agent is a model configured with instructions and permitted tools so it can decide which actions to use for a task. It still needs boundaries, state, permissions and failure handling around those actions.
Is an AI agent better than n8n automation?+
They solve different problems. n8n can run deterministic workflows and can also orchestrate AI steps or agent tools. If the process can be described reliably with rules, an agent may add unnecessary cost and uncertainty.
What is the difference between an AI agent and an AI workflow?+
In a workflow, the developer largely determines the path. In an agentic system, the model may decide which permitted tools or actions to use at runtime. Many production systems combine both.
Can AI agents connect to our CRM and business software?+
Yes when the relevant APIs, integrations or tools are available and the required permissions can be granted safely. The agent does not magically gain access to the systems; those capabilities must be explicitly connected.
Do we need several AI agents?+
Usually not. Multiple agents are useful when separate roles genuinely need different instructions, permissions, tools or evaluation. Creating several agents for work that could be handled by functions or one workflow only adds complexity.
Can an AI agent run the business without people?+
That is not how OJC Labs designs production automation. Some bounded processes can run without intervention, but important exceptions, permissions and high-impact decisions should retain an appropriate human owner.
Can OJC Labs automate a process that has no API?+
Sometimes. RPA or computer-based automation can interact with a user interface when an API is unavailable, but it is generally more fragile and should be used deliberately.
How much does business automation cost?+
It depends on the number of systems, integration methods, process complexity, AI usage, data architecture, permissions, exception handling and ongoing operational requirements. OJC Labs defines the architecture before estimating the implementation.
NEXT STEP
Start with one workflow worth improving.
Bring us the trigger, the current tools, the repeated actions and the exceptions your team handles today. We will determine what should remain deterministic, where AI can help and where a person should stay in control.