Email inbox handling
AI reads messages and attachments, recognizes the type of case (inquiry, complaint, invoice), extracts the data and creates a task in your CRM or prepares a reply for approval.
02 / n8n · AI
Almost any repetitive process can be automated, even when the input data looks different every time. The process sets the rules, and AI handles the exceptions. We build it in n8n, on your server.
AI reads messages and attachments, recognizes the type of case (inquiry, complaint, invoice), extracts the data and creates a task in your CRM or prepares a reply for approval.
Data from PDFs and photos goes to your accounting system or Alfresco, with verification and an approval flow.
Automatic replies based on your company’s knowledge base. Difficult cases go to a person, with a description of the situation.
A customer inquiry turns into a calculation and a ready offer, with no manual number-crunching.
Data from many systems is combined into one summary, sent by email or messenger at a set time.
Watching prices, tenders and contract expiry dates: you are notified before it is too late.
Creating accounts, sending documents and reminders for new employees or customers.
Together we pick the process that gives the biggest return at the lowest risk, and start with it.
We describe what goes in, what should come out, what the exceptions are and where a human decision is needed.
We connect the systems you use: email, CRM, ERP, spreadsheets, messengers, databases and APIs.
Classification, data extraction, summarization, draft replies. With result validation and human approval where the risk is significant.
Failure notifications, logs, retrying failed steps. You know when something has stopped.
Installing n8n on your server, backups, updates and security.
A description of how the flows work and short training, so the team understands what is happening.
Further processes, changing requirements, new systems. We expand the automation step by step.
We build in n8n at the level of nodes, expressions and code, not just “click-together” flows. These are the elements we work with every day.
Triggers and Webhook
Starting a flow on an event: Webhook (test and production URLs), schedule (Cron), email, a change in a database.
IF / Switch (routers)
Branching by conditions: different paths for different types of messages or documents.
Merge and Loop Over Items
Joining branches and processing lists record by record or in batches.
Sub-workflows (Execute Workflow)
Splitting large processes into smaller, reusable ones.
Wait and Stop and Error
Waiting for an approval, a deadline or a reply, and a controlled stop with an error.
Code node (JavaScript / Python)
Custom logic where ready-made nodes are not enough: transformations, validations, calculations.
Expressions
{{ $json… }} expressions, references to earlier nodes, dates, formatting, helper functions.
Data model (items, JSON)
Correct field mapping, working with arrays and nested objects.
Binary data
Files, attachments, PDFs and images passing through a flow.
HTTP Request node
Any API: authentication, pagination, headers, retries and rate limits.
Custom and community nodes (TypeScript)
Custom nodes for systems that are not available out of the box.
Credentials
Secure storage of access data: OAuth2, API keys, tokens.
Email and document connectors
Gmail, Google Drive, Google Sheets, IMAP/SMTP, Telegram.
Databases
PostgreSQL, MySQL and others: reading, writing, synchronization.
Alfresco (REST / CMIS)
Fetching and saving documents and metadata in the repository.
Polling or webhook
Choosing how to receive data from a system that exposes it.
AI Agent and LLM chains
Classification, data extraction, draft replies, agents with tools.
Structured output
Model output in a defined JSON format, validated and ready for further processing.
Embeddings and vector store (RAG)
Answers based on your company’s documents.
Local models (Ollama)
When data cannot leave your infrastructure.
Human-in-the-loop
Human approval where an error is costly.
Error Trigger and Error workflow
Error notifications and fallback paths.
Retry on fail
Retrying failed steps with a delay, limits and handling of 429.
Queue mode (Redis + workers)
Scaling executions across multiple processes.
Git as the source of truth
Exporting and versioning flows, test and production environments.
Docker, nginx, HTTPS, backups
Installation on your server, updates, encryption key and recovery.
Pruning execution history
Controlling the amount of stored data (GDPR and disk space).
If you are trying to set this up yourself and got stuck on one of the points above, tell us where you are. We will point you in the right direction, or do it for you.
You describe how it works today, and we ask questions: where the data comes from, how many exceptions there are, who approves.
We build a working prototype on sample data so you can see the result before deciding on next steps.
We launch the process with human approval. Only when the results are stable do we increase autonomy.
We watch errors and answer quality, and refine the rules and prompts.
Once the first one works, we pick the next.
It does, which is why we do not leave it unchecked. We use rules and result validation, and where an error is costly, the result goes to a person first. We reduce oversight only once quality is confirmed.
That is exactly what AI is for: it copes with variable structure. In addition, the flow includes a fallback path: uncertain cases go to a person.
n8n runs on your server, so data does not leave your infrastructure. If we use an external model, we minimize or anonymize the data before sending it and use providers with a data processing agreement. Locally hosted models are also possible.
No. We install n8n on your server or on one you designate.
It depends on the number of processes and integrations. We usually start with a small pilot with a clear scope, so you can see the result before we plan more.