An enquiry reaches the right person
The type is recognised, the owner is found and a task is created with the details.
n8n automations for business
How much work does your team do between the programs you use?
Moving data, handling documents, sending notifications and compiling reports. We build n8n automations that connect your systems and carry out the repetitive actions according to your business rules.
n8n automations
We connect your tools into a working process.
An online store, CRM, ERP, email, documents or a task platform — we build the connections you need with n8n. We define when the process starts, which data it uses, which actions it performs and when a person has to approve.
What we automate
Possible uses
Every automation is designed around the process, data and systems of the specific business.
01Inputs
- Telegram
- Forms
- PDFs and scans
02Tools
- n8n
- LLM API
- RAG
- PostgreSQL
03Systems
- ClickUp
- ERP
- CRM
- Shopify
04Output
- Tasks
- Data in ERP
- Reports
Data exchange between systems
Records move between your programs by rule, without being entered twice.
Enquiry handling
An enquiry is recognised and routed to the person responsible.
Documents
Extracting, checking and recording information.
Orders and notifications
Actions on a new order or a status change.
Reports
From the business data you already have.
AI steps
Where the process needs to understand text, documents or voice.
n8n
What a workflow looks like
Actions on a new order
The data is checked, recorded in the other system and the team is notified.
A report that arrives by itself
On a schedule the workflow collects the data, counts it in code and sends the summary.
Sample workflows. The actual steps are defined by your systems and rules.
The approach
The model proposes. Code checks. A person approves.
Much of the operational work in any company is retyping: data from a contract into a spreadsheet, from a waybill into a warehouse system, from a customer email into a task for a colleague. Language models can take over the retyping, but not the responsibility for the result. That is why every system we build draws a clear line between what the model does, what the code checks and what a person approves.
Clear rules are handled by code. Models can help interpret unstructured inputs. Consequential actions follow an agreed approval step.
Technology
How we build them
The tools are chosen to fit the task, not the other way around. These are the ones we use most often.
Orchestration with n8n
Every process is a visible diagram of steps: trigger, check, model, write, notification. We host n8n on a server in the EU under your account, so the processes and data stay with you, and the diagram can be read by someone who doesn't write code.
RAG — answers from your documents
Manuals, contracts, product catalogs and internal policies are indexed in a vector database. The model answers only from the passages it finds and cites the source; if there is no passage, it says it doesn't know. Access follows the user's permissions.
Language models through business APIs
We work with OpenAI, Anthropic (Claude) and Google (Gemini) and choose the model for the task: a smaller, cheaper one for classification, a stronger one for complex documents. Where requirements are stricter, an open model runs on your own server.
Structured outputs and functions
The model doesn't return free text but data in a schema, which the code checks before anything is saved. When it needs to act — create a task or look up an order — it does so through a limited set of functions with clear permissions.
Integrations with your systems
Telegram, email, web forms, ClickUp, Shopify, ERP and CRM — via APIs and webhooks. Where there is no API, we use file exchange in an agreed format.
Data and monitoring
PostgreSQL for records and the log, pgvector for semantic search, error alerts and daily reconciliations. Model costs are tracked for every process.
Honestly
When AI is the wrong tool
When the rule is clear
Calculating VAT, a discount above a threshold, routing by postal code — that is a job for code. It is cheaper, gives the same result every time and can be verified.
When a mistake is irreversible and nobody sees it in time
Issuing a payment order, sending a legal document, a final decision on a complaint.
When the decision affects people
Candidate screening, employee evaluation, creditworthiness of individuals. Under the AI Act these are high-risk applications, and under GDPR a person has the right not to be subject to a fully automated decision that has legal or similarly significant effects on them.
When the volume is small
If manual processing takes a few minutes a week, the system won't pay for itself.
Control
Keep control
Confidence threshold and independent checks
Fields below the threshold go to review. The model's self-assessment isn't fully reliable, so we add rule-based checks: totals, formats, matches against existing records.
Human review
Where an action has consequences, the system shows the proposal and waits for confirmation.
Log
For every action we keep the source, the result, who approved it and which version of the model and instructions was running.
Rollback
Every system can fall back to rules or manual mode without losing data.
Test cases
A set of real cases is run on every change of model or instructions, before the change reaches production.
Separately, we reconcile reports against the source. Automations have errors that don't raise any message: a misordered step can multiply the number of records, and a second process can quietly take over the first one's input. That is why every system has a single entry point.
Process
Start with a bounded pilot
01
One process with a clear metric
We pick one process where manual work can be measured: number of documents, processing time, share of errors corrected. Without a baseline, there is no way to prove the system helps.
02
Pilot on real, anonymized data
We test on your real examples with personal data removed and compare the result with manual processing. At this stage we set the confidence thresholds and the fields that always go through a person.
03
Rollout with human review
The system proposes, the employee confirms, and every action goes into the log. Falling back to manual mode is planned from day one.
04
Scaling and monitoring
We move on to the next process only once the first one runs reliably. We track the share of manual corrections and rerun the test cases on every change of model or instructions.
Data
Data protection and compliance
No training on your data
We work through business APIs whose default terms exclude the use of inputs and outputs for model training. Where requirements are stricter, the model is installed on your own server.
Processing in the EU
The servers and databases are in the EU. If the model provider processes data outside the EEA, it does so only on a legal basis under GDPR — an adequacy decision or standard contractual clauses.
Minimal access
The model receives only the fields the task requires; unnecessary personal data is removed beforehand.
Encryption, retention, access log
Data is encrypted in transit and at rest, retention periods are set in advance, deletion is automatic, and every access is logged.
Accounts in your name
You are the data controller, and we are the processor under an agreement per Article 28 of GDPR.
The AI Act (Regulation (EU) 2024/1689) applies in stages. Since February 2, 2025, the prohibited practices and the AI literacy requirement have been in force; after the amendment by Regulation (EU) 2026/1744, the latter requires measures that support staff literacy, with no certificate and no prescribed level. From August 2, 2026, the transparency rules apply: people must be informed when they interact directly with an AI system. The rules for high-risk systems under Annex III, which include recruitment and employee evaluation, apply from December 2, 2027. Document processing and internal reports are not on that list unless they are used in one of the listed areas. How your specific system is classified is a question for your lawyer; we provide the technical description they need.
Experience
Examples from our work
Telegram, ClickUp and daily reports
Built for a company with three teams. The manager writes tasks in free-form Bulgarian in Telegram; the system reads them, shows a preview and waits for confirmation. The tasks then go into ClickUp and into each employee's private chat with status buttons, and a report arrives at the end of the day. The model is used in two places: to read the text and to describe numbers that have already been counted. Counting and assignment are done by code. As of today, the system runs end to end in a test environment, and the text is read with rules until the client switches on a model; the switch is prepared and reversible. The server and accounts are in the client's name. It also handles voice messages and is hosted on the client's server in the EU, under the client's account.
Explore the automation →Cash Runway
Our own product for online merchants who sell with cash on delivery (COD), currently in development. It models cash flow across the whole chain: ad spend, orders, share of delivered shipments, delay in the courier's payout, stock running out, payments to suppliers, fixed costs and VAT. It shows when the money will be in the account, not when the sale was made. It is a forecast for management decisions, not accounting software.
Learn more →Questions
Frequently asked questions
Does every step need AI?
No. Calculations and clear rules are usually better handled by conventional software.
Do you use our data to train AI models?
No. We work through business APIs whose default terms exclude the use of inputs and outputs for training — for OpenAI this has applied since March 1, 2023, and for Anthropic across all commercial products, unless the customer explicitly agrees otherwise. When requirements are stricter, an open-weight model can run on your own server, so the data never leaves your infrastructure.
Where is data processed in AI automation?
The processing location is set before launch and written into the contract. The servers and databases are in the EU, and with providers that offer it, the model itself can also process and store data in Europe. If data leaves the EEA, it does so only on a legal basis under GDPR, such as an adequacy decision or standard contractual clauses.
What happens if the AI model makes a mistake?
Models make mistakes, so the system is built to catch an error before it reaches a customer or accounting. Results below the confidence threshold, and those that fail the checks on totals, formats and numbers, go to an employee for review. Every action is in the log, so an error can be traced back to its source, and the system can fall back to manual mode.
Do we need our own server for an AI system?
Not necessarily. For most processes, a virtual server in the EU in your name and a business account with the model provider are enough. Your own server with a local model makes sense for especially sensitive data or when data must not leave your infrastructure, but it costs more to maintain and takes more setup.
Will it work with our tools?
We assess available interfaces and limitations before promising a connection.
Is AI automation compliant with GDPR and the AI Act?
We design it with a data processing agreement, minimal data, retention periods, an access log and a person in the loop for decisions with legal consequences for people. Under the AI Act, the transparency rules apply from August 2, 2026, and the rules for high-risk systems under Annex III from December 2, 2027; document processing and internal reports are generally not on that list. How your specific system is classified is a question for your lawyer, and we provide the technical documentation they need.
How much does AI automation of a process cost?
The price depends on the number and type of sources (PDFs, scanned documents, emails, systems with APIs), the number of integrations, the hosting and security requirements, and the volume. On top of that there are running costs for the server, licenses and the model provider's fees, which are billed by volume and are usually in your name. After the pilot we know the real volume and give a price based on it, not on assumptions.
Which task repeats every day?
Describe its input, required output and who handles it today.