AI automation
AI process automation for businesses: invoices, emails, reports
We connect your email, documents, accounting and CRM, and let AI read what is unstructured. We start with the process that takes the most hours and measure the result.
What to automate first, and how to know it paid off
We automate repetitive work in businesses: incoming invoices, email, reports, data extraction from documents and CRM updates. AI steps in where the data is unstructured, such as a supplier's PDF or a free-form email, and the rest of the workflow runs on clear rules. A single automated workflow costs €400–1,200, and a pipeline that connects several systems starts at €1,500.
Which processes to automate first
Pick a process that repeats often, follows the same steps, has accessible data and does little harm if one case goes wrong and is caught at review. Some good candidates:
- Incoming invoices. Invoices from RO e-Factura arrive as structured XML files, so they can be read without AI and sent straight to accounting or an expense sheet. AI helps with the rest: PDF invoices from foreign suppliers, receipts, scanned documents. Background is in our e-Factura guide.
- Email. Messages to your contact address are classified (quote request, complaint, invoice, spam), routed to the right person and, where it makes sense, given a draft reply that someone approves.
- Reports. Shop sales, calendar bookings or CRM leads land in a spreadsheet and a short email every Monday morning, with no copying by hand.
- Document extraction. Contracts, quotes or scanned forms become fields in a table: parties, amounts, deadlines. Fields that fail the checks, such as a total that does not match the sum of the lines, go to a person.
- CRM updates. A new form submission or a customer email creates or updates the contact and the task in your CRM, with a summary of the conversation.
How to measure whether it pays
The sum is simple. Take the hours saved per month, multiply by the full cost of an hour of work, then subtract the monthly running cost (subscriptions, AI model usage, hosting), the time spent on review and the maintenance. Before we build, we measure for two weeks how long the process takes today and how many cases come in. Without that baseline you cannot tell after launch whether you gained anything. A worked example is in our article on AI agents ROI.
A person stays in the loop
Any step that moves money, sends something on the company's behalf or decides something about a person goes through a human. Article 22 of the GDPR gives people the right not to be subject to a decision based solely on automated processing that significantly affects them. Automated CV screening is also high-risk under the AI Act, with duties that apply from 2 December 2027, as we explain in our AI Act guide. Every workflow keeps a log: what came in, what it decided, who approved. When something fails you get an alert and the case stays in a queue to be retried.
n8n, Make or custom code?
| Tool | When it fits | What to watch |
|---|---|---|
| Make | Simple workflows between cloud apps, quick start | You pay by volume of operations, and data passes through their servers |
| n8n | More complex workflows, with the option to host on your own server | Self-hosting needs updates and backups; read the licence before commercial use |
| Custom code (Python, TypeScript) | High volumes, specific logic, data that must not leave the company | Higher upfront cost, no per-operation fee |
We pick the tool to fit the process. A workflow can start in n8n or Make and move to code once volume grows.
What we automate
Fewer copy-paste hours, with checks built in
Each workflow is measured before and after, and keeps a person where judgement matters.
Document extraction
PDF invoices, receipts, contracts and scanned forms become fields in a table, with amount and date checks before anything enters your systems.
Email triage
Messages are classified and routed to the right person, with draft replies that only go out after approval.
Automatic reports
Sales, bookings or leads pulled from several sources into one sheet and a short email, at the time you choose.
CRM kept up to date
Contacts, tasks and conversation summaries created from forms and emails, without retyping.
n8n, Make or code
We choose the tool by volume, budget and where your data is allowed to live, and the documentation stays with you.
Human review and logs
Steps that move money or affect people need approval. Every run leaves a trace, and errors trigger alerts.
How it works
How an automation project runs
- 1
Map and measure
We write the process down step by step and measure for two weeks how long it takes and how many cases arrive. That becomes the ROI baseline.
- 2
Prototype on real data
We run the workflow on 50–100 past cases and compare the output with what a person did. Together we set the accuracy threshold.
- 3
Go live with review
The workflow goes into production with human approval on sensitive steps, alerts on errors and a queue for uncertain cases.
- 4
Handover and a one-month check
You get the documentation and all access. After the first month we compare hours and costs with the baseline.
Budget
Automation prices
Each workflow is quoted in writing after we map the process; the bands are indicative and negotiable. Tool subscriptions, AI model usage and hosting are paid to those providers.
Single workflow
€400–1,200per workflow
One process automated end to end, the usual place to start.
- One process, two or three apps connected
- AI for text or documents where needed
- Human approval and error alerts
- Documentation and handover
Multi-system pipeline
from €1,500one-off
Several systems working as one, for processes that cross departments.
- Email, accounting, CRM and shop connected
- Queues, retries and a full log
- Hosting on your own server on request
- Dashboard of runs and errors
AI extraction feasibility
€1,200–2,500one-off
For when it is unclear how well AI can read your documents.
- Evaluation set from your real documents
- Accuracy measured field by field
- Cost estimate per document
Automations we run
The Danube Labs website runs real automations: an external cron job checks Google Calendar bookings every 5 minutes and sends the confirmation, a reminder 24 hours before and an email when the meeting starts, while each booking's state lives in the event's private properties, with no database. My Claude Code routes requests to AI models with fallback chains, key rotation and the cost of every request logged. AI Transcription Studio, in development, has a job queue that takes whole folders, skips files already processed and isolates failures per file, the same principles we apply to any document workflow. More in the portfolio.
FAQ
Questions, answered
How much does an AI automation cost?
A single workflow costs €400–1,200, depending on how many apps it connects and whether AI has to read documents. A pipeline across several systems starts at €1,500. If it is unclear how well AI can read your documents, a feasibility study with an evaluation set costs €1,200–2,500. Running costs, such as tool subscriptions, model usage and hosting, are paid to the providers.
Which process should I automate first?
One that repeats often, follows the same steps each time, has data you can access and does little harm if a case is wrong and caught at review. Supplier invoices, sorting contact-form emails and a weekly sales report are typical first picks. Avoid starting with a process that changes every month or depends mostly on someone's judgement.
Should I use n8n, Make or custom code?
Make suits simple workflows between cloud apps and a quick start, priced by volume of operations. n8n suits more complex workflows and can run on your own server, which means someone handles updates and backups. Custom code suits high volumes, specific logic or data that must stay in the company. We recommend one after mapping the process, and the workflow is documented whichever you choose.
What happens when the AI gets something wrong?
The workflow is built to expect it. Extracted data passes checks, such as totals matching line items and required fields present, and anything that fails goes to a review queue. Steps that pay, send or decide about people need a human approval. Every run is logged, so a wrong result can be traced, corrected and used to improve the instructions.
Is my company data safe in an AI automation?
We send the model only the fields it needs, use providers under API or business terms that exclude training by default, sign processing agreements where the GDPR requires them and keep keys in accounts you own. For data that must not leave the company, the workflow can run on your server or with local models. Our cybersecurity basics guide covers the rest.
Can I automate invoices from e-Factura?
Yes. Invoices received through RO e-Factura are structured XML files, so they can be read reliably without AI and sent to your accounting software or an expense sheet. AI is useful for documents outside the system, such as PDF invoices from foreign suppliers or scanned receipts. Agree with your accountant who downloads what and where it ends up before we build.
How long does it take to build an automation?
A single workflow usually takes 1–3 weeks of work, and the two-week baseline measurement can run alongside the prototype. A multi-system pipeline usually takes 4–8 weeks, depending on how quickly we get access to each system and on the quality of the data. The timeline in the quote is tied to receiving that access.
Who maintains the automation after handover?
Apps change their APIs and tools release updates, so automations need occasional care. Alerts tell you when a run fails. We do maintenance on request, with an estimate first, and the documentation and accounts in your name mean anyone competent can take over. Workflows that use AI models should be retested when the provider retires a model.
Tell us which process eats your hours
In a free call we pick the right process to start with and tell you what it would cost and what you could save.