Case study: a training centre’s back office with no manual data entry, from registers to invoices

Dabster Labs · Case study · Last updated 27 October 2026

A vocational training centre in Nicosia runs with a team of a few people, hundreds of enrolments a year, government-subsidised seminars, and a client base of professionals who must complete continuing education hours every year. The same back-office work was done by hand every day, by the same people. This is what we automated, what it measures, and how it was built. The client is not named at their request, and the third-party systems involved are described only in general terms.

The problem

• Watching competitors’ seminars and professional registers, copying records into spreadsheets.
• Answering every course registration form with the right instructions for the state subsidy platform.
• Entering participants twice: once in the state platform, once in the CRM.
• Tickets that stayed open for months after the person had already enrolled or the course had ended.
• Duplicate companies and contacts in the CRM, so the same person received the same email twice.
• Supplier invoices the accountant opened one by one to decide the transaction type, the expense account and the VAT code, then keyed in.

What we built

Six automations running on the client’s own server in the EU, with AI used only where judgement is needed. Every decision with a consequence, such as sending an email or posting to the accounts, goes through a person with one click.

1. Market watch robot

It signs in to public registers and seminar listings, reads the records, matches them against existing CRM contacts and creates the new ones. A monthly report goes out only when something has changed. First run: 4,389 records in 55 minutes with zero failures, 2,015 new contacts, 2,312 matches to existing ones, and 118 of 118 seminars linked to the right provider.

2. Registrations without a secretary

Every 10 minutes it reads the new forms from the website and replies with the right subsidy instructions, in the wording the office would use. Anything unusual becomes a callback for a human. Courses and participants sync hourly from the state platform into the CRM, and a daily reminder goes to everyone who has not completed their registration before the deadline.

3. Ticket lifecycle

Each morning it closes the tickets of people who enrolled or whose course has ended, handles cancellations, flags tickets that have stalled, and moves interested people who did not enrol to the next run of the same course. First run: 16 leads recycled and more than 22 tickets that had been open for months cleared.

4. A clean CRM

A one-time merge of 68 duplicate clusters took the company list from 1,982 to 1,912. Since then a daily detector finds duplicate contacts and a person approves each merge.

5. AI bookkeeper

Every supplier invoice is read, the supplier is recognised even when its name appears only in a logo, and the agent proposes the transaction type, the expense account and the VAT code according to Cyprus VAT rules and the company’s own habits. The accountant approves or corrects on one page with the PDF beside it. Each correction is stored per supplier, and after two identical approvals that supplier moves to a one-click lane.

Measured on 36 real invoices from 8 companies, with the accountant’s actual postings as the correct answer and before the agent had learned anything from corrections: VAT right on 31, expense account on 28, transaction type on 27, and the one-click lane 7 for 7. The production accounting system is never written by the agent. Every morning the backup is restored into a test copy, the agent’s knowledge is refreshed and 539 past invoices are replayed. If the success rate drops, the new knowledge is not deployed and the report says why.

6. Reports you read in 90 seconds

A morning digest with 10 checks on servers, security and the business, a weekly management brief, and a daily activity report per employee. Together they replace five or six manual checks a day.

Results

• More than 200 hours a month of manual work that is no longer done by hand, by the team’s own estimate.
• 198 enrolments since 1 May, with a recorded first contact for 100 of them. Of those, 35 came from the automated outreach.
• Every enrolment is credited to the channel that brought it: call, newsletter, email or chatbot. Management sees what works.
• No data leaves the EU: a local model reads the registers, a European AI provider reads the invoices, and everything runs on the client’s server.
• The accountant looks at one page and presses one button instead of opening every PDF.

How it was built

Python and n8n on the client’s server, a Dolibarr CRM, local and European AI models, email through their own SMTP. Each automation has a log, a report and an off switch. We started with a process audit, delivered in stages from April, and every piece went into production only after it had run alongside the person doing the job.

Why the numbers are the way they are

Record counts, match rates and accuracy scores come from the systems’ own logs and from evaluation runs against the accountant’s real postings. The hours figure is the team’s estimate, not a stopwatch, and we say so. Where a number could not be measured, we left it out rather than invent it.

What this looks like for you

If your team spends hours copying data between systems, answering questions that are always the same, or keying in invoices one at a time, this is the work we automate. It starts with a process audit and a fixed price in writing. See what we build, read what to automate first, or book a call.

See also: Google AI Mode names Dabster Labs an AI leader in Cyprus and Greece