QubicQubıc
AI · Accounting

From 5 hours of manual classification daily — to 25 minutes of review.

24 employees · Istria · 4 months

An accounting firm was spending a full position on document classification. An AI solution took over 92% of the work — without layoffs, with growing capacity for new clients.

−92%

5h / day25 min

Document classification

Context

An accounting firm from Istria with 24 employees — serving around 180 small and mid-sized companies across the local market. Daily inflow of documentation is about 800–1,200 items — invoices, statements, contracts, expense claims.

One person in the firm was working exclusively on classification — sorting documents by client, type, and priority, and routing them to the responsible accountant. That's a full-time position that doesn't deliver direct value to the client — it's just a prerequisite for the actual work.

The owner was weighing two directions: hire an additional classifier to expand capacity (more clients), or find a way to make the existing process more efficient.

What didn't work first

They considered hiring another classifier to scale — which would have meant another salary on work that doesn't deliver client value. They considered outsourcing — which would have meant losing control over sensitive documents. Only the third option — automate what's actually structured — turned out to be the right one.

Approach

We started with AI & Automation — analysis showed classification was an ideal AI task: clearly defined, repetitive, and with historical data available for training (3 years of classified documents in the system).

We set up an AI classifier that reads incoming documents (PDFs, invoice images, email attachments), recognizes the document type, identifies the client, and routes it to the responsible person in the system. Accuracy on routine documents (75% of the volume) is above 95%. On more complex documents the system flags "needs review" rather than getting it wrong.

In parallel we worked on Training & Adaptation — workshops with the accountants and the classifier so they understood how the system works, when they can trust it, and when they need to step in. The classifier didn't lose her job — she took on a new role as "AI quality controller", reviewing borderline classifications and improving the model.

Delivered

  • AI document classifier integrated into the existing system.
  • Dashboard for tracking accuracy, escalations, and model improvement.
  • Pre-launch training for the whole team — understanding the system before it went live.
  • 30 days of practical support after launch.
  • Roadmap for expansion into adjacent processes (invoice processing, automatic statement ingestion).

Result

  • 92% of documents are classified automatically — routed to the responsible person in seconds, not hours.
  • 40 hours freed per week — equivalent to one full-time employee. The owner used that capacity to take on 12 new clients in the first year.
  • The classifier, who had previously been in "low-value" admin work, now has a role that's more interesting and technically demanding. No layoffs, and a real lift in the value of the role.
  • Classification accuracy keeps improving — the borderline cases the classifier manually reviews flow back into the model as additional training data.

At a glance

  • 92%

    Documents auto-classified

  • +40

    Hours freed per week

  • +12

    New clients in year one

I was afraid of layoffs. AI didn't eliminate the roles — it changed what they spend their time on.
— Managing partner, accounting firm

INTERFACE EXAMPLE

This is what AI document processing looks like in practice — the console we built for an accounting firm with 24 employees.

AI CONSOLE ILLUSTRATION · ANONYMIZED

What we learned

AI adoption works when it doesn't replace people but gives them roles where they're more valuable. Classification is boring work; reviewing and improving the system is interesting. The difference in motivation shows up within the first few weeks.

Client identity anonymized per contractual discretion. Numbers and context are accurate.