In the past 18 months we've integrated AI models into dozens of business flows, from the simplest (document classification) to the most complex (production-line anomaly detection). Four projects deserve telling for lessons that hold across verticals.
Case 1 — Invoice classification at an accounting firm
Problem: 4 staff each spent 2-3 hours daily classifying inbound e-invoices into accounting categories. Frequent errors, growing backlog.
Solution: AI model that, given supplier + description + amount, suggests the category with a confidence level. Above 90%, automatic. Below, human escalation.
Results at 6 months: 84% invoices auto-classified, human time down from 10 to 1.5 daily hours total, classification errors from 3% to 0.4%.
Case 2 — Product description generation for grocery e-commerce
Problem: 18,000-product catalog, inconsistent descriptions, many missing. A 2-person team couldn't keep up.
Solution: AI pipeline generating description + bullets + SEO meta from: product name, category, attributes, image. Output always human-reviewed before publishing.
Results: time per product from 12 minutes to 90 seconds, catalog coverage from 40% to 98% in 4 months, +12% organic traffic on product pages.
Case 3 — Anomaly detection on a food production line
Problem: line with 8 unplanned stops per month, only found after they happened.
Solution: streaming AI model analyzing vibration, temperature and power draw, flagging predictive anomalies 4-6 hours ahead. Trained on historical line data via our Badil MES.
Results: unplanned stops down to 2/month (-75%), €35k/year recovered production, lower operator safety risk.
Case 4 — WhatsApp customer-care routing across brands
Problem: customer-care hub handling WhatsApp for 6 brands. Operators spent lots of time figuring out "which brand is this for" and pulling customer context.
Solution: AI that, given an incoming message, identifies the brand, customer intent (purchase, complaint, info, password reset…), and drafts a reply based on FAQ + customer history.
Results: avg conversation handle-time from 8.5 to 3.4 minutes, higher customer satisfaction (faster replies), operators report better experience.
What we learned (true for any AI project)
- AI works when it integrates into an existing flow, not when it's a separate app. UX of integration is often the success/fail line.
- Keeping humans in the loop in the first months is non-negotiable: builds trust, gathers feedback to improve the model.
- Gains come from cutting low-value work, not "replacing" people. All clients above kept the same headcount, moved to higher-value work.
- Measure before and after: AI projects without baseline KPIs vanish after 12 months because nobody knows if they worked.
Have a repetitive flow to automate with AI? Let's talk.
Get in touch
