What this looks like in practice

The situation

A Czech regional logistics company operating three warehouses was processing 800–1,200 supplier invoices per month. Each invoice required manual cross-referencing against delivery records in their ERP system — a process that took two full-time staff members roughly 120 hours per month combined.

The error rate on manual matches was around 4%, which translated to frequent disputes with suppliers and delayed payment cycles. The finance team was aware of the problem but had no clear path to fixing it without replacing their existing ERP, which they were not prepared to do.

What we built

A document-processing pipeline that extracts structured data from PDF invoices (including handwritten and scanned variants), matches it against ERP records using a combination of exact and fuzzy matching, and flags exceptions for human review.

The system integrates directly with their existing ERP via API. No replacement, no migration — the AI sits alongside the tools they already use and handles the matching layer.

The outcome (6 months post-deployment)

• Processing time reduced from 120 to 11 hours per month
• Automatic match rate: 98.4% of invoices (up from 0%)
• Manual review queue reduced to 1.6% of volume
• Supplier dispute rate dropped by approximately 70% over the following two quarters

Note: the company asked us not to disclose their name. The metrics above are from their internal reporting at the 6-month mark and have not been independently verified.

The situation

A Vienna-based B2B SaaS company with 40 employees was receiving 300–400 support tickets per week. Their three-person support team was handling everything manually — first response times averaged 14 hours, and approximately 60% of tickets were repetitive questions with well-documented answers.

Their customers operated in German, English, French, and Czech. The support team was not fluent in all four languages, which added translation overhead and increased the risk of miscommunication.

What we built

An AI agent that handles tier-1 support tickets end-to-end — reading the incoming ticket, identifying the issue type, retrieving the relevant answer from a structured knowledge base, drafting a response in the customer's language, and sending it without human review (for tickets above a defined confidence threshold).

Tickets below the confidence threshold are drafted and queued for human review, with the AI's reasoning visible to the support agent. This gave the team a way to catch errors and identify gaps in the knowledge base rather than just trusting the output blindly.

The outcome (4 months post-deployment)

• 64% of tickets resolved end-to-end by the AI agent without human review
• Average first response time reduced from 14 hours to under 4 hours
• Customer satisfaction score (CSAT) increased from 72 to 81 over the same period
• Support team redirected approximately 60% of their time to complex escalations and proactive outreach

Note: the company asked us not to disclose their name. Metrics are from their internal reporting. CSAT methodology was unchanged throughout the period.