AutomationSTAR

Test Automation Conference Europe

Deep Dive DD4

AI Metrics: When Numbers Hide Real Risk

Maryna Didkovska

09:45-11:45 CET, Thursday 5th November

AI agents are already in production – making decisions, calling APIs, and changing real systems.

Yet many teams still judge their quality using metrics designed for predictable, deterministic software. Accuracy scores and static thresholds are applied to systems that reason and act autonomously, creating a quiet but dangerous illusion of control. In several production reviews I’ve been part of, the most damaging incidents happened when all reported metrics looked healthy. Systems showed 90%+ accuracy, stable dashboards, and green CI/CD gates – while quietly drifting into unsafe behavior, deleting critical data, or reinforcing the wrong decisions over time.

These failures were not invisible—we were simply looking in the wrong place.

This Deep Dive challenges an assumption many teams don’t even realize they are making: that if something is measurable, it is controlled.
Using production-inspired scenarios, we explore why traditional automation and AI quality metrics fail when applied to autonomous agents—and how those blind spots turn into real production decisions.

One recurring issue is metric illusion. Accuracy often rewards inaction, and stability can hide real damage, especially when agents operate unattended. Another problem emerges over time. AI agents rarely fail suddenly—they degrade gradually. Through week-by-week timelines, we examine how systems cross safety boundaries long before dashboards turn red.

The hardest part is decision-driven control. Instead of asking “Is the metric good?”, we focus on how combinations of signals should drive concrete actions: when an agent must stop, when it should escalate to a human, and when it is safe to continue.

Participants will work through realistic and uncomfortable scenarios from healthcare, finance, and enterprise systems.

They will decide whether to let an agent run overnight, whether to release to production, and which metrics create false confidence.

The goal is not more metrics, but better decisions before an agent causes damage.

BACK TO PROGRAMME

What you will Learn

  1. How to connect AI and automation metrics to concrete production decisions - when to stop an agent, when to escalate to human, and when it is safe to continue.
  2. How to read metrics over time rather than as snapshots, spotting early warning signs of agent drift that stable dashboards and CI/CD gates can miss.
  3. How to identify common 'metric illusions' (high accuracy masking added value, stable metrics hiding damage, green dashboards creating false confidence) and challenge them before incidents occur

Session Details

  • Intermediate
  • 90 minutes
  • Hands-on workshop
  • AI-Augmented Automation (Gen AI)
Buy Conference Ticket

Check out this video from Maryna Didkovska to hear more about her talk on AI Metrics: When Numbers Hide Real Risk. We look forward to welcoming you to AutomationSTAR 2026.

Session Speaker

Maryna Didkovska

Senior Director- EPAM, Ukraine

Maryna Didkovska is a Senior Director, Quality Architecture at EPAM Systems. As a Test Manager and Quality Architect Maryna has managed teams to build scalable testing process on a projects with over 600 participants, develop testing ecosystems on Enterprise level and coached others be experts and leaders. Having an academia background in AI domain, she leverages theory with hands-on experience to interact with teams to improve software quality and drive digital transformations. Maryna is a regular speaker at conferences such as STAREAST, EuroSTAR, SEETEST, STARCANADA, ProQuality etc, covering a variety of testing topics.

part of the