AI Inside and Out: New Guardrails for When AI Is Both Your Tool and Your User

AI is changing our systems from two directions simultaneously - and most teams are only watching one.

From the outside, AI-powered consumers interact with platforms in ways humans never did: relentless parallel requests, creative exploitation of undocumented behaviors, and usage patterns that make traditional load tests irrelevant. Your stress tests assumed human patience. AI has none.

From the inside, AI is now embedded in development workflows - writing code, finding bugs, generating fixes. Engineers are shipping faster, but the failure modes are new: sycophantic overconfidence in AI outputs, quantity-over-quality optimization, dissolved ownership of generated artifacts, and review pipelines drowning under machine-rate output.

This talk explores the guardrails required when AI operates on both sides of your platform, organized across three dimensions:

• People - How trust calibration evolves when AI is your collaborator. We'll share findings on how engineers relate to AI-generated work: the ones who trust it too much, the ones who reject it entirely, and the behavioral shifts that actually improved outcomes.
• Process - How development workflows change when output exceeds human review capacity. We'll cover blind validation techniques, two-stage sign-off gates (where AI validates against evidence before SMEs reconcile), and the process redesigns that prevent reviewer burnout while maintaining quality bars.
• Technology - How to improve AI confidence at scale using structured knowledge bases, multi-model disagreement detection (not consensus - disagreement is the signal), evidence-grounded scoring across four dimensions, explicit abstention when confidence is low, and staleness-aware systems that know when their own knowledge is out of date.

The through-line: traditional engineering assumed a human author and a human user. When AI is both, every assumption about quality, resilience, and trust needs to be re-examined - and the guardrails you need aren't just technical. They're psychological and procedural too.