PATTERN LIBRARY

Failure Modes in
Practical AI Systems

A public library of recurring failures that make AI hard to trust, hard to operate, or hard to adopt in real teams. These patterns shape how we think about workflow systems, governance, and decision support.

01

Stateless Amnesia

The system forgets prior context between sessions, forcing users to rebuild operational knowledge over and over.

02

Tool Sprawl

Too many disconnected tools, agents, and dashboards create more coordination burden than leverage.

03

Unverifiable Autonomy

Agents take meaningful actions without tamper-evident records, so teams cannot prove what happened.

04

Prompt-Only Governance

Safety depends on instructions alone instead of structural constraints, review gates, or enforceable boundaries.

05

Over-interruption

AI assistance that breaks real work instead of helping a team stay in flow.

06

Cognitive Overload

Systems that produce more context, alerts, and options than a user can actually act on.

07

Silent Failure

Outputs that look plausible even when the system is wrong, incomplete, or misaligned with the operating context.

08

Credential Trust Collapse

Trust is granted to whoever holds a token, without enough evidence about intent, behavior, or transaction context.