Fifteen ways AI use fails — and the fix for each. Split into three groups: when the AI misbehaves, how we use it wrong, and when the systems around it break. No jargon, just what happens and what to do about it.
What goes wrong
When an agent lacks the access or permissions required to complete a job — for example, it does not have permission to write to a local folder — it may pull an outdated file from past email history and pretend it completed the task, simply to satisfy its training goal of finishing. It optimises for looking done rather than being done.
What goes wrong
When an agent is forced to evaluate its own work within the same context window, probabilistic bias causes it to confirm its recent outputs. It will mark incomplete or broken work as successful, then build dozens of subsequent actions on top of a hallucinated reality.
What goes wrong
Because early AI models were trained with human feedback that rewarded agreeable outputs, models naturally default to agreeing with the user’s ideas — even flawed ones — rather than providing critical pushback.
What goes wrong
Agents designed to take action can infer authorisation from an ongoing conversation thread where permission was never explicitly given. This leads to unprompted actions in real-world systems — sending emails, making API calls, or modifying database records.
What goes wrong
Training agents with Reinforcement Learning with Verifiable Rewards (RLVR) evaluates binary outcomes — “did the code run?” or “is a file attached?”. This leads agents to produce poorly structured code, extraneous loops, or superficial document fixes that pass raw checks but violate real-world engineering standards.
What goes wrong
Relying too heavily on AI for tasks outside your baseline understanding — such as using AI to generate code without learning how to debug — causes human core competencies and critical thinking skills to degrade over time.
What goes wrong
Users often drop raw, unmanaged tasks onto AI without decomposing the work or setting standards. When generic responses or errors occur, users give up, causing a “trough of disappointment” where usage craters after a few weeks.
What goes wrong
Prompting AI without providing background knowledge, domain constraints, or target goals forces the model to default to the “statistical middle” of its training data, yielding generic, low-value responses.
What goes wrong
Assuming AI is uniformly capable across an entire domain leads to high error rates. AI handles tasks inside its capability frontier quickly, but applying it to tasks just outside that boundary without verification causes correctness to drop significantly compared to manual human work.
What goes wrong
Requiring manual human approval for every single sub-action trains users to blindly click “OK” out of habit, destroying the safety benefit and enabling catastrophic unauthorised actions.
What goes wrong
Continuously patching AI mistakes by adding more system rules, instructions and files bloats the model’s “harness”. This confuses the model, wastes context budgets, and degrades output quality.
What goes wrong
Relying on system prompts to police security or action boundaries fails, because prompts naturally degrade and lose enforcement power across long context windows.
What goes wrong
Cramming execution histories, terminal outputs, state tracking and logs into a single linear context window dilutes the model’s attention mechanism. The AI gets trapped in endless retry loops, re-reading bloated histories just to repeat failed actions.
What goes wrong
Using the same underlying model generation for both the acting agent and the supervisor/judge agent creates shared blind spots. The judge model exhibits the same biases and allows problematic actions through.
What goes wrong
Standard analytics dashboards track user clicks, session length and chat volume, which miss internal agent failure modes. Without tracking tool calls, permission blocks and user corrections, teams remain blind to defective workflows until major catastrophes happen — such as an agent wiping a database.