---
title: Why AI decisions drift
summary: >-
  AI systems drift when they keep making decisions after the original intent,
  evidence, and boundaries have become implicit.
agent_summary: >
  This page explains the major sources of AI decision drift and how JudgmentKit
  responds by making workflow decisions, evidence, and ownership explicit.
canonical_url: /docs/start/why-ai-decisions-drift
page_type: start
related_resources: []
related_schemas:
  - /schemas/decision-record.schema.json
  - /schemas/verdict.schema.json
last_reviewed: '2026-04-09'
---
# Why AI decisions drift

AI systems drift when they keep making decisions after the original intent, evidence, and boundaries have become implicit.

> Agent summary: This page explains the major sources of AI decision drift and how JudgmentKit responds by making workflow decisions, evidence, and ownership explicit.


## Headings
- ## Problem in human terms
- ## The main sources of drift
- ### Intent drift
- ### Evidence drift
- ### Boundary drift
- ### Ownership drift
- ## What JudgmentKit does about it
- ## Practical signal that drift is happening
- ## Related pages

## Problem in human terms

An AI system can look stable in a demo and still drift in production. The model changes, prompts accrete, retrieval sources get noisier, or a new surface changes what “good” should mean. The output still feels locally plausible, so teams miss the real problem: the system no longer knows which decision it is making.

## The main sources of drift

### Intent drift

The workflow started with one purpose, but the live system now optimizes for another. This often shows up as assistants sounding too sales-like in support contexts or generation tools over-indexing on novelty instead of system safety.

### Evidence drift

The system uses stale or mismatched evidence. It may cite the wrong account context, apply a generic pattern to a high-risk flow, or answer without the trace needed to justify the result.

### Boundary drift

Allowed variation widens over time until hard stops are no longer obvious. A team keeps shipping “small exceptions” until the exception becomes the norm.

### Ownership drift

No one can tell who owns the decision, the risk, or the runtime behavior. When incidents happen, the handoff path is incomplete.

## What JudgmentKit does about it

JudgmentKit reduces drift by forcing the system to answer five questions:

1. What decision is being made?
2. What does good judgment look like?
3. What evidence is allowed?
4. What are the hard stops?
5. Who owns the response when the system should not continue?

## Practical signal that drift is happening

- outputs sound more confident than the evidence supports
- generated UI no longer maps to system primitives
- users see inconsistent answers across surfaces
- latency and cost rise without a product reason
- escalation happens too late or without enough context

## Related pages

- /docs/start/what-is-judgmentkit
- /docs/how-it-works/decision-lifecycle
- /docs/guardrails/brand-and-tone

## Related pages
- /docs/start/what-is-judgmentkit
- /docs/how-it-works/decision-lifecycle
- /docs/guardrails/brand-and-tone

## Related schemas
- /schemas/decision-record.schema.json
- /schemas/verdict.schema.json