Artificial Narrow Intelligence (ANI) — sometimes called "weak AI" — refers to AI systems designed to perform a specific task or narrow class of tasks at or above human level. Every AI system that exists today, including the most capable large language models, falls into this category.
## What makes AI "narrow"
ANI systems are not general reasoners. They excel within their training distribution and degrade outside it. A chess engine is superhuman at chess and useless at driving. An image classification model trained on radiology scans fails on dermatology. Even large language models — despite their apparent versatility — remain ANI: they predict tokens based on statistical patterns; they cannot learn from a conversation (without fine-tuning), maintain persistent beliefs, or reason reliably about novel physical situations.
The characteristic failure mode of ANI is **distribution shift**: performance collapses when inputs differ meaningfully from training data.
## ANI vs AGI
**Artificial General Intelligence (AGI)** would exhibit human-level (or above) performance across arbitrary cognitive tasks without task-specific training. No AGI system exists today, and serious disagreement exists in the research community about whether current architectural approaches (transformer-based LLMs) can scale to AGI or whether fundamentally different systems are required.
The business relevance: enterprise AI strategy should be built around ANI capabilities that exist today, not AGI capabilities that may or may not arrive this decade. Organisations that anchor their roadmap to "when AI becomes general" are deferring decisions they should be taking now.
## Practical implications
- **Tool selection**: match the right narrow system to each task. A specialised document-intelligence model will outperform a general-purpose LLM on invoice extraction. A purpose-built fraud model will outperform a general classifier on payment risk.
- **Evaluation**: ANI systems require narrow, task-specific benchmarks. General-purpose accuracy scores mislead — test on the actual distribution your system will encounter in production.
- **Risk management**: ANI brittleness at distribution boundaries means you need monitoring for data drift and out-of-distribution inputs. Deploy with a human-in-the-loop escalation path for the cases the ANI cannot handle.
## The 2026 landscape
Frontier models (GPT-4o, Claude 3.7, Gemini 1.5 Pro) are demonstrating multi-domain competence that blurs the ANI/AGI boundary at the surface level. They can write code, summarise contracts, analyse images, and hold complex conversations within a single inference call. But each capability remains narrow in the sense that performance degrades on inputs far from training distribution — they are very wide ANI, not AGI. The practical difference matters less than the monitoring implications: even the widest ANI systems need task-specific evaluation, not general-purpose benchmarks.
Where AIMenta applies this
Service lines where this concept becomes a deliverable for clients.
Beyond this term
Where this concept ships in practice.
Encyclopedia entries name the moving parts. The links below show where AIMenta turns these concepts into engagements — across service pillars, industry verticals, and Asian markets.
Other service pillars
By industry