04 — Understanding Artificial Intelligence
Cross References: https://app.notion.com/p/3a84d77839538161956bda6bcd498723?pvs=21, https://app.notion.com/p/3a84d778395381f1bebddca90d35bac8?pvs=21, https://app.notion.com/p/3a84d778395381d98790dc4d6c42c0e0?pvs=21, https://app.notion.com/p/3a84d778395381d0843de2f020b7acb0?pvs=21 Number: 4 Owner: CSI Office of the CTO Review Date: July 26, 2027 Standalone Summary: AI is pattern-based software that produces probable outputs; it requires context, controls and verification. Status: Done Version: 1.1
Plain-language definition
Artificial intelligence is software that learns or applies patterns from examples and data to produce a useful output. Depending on the system, the output may be a classification, prediction, recommendation, extracted field, summary, draft, image, alert or conversation.
It does not “understand” a business in the same accountable way as an owner, manager, doctor, teacher, accountant or engineer. It operates within the information, instructions, tools and controls provided to it.
A simple operating model
flowchart TD
A["Business input"] --> B["AI system"]
C["Approved data and instructions"] --> B
B --> D["Proposed output"]
D --> E["Validation or business rule"]
E --> F["Human decision or controlled action"]
F --> G["Outcome measurement"]
Main business forms
| Form | Typical output | Example | Primary control |
|---|---|---|---|
| Recognition | Category or detected item | Identify defective product image | Known accuracy by defect type |
| Prediction | Probability or forecast | Estimate demand or equipment risk | Quality and relevance of historical data |
| Language assistance | Draft, summary or extracted information | Summarise service notes | Source grounding and review |
| Recommendation | Ranked options | Prioritise leads or maintenance cases | Fairness, criteria and override |
| Generative media | Text, image, audio or video | Create a first draft of training material | Rights, factual review and disclosure |
| AI-enabled automation | Suggested or executed action | Route a document after classification | Permissions, limits, logging and rollback |
Three essential truths
- AI output can be useful and still be wrong.
- A system that performs well in a demonstration may fail on local data or exceptional cases.
- Human review is a control only when the reviewer has time, authority, skill and the source evidence needed to detect an error.