Artificial intelligence (AI) is the field of building machine-based systems that use models to infer outputs from inputs in pursuit of objectives. Those outputs can be predictions, generated content, recommendations or decisions. In finance, that could mean extracting information from an issuer filing, identifying unusual transactions or supporting a portfolio decision. Each task needs its own evidence of reliability; the AI label supplies none.
The OECD’s technical explanation includes both learned models and symbolic or knowledge-based approaches. AI therefore extends beyond machine learning and neural networks. It also need not operate without supervision or keep learning after deployment. The definition is a reference for the OECD Recommendation; its application depends on context.
Four terms that answer different questions
| Term | What it describes | Finance example |
|---|---|---|
| AI | The broader family of systems and techniques, including learning and knowledge-based reasoning. | A system that interprets financial documents and proposes a structured summary. |
| Machine learning | Constructing a model from data through training, rather than specifying every response by hand. | Fitting a classification model to labelled historical transactions. |
| Inference | In deployed ML, applying a trained model to an input to produce an output. | Applying that model to a newly received transaction. |
| Forecasting | Predicting a future quantity or event using information available at the forecast date. | Estimating next month’s trading volume. |
The ML and inference meanings follow the official Google glossary; forecasting follows Hyndman and Athanasopoulos. A model can infer the language of an existing filing without forecasting anything. Conversely, a simple forecast can repeat the most recent observation without using a sophisticated learned model. Inference here means model use; in statistics, the same word also describes drawing conclusions about an underlying population or model from data.
Knowledge-based AI represents relationships explicitly and reasons with them. Such representations can be specified by people or learned. This does not mean that every spreadsheet condition, calculation or software branch is automatically AI. Describe the mechanism and the task before choosing a label.
A loan file: where the model stops and the rule starts
Consider a fictional document-assistance workflow. A trained document model reads an application and proposes structured fields, with references to the source documents. A separate, explicitly programmed check then tests whether the required fields are present. A reviewer handles omissions and verifies the extracted contents.
For this example only, the schema requires four fields: applicant name, requested amount, term and income evidence. The proposed record contains the first three; income evidence is absent. Comparing the required field names with the populated field names finds one missing field. The workflow requests that evidence before further assessment.
Original fictional workflow. Field presence is a schema check; extracted values still require verification. The model stage and the explicit validation rule do different jobs.
That count is deterministic data validation. It neither measures the applicant’s creditworthiness nor demonstrates that the presence check itself qualifies as AI. The broader hypothetical workflow contains a trained model; it also contains ordinary software rules and human review. Calling the whole workflow AI-assisted must not hide these different responsibilities.
Presence is also weaker than correctness. A model might extract the wrong amount from a document or mistake an unsupported statement for income evidence. Populating every field would make the record structurally complete, while leaving those errors unresolved. A missing value should not become an invented fact merely to satisfy the schema.
Judge the task, then the system
For a document model, test extraction accuracy against verified records and preserve the evidence needed to inspect mistakes. For a return forecaster, test later observations using only information available at each decision time. For a trading workflow, a good prediction still needs a position rule, execution assumptions and realistic transaction costs. These are distinct evaluation problems.
A human approval step defines who may act; it does not by itself establish model accuracy. Likewise, a deterministic rule can enforce a useful constraint without validating the model upstream. Evaluate the components and their combined behaviour, including what happens when inputs are incomplete or outside the conditions under which the model was tested.
For the learning part, The Supervised Learning Problem and the ML Pipeline develops labelled examples, objectives and held-out evaluation. It is an ML next step, with a public preview and membership access to the full lesson.
Sources
- OECD (2024), Explanatory memorandum on the updated OECD definition of an AI system, pp. 4, 6, 8–9: system definition, human involvement, learning and knowledge-based approaches, inference and output types.
- Google for Developers, Machine Learning Glossary, entries “machine learning”, “inference” and “model”: training versus applying a model.
- Hyndman, R. J. and Athanasopoulos, G., Forecasting: Principles and Practice, §1.2: forecasting versus goals and planning; §1.1 discusses simple forecasting methods.
- Géron, A. (2019), Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, second edition, Chapter 1, “What Is Machine Learning?” and “Why Use Machine Learning?”. Local text consulted; definition verified on PDF page 22. The application fields and workflow illustration are original teaching assumptions.