Artificial narrow intelligence (ANI) is an informal term for AI built to do one bounded job, such as scoring payments for unusual activity, in contrast with the still-hypothetical systems that would match people across many tasks. Financial AI systems described in public research, such as payment fraud screens, fit the narrow description. A Bank of Canada team reported in 2024 that a two-layer machine learning screen reached a 93% detection rate on test payments. This page is general information, not financial advice.
Key takeaways
- ANI is not a formal technical category; regulators such as the Canadian Securities Administrators regulate the activity an AI system performs, not a label for the system.
- A narrow system performs well on the task and data it was built for and has no built-in ability to judge unrelated tasks.
- Bank of Canada researchers proposed a layered machine learning screen for large-value payments, tested on artificially manipulated transactions.
- The Ontario Securities Commission identifies three retail-investing use cases: decision support, automation, and scams and fraud.
- Narrow does not mean harmless: bias, poor data, and opaque models remain documented risks.
What "narrow" means and where the term comes from
No statute or standard defines artificial narrow intelligence. The phrase is common in popular and academic discussion to describe systems that handle a specific task, usually a prediction, classification, or recommendation, and nothing beyond it. The formal definition regulators use is broader. Canada's securities regulators, in Staff Notice 11-348, define an "artificial intelligence system" by pointing to the OECD's definition, and the notice does not create a separate category for narrow systems.
Researchers led by Meredith Ringel Morris, in a 2023 paper on levels of artificial general intelligence (revised September 24, 2025), propose classifying systems by the depth of their performance and the breadth of their generality. That framing makes "narrow" a matter of degree, not a bright line. A spam filter and a large language model both rank as narrower than a hypothetical general system, although the language model handles many more kinds of tasks than the filter.
One documented example: screening large payments
In May 2024, Bank of Canada researchers Ajit Desai, Anneke Kosse, and Jacob Sharples published Staff Working Paper 2024-15. It proposes a real-time monitoring framework for Canada's high-value payment system. A supervised model first separates typical payments from unusual ones, and only the unusual ones go to an unsupervised anomaly-detection model.
The authors tested it on data from the Canadian system in which some transactions had been manipulated artificially. The first layer achieved a 93% detection rate, and the second layer scored the manipulated transactions as nearly twice as suspicious as the originals. The framework is aimed at system operators and overseers trying to catch anomalies from cyber attacks or operational outages. It is a research proposal tested on artificial manipulations, not a report on a deployed production system.
This is a clean illustration of narrowness. The model knows the shape of payment data, and it has no way to read a contract, understand a news story, or explain to a customer why a payment was held.
Narrow AI in retail investing
The Ontario Securities Commission, working with the consultancy BIT Canada, reviewed AI tools aimed at retail investors and grouped them into three uses. Decision support covers systems that provide recommendations to guide investment choices. Automation covers systems that manage portfolios or funds for retail investors. The third is scams and fraud, meaning AI used against investors.
The report lists benefits such as reduced cost and wider access to advice, and risks such as bias, herding when many investors rely on a few providers, poor data quality, and the opacity of black-box models. The CSA notice adds that securities laws are technology-neutral and that market participants stay responsible for the systems they use, with a degree of explainability appropriate to the setting.
Comparing narrow systems by task
The table pairs each narrow task with its built-in limit. The rows are generic categories drawn from the reports above, not product reviews.
| Task | What the narrow system does | What it cannot do |
|---|---|---|
| Payment screening | Flags payments that differ from typical patterns | Decide on its own whether a flagged payment is an attack, a mistake, or legitimate |
| Investor decision support | Produces recommendations from the data it is given | Account for facts missing from that data, or guarantee outcomes |
| Portfolio automation | Applies a set strategy to a fund | Explain every step if the model is opaque |
| Scam-warning tools | Marks offers that look high-risk | Catch every scam, or avoid warning on legitimate offers |
Information, not advice
This page explains a term and summarizes published research. It does not recommend any AI service, product, or investing approach, and a system described as "AI-powered" carries no guarantee of accuracy or returns. The OSC report warns that false claims about "AI-enhanced" investments are a scam pattern in their own right. Anyone with a money decision to make can talk to a registered adviser and check a firm's registration with the provincial securities regulator or, in the United States, with the relevant federal and state regulators.
FAQ
Is every AI tool in use today narrow?
By common usage, yes: none of the sources reviewed describes a system that matches people across all tasks. How narrow a given tool is varies, because a language model handles more kinds of tasks than a payment screen does.
Are narrow systems regulated differently from general ones?
In Canadian securities law, the CSA states that existing laws apply to the activity being conducted, regardless of the technology. The notice creates no new legal requirements.
Does a 93% detection rate mean the system is 93% accurate?
No. A detection rate measures how many true cases were caught. It says nothing about false alarms, and the Bank of Canada test used artificially manipulated payments.
Who decides when a flagged payment is a problem?
In the Bank of Canada design, the output is meant for system operators and overseers, who investigate what the screen flags.
Related reading
- What artificial intelligence means in general terms
- AI fintech companies in Canada
- Limited memory AI and its limits
- How AI problems are framed
More in AI in finance.
Last verified: October 2026
Sources
- Bank of Canada, Desai, Kosse, and Sharples, "Finding a Needle in a Haystack: A Machine Learning Framework for Anomaly Detection in Payment Systems," Staff Working Paper 2024-15, May 2024: source page
- Canadian Securities Administrators, Staff Notice and Consultation 11-348, "Applicability of Canadian Securities Laws and the use of Artificial Intelligence Systems in Capital Markets," December 5, 2024
- OSC research with BIT Canada on AI use cases for retail investors ("Use Cases and Experimental Research"), 2024
- Morris et al., "Levels of AGI for Operationalizing Progress on the Path to AGI," arXiv:2311.02462, 2023 (revised 2025): source page