Introduction

In the rapidly evolving landscape of technological innovation, artificial intelligence (AI) has emerged as a potential solution to Canada’s persistent productivity challenge (Billy-Ochieng’ et al., 2024). With productivity growth stagnating at just 0.2 per cent annually over the past decade (Caranci & Marple, 2024), and mounting economic pressures from potential trade disruptions, the need to enhance productivity has become increasingly urgent. Generative AI — artificial intelligence systems capable of creating new content, such as text, images, music or code — offers particularly promising opportunities for productivity gains, while simultaneously raising important questions about workforce adaptation.

AI is an umbrella term, used to describe a set of technologies able to perform tasks commonly associated with natural intelligence, such as identifying objects from visual data (vision) or processing natural language (speech) (Oxford University Press, 2023). Approaches vary across technologies, but a common throughline between them is that, generally, AI algorithms are built to be able to modify and refine the way that they work based on exposure to large amounts of data.

Its potential to boost productivity through task automation, process optimization and augmentation of human capabilities has led economists to recognize it as a general, purpose technology with significant economic, social and policy implications (Acemoglu, 2024a; Agrawal et al., 2019; Bick et al., 2024; Council of Economic Advisers, 2024). Due to AI’s expected impact on society, it has been described as a “Gutenberg moment,” likening its influence to that of the printing press (Nuño, 2023).

However, realizing these productivity gains requires significant implementation challenges to be addressed. Canada currently lags behind other G7 countries in AI adoption, with only 3.1 per cent of companies having adopted AI technologies by 2022, due in large part to infrastructure limitations and skills gaps (Pamma, 2024). This underscores the importance of understanding how generative AI might reshape labour market demands and identifying the skills needed to effectively harness these technologies.

Context and Approach

The potential impact of digital technologies on work started to receive significant attention in 2013 with a study by Frey and Osborne that found that 47 per cent of all occupations were at high risk of being replaced by computerization. The authors reasoned that, in essence, computerization is a form of automation. The process of automation, in turn, tends to replace lower-skilled occupations and augment higher-skilled occupations, raising concerns about increased unemployment and rising income inequality. Unlike past technological shifts that primarily affected manual labour, computerization threatens roles previously deemed difficult to automate, including routine-based service occupations in sectors like logistics, office support and certain service roles.

Later studies on the effects of computerization and machine learning — the study of an algorithm’s ability to improve performance on a given task without specific instructions — on work demonstrated that the impacts are more nuanced. A common practice in this approach is to view individual occupations as collections of tasks and assess which tasks could be transformed by technology (Acemoglu & Restrepo, 2022; Brynjolfsson et al., 2018; Moll et al., 2022). Studying the impact of machine learning (ML) on occupations, Brynjolfsson et al. (2018) utilize a task-based approach to analyze how technology can alter job functions within occupations. They used a crowdsourcing platform to obtain ratings for a series of job tasks within U.S. occupations, with regard to their suitability for machine learning on a given scale. Using these ratings, the authors then calculated a “suitability for machine learning” score for each task. This approach allows for a better understanding of how technology can reshape job responsibilities rather than merely replacing entire occupations. Applying it to the Canadian labour market, Frank and Frenette (2021) find that new technologies have shifted the nature of work toward more non-routine tasks between 2011 and 2018. Yet they also note that these changes were rather modest in scale.

Accelerated development and adoption of machine learning technologies have led to an increased focus on AI as a so-called general-purpose technology. This term is often used for innovations that have widespread applications across various occupations and industries.

Felten et al. (2021) applied the task-based approach to explore specifically how AI technologies might affect occupations in the United States. The increased focus on AI is tied to its role as a general-purpose technology (Acemoglu, 2024; Agrawal et al., 2019; Bick et al., 2024, Council of Economic Advisers, 2024). Examples of general-purpose technologies include the steam engine, electricity and microchips. Due to their applicability across the economy, general purpose technologies commonly cause significant economic disruption, thereby creating winners and losers (Trajtenberg, 2018). Using a crowdsourced survey, Felten et al. linked common AI applications, such as image recognition and language processing, to workplace abilities. They then used occupational data to determine the level of occupational exposure to AI technologies. This exposure measure is agnostic with regard to AI’s ultimate impact on a specific occupation. In some cases, high exposure can lead to automation; in other cases, an AI technology can complement a human worker.

More recently, the rise in popularity and availability of AI tools capable of generating content like text, audio or visuals — popularly referred to as “generative AI” — has led to renewed interest in the impact of technology on labour and skills. This is especially the case since generative AI’s capability to create new content allows it to take on a host of cognitive and creative tasks previously perceived as the prerogative of humans. As a consequence, occupations formerly thought of as immune from computerization may now also be experiencing some transformation (Gmyrek et al., 2023).

Focusing primarily on generative AI, Eloundou et al. (2023) measured occupational exposure to large language models (LLMs) — a type of machine learning model that is trained on large amounts of language data (see box 1). In particular, they assessed the potential time savings that generative AI tools could provide for various tasks and work activities within a given occupation. They found that approximately 19 per cent of the U.S. workforce might face significant disruption to their roles. Notably, the authors pointed out that, while generative AI can enhance productivity by streamlining tasks, it may also require workers to adopt new skills to remain relevant in a rapidly evolving job market.

Pizzinelli et al. (2023) determined both the level of exposure and complementarity of U.S. occupations to generative AI. Similar to Felten et al. (2021), the study applied an AI exposure index to assess the level of automation risk of different jobs due to generative AI based on task composition and automation potential. In addition, it determined complementarity — whether AI could enhance productivity in specific jobs — or substitution — where AI might displace jobs. One of the main findings is that higher-skilled occupations might experience more complementarity with generative AI, while lower-skilled jobs could face higher displacement risks.

From a policy standpoint, understanding how generative AI will change the demand for skills is particularly important for guiding people about to enter the labour market. Individuals currently in school or university need insight into which skills will gain or lose importance, so as to make informed career choices. In addition, assessing the impact of generative AI on occupations is crucial for policy interventions aimed at the current workforce, enabling support through retraining and upskilling.

Two recent studies analyze the potential impact of generative AI on Canada’s labour market. Using a combination of qualitative analysis, expert insights and case studies, Burt (2023) analyzed the effects of generative AI on routine and cognitive tasks. It found that the most significant occupational impact from the deployment of such tools appears to be on writing and programming skills. Their results show that the top 10 occupations where these skills predominate are either associated with STEM occupations, such as computer network technicians, software engineers and designers; or with cognitive occupations, such as journalists.

More recently, Mehdi and Morissette (2024) assessed the potential impact of AI on Canadian workers. Applying the methodology developed by Pizzinelli et al. (2023) to Canadian occupations, the study categorized workers into three groups based on their exposure to AI: those whose jobs may benefit from AI due to high complementarity, those at risk of having their tasks replaced by AI, and those less affected. The authors determined that around 31 per cent of Canadian employees, equivalent to 4.2 million workers, are in occupations that could be negatively impacted by AI.

Our study takes a novel approach to understanding generative AI’s potential impact on the Canadian workforce. We first analyze the automation risk of generative AI across skills and work activities, then examine the likely effect on Canadian occupations and industries. To do this, we introduce two methodological innovations.

First, we leverage the capabilities of one of the largest and most popular all-purpose generative AI tools (ChatGPT — an LLM-based chatbot by company OpenAI) to assess the automation risk of generative AI across different skills and work activities. Trained on vast amounts of data encompassing diverse fields, including labour market trends and technological advancements, tools like ChatGPT can synthesize and analyze information to provide a comprehensive assessment of generative AI’s capabilities. Specifically, recent research has demonstrated that LLMs can effectively analyze structured data formats when provided with appropriate frameworks (Jiang et al., 2023). We define “automation risk” as the technical feasibility of a generative AI system replacing or significantly transforming a specific occupational skill or work activity. Importantly, our analysis focuses solely on technical capabilities, without considering potential implementation barriers such as organizational, regulatory or financial constraints.

Second, our methodology draws on the Occupational and Skills Information System (OaSIS), a comprehensive database developed by Employment and Social Development Canada that provides detailed information on skills, abilities and competencies across nearly 900 Canadian occupations. The systematic, structured nature of OaSIS makes it particularly well-suited for AI-driven analysis, as it allows us to distill occupations into their major elements, which can then be individually evaluated by an AI chatbot.

LLMs offer distinct advantages in this assessment. They can systematically process standardized task descriptions and skill requirements, maintaining analytical consistency across multiple occupational categories. This approach allows us to identify which tasks involve routine, rule-based actions that align with current AI capabilities, and which require complex human skills that remain difficult to automate.

The reliability of this approach is supported by emerging research showing LLM assessments to be in line with traditional expert evaluations. Eloundou et al. (2023) compared occupation ratings of automation risk done by humans familiar with LLMs to responses from ChatGPT (GPT 4) and found strong correlations between the two. While AI ratings were lower on average, responses from both sources have a very similar trend.

To our knowledge, this is the first study of its kind using OaSIS to examine the impact of generative AI on Canada’s labour market.

Methodology

We assess the impact of generative AI on Canada’s workforce by analyzing how it affects skills and work activities. In doing so, we determine the risk of automation for each skill and work activity related to Canadian occupations.

As noted earlier, there are two important caveats with our approach. First, the risk of automation from generative AI focuses solely on technical feasibility, consistent with other literature on the topic (Acemoglu & Restrepo, 2019; Eloundou et al., 2023; Frey & Osborne, 2017). As such, it does not take into account regulatory frameworks, cultural acceptance, and other factors that would influence the adoption of generative AI. For instance, while self-driving transport trucks may be technically feasible, public skepticism and regulatory hurdles could significantly delay or limit their deployment, reducing the actual impact on employment in that sector.

Secondly, we focus on changes over the next five years, as predictions beyond this time horizon likely have a higher degree of uncertainty given the rapid technological developments in this space. Throughout our analysis, it is important to keep in mind that adoption barriers may result in slower, more uneven implementation of generative AI across different sectors and regions.

We note also that the inclusion of both skills and work activities in our analysis is deliberate. The way in which OaSIS defines skills and work activities does not allow for a clear delineation between what might be an innate human ability (skill), and an action that might be more highly correlated with a job description (work activity). For example, Quality Control Testing is listed as a skill in OaSIS but could quite as easily be explained as a work activity. As such, any separation in analysis between work activities and skills would not yield a clearer or comparable analysis due to the ambiguity and fluidity of the way the items are defined. Furthermore, OaSIS organizes its descriptions into eight categories, five of which pertain to individual traits and requirements and three to the work environment. “Skills” is included from the individual characteristics and requirements category while “work activities” comes from the work environment; as such, including both in the analysis enables a more comprehensive analysis.

As it is our goal to determine the impact of generative AI on Canada’s workforce, we apply occupational information from the newly established OaSIS. Developed by Employment and Social Development Canada (ESDC) with support from Statistics Canada and the Labour Market Information Council (LMIC) in 2021-22, OaSIS is Canada’s database on occupations and associated competencies. The database was constructed using best practices from international examples such as the Occupational Information Network (O*NET) in the United States, and the European Skills, Competencies, Qualifications and Occupations (ESCO).^[1]

OaSIS links the existing Skills and Competencies Taxonomy to Canada’s National Occupational System (NOC). It provides detailed information on the skills, abilities, personal attributes, knowledge and interests needed for over 900 occupations in Canada. For our analysis, we extract data on the 34 unique skills and 41 unique work activities across all occupations.

For each work activity and skill, OaSIS provides a score, on a scale from 0 to 5, measuring the proficiency required for competency in a given occupation. The ratings are conducted by trained HR analysts. On this scale, 0 means that the competency does not apply to the occupation and 5 means that it is essential. We consider only those competencies most relevant for a given vocation. For this reason, we include only the subset of skills and work activities with a proficiency weight of 3 and higher for each occupation.^[2]

Applying the information on skills and work activities for each Canadian occupation provided by OaSIS allows us to determine automation risk scores in the following way.

We asked ChatGPT to rate how easy or difficult it would be for a specific skill and work activity to be performed by generative AI. When using a chatbot, results can differ based on three components. First, the output generated by a chatbot can depend on how a specific question is phrased. Second, results may depend on which particular model of chatbot is used. Finally, each model allows the user to specify certain parameters that influence the response, such as the level of randomness and the length of the generated output. To account for this, we use two different models of ChatGPT, vary the way we phrase our question by using different prompt structures, and modify specific parameters. Box 2 describes this approach in more detail, and further details on the prompt structure are provided in appendices B, C and D. Ultimately, we leverage these components to generate an average score of the risk of automation to provide a more comprehensive and robust estimate than if we were to simply report the output from one version of the question phrasing, one model or one set of inputs to the model parameters.

In total, we obtained 12 responses for the 34 unique skills and 41 unique work activities associated with all recorded occupations (listed in appendix E). This mimics the approach taken by Brynjolfsson et al. (2018) who surveyed 7 human experts across each of O*NET tasks. Next, we calculate the average automation risk scores for each work activity and skill which provides us with the potential automation risk from generative AI by skill and work activity for the over 900 occupations contained in the OaSIS.

The Impact of Generative AI on Occupations, Skills and Work Activities

Applying our methodology provides us with an average risk of automation score for each unique skill and work activity. As described, this score is averaged over 12 different prompts fed into the API, which varied in certain parameter values and vocabulary used, and across two different models of ChatGPT. This allows us to determine the likelihood of the risk of automation for each of the 900 occupations contained in OaSIS.

Our comprehensive analysis of generative AI’s automation risk for the Canadian workforce reveals three significant patterns in the data. First, clerical and data-processing skills and work activities demonstrate the highest automation risk from generative AI, with activities like entering and storing information scoring above 4 out of 5 on our risk scale. These are followed closely by activities involving operation monitoring, data analysis and scheduling, which all score above 3.5. Writing skills and activities also emerge as having a high risk of automation, aligning with recent empirical studies on generative AI’s capabilities in content creation and documentation (Burt, 2023).

Second, skills and work activities involving human interaction, social perception and instruction show markedly lower automation risk from generative AI. This pattern indicates that social, managerial and leadership skills will remain predominantly human domains in the near term, with generative AI having limited capability to automate these inherently interpersonal activities.

Third, our findings suggest that generative AI is more likely to transform the composition of skills and work activities within occupations rather than completely automate entire occupations. This is evidenced by our analysis of occupations representing 50 per cent of total Canadian employment in 2021, which show moderate automation risk scores between 2.77 and 3.3. For instance, retail salespersons demonstrate lower automation risk scores, particularly in activities requiring negotiation and persuasion, suggesting their roles will evolve to emphasize these human-centric skills while potentially seeing automation of more routine activities.

Overall, our analysis reveals distinct patterns of automation risk that vary significantly across skills, work activities and occupations. At the skill and work activity level, we find a clear divide between highly automatable activities such as data processing and monitoring, and more resilient human-centred skills such as instruction and social perception. At the occupational level, these risks manifest in three distinct patterns. Occupations requiring high levels of social interaction or manual skill show the lowest automation risk, suggesting that there will be minimal changes to their skill requirements and work activities. A second group of occupations shows high automation potential for certain activities but maintains important human-centred skill requirements, indicating likely shifts in work composition rather than a complete transformation. Finally, occupations centred primarily on routine information processing and monitoring activities show the highest automation risk, suggesting a more substantial transformation of their required skills and work activities.

Level of generative AI automation risk by skills and work activities

Table 1 shows the results for the skills and work activities with the highest risk of automation scores.^[3] In short, amid the skills and work activities listed in table 1, clerical activities — which include entering, transcribing or storing information — carry the highest automation risk from generative AI, with an average score of 4.29 out of 5. Skills and work activities related to monitoring, scheduling and data analysis also rate relatively high. Moreover, and in line with recent empirical studies, writing is among the top 5 skills/work activities with the highest automation risk.

Table 2 shows our results for the skills and work activities with the lowest automation risk scores. Instructing, meaning the capability to teach others knowledge, has the lowest overall risk score at 2.04, followed by social perceptiveness and assisting and caring for others. These results demonstrate that social, managerial and leadership skills are among those with the least risk of automation by generative AI.

Generative AI automation risk by occupation

Table 3 lists the top 10 occupations with the highest average automation risk from generative AI. According to our results, data entry clerks, general office support workers, and shippers and receivers exhibit the highest automation risk. This aligns with our previous findings, as these occupations comprise a relatively large share of clerical skills and information-processing work activities.

As described, the score for each occupation is obtained by averaging only the subset of skills and work activities that are the most relevant to it (i.e., with a proficiency weight of 3 and higher). So, the fact that “Data entry clerks” exhibit a high automation risk score in table 3 means that all of the skills and work activities most necessary for that occupation exhibit, on average, a high risk of automation.

A closer examination of these high-risk occupations reveals how generative AI might transform their skills and work activities. For instance, health information management occupations (automation risk score: 3.41) combine both highly automatable skills and work activities, like data processing and documentation, with skills and work activities requiring human judgment and interpersonal capabilities. While generative AI shows high potential for automating their information-processing and record-keeping activities, others, such as co-ordinating with health care providers and ensuring compliance with regulations, may exhibit lower automation risk. Similarly, for general office support workers (automation risk score: 3.67), while routine documentation and data entry work activities show high automation potential, their skills in facilitating workplace communication and providing personalized administrative support are likely to remain important human-centred components of the occupation.

In general, it is important to note that automation risk scores reflect the technical feasibility of generative AI replacing or transforming tasks within occupations. However, real-world adoption also depends on economic factors, business incentives and investment costs. With regard to bakers, for example, while large-scale commercial baking operations may find automation cost-effective for streamlining production (e.g., to monitor ingredient usage or predict demand fluctuations), small bakeries may lack the financial incentive to invest in AI-driven systems.

A similar dynamic applies to other occupations with a high average risk score. While generative AI has the potential to automate or assist with cognitive-heavy tasks, full automation would likely require additional advancements in robotics and physical automation. For example, shippers and receivers, who are responsible for processing shipments, tracking inventory and moving materials, may see AI assist with logistics optimization and automated record-keeping. However, the physical aspects of loading, unloading and transporting goods would require robotics rather than generative AI alone. Similarly, while general office support workers could see many of their administrative duties — such as document generation and email drafting — automated by AI, tasks that require co-ordination across multiple departments or handling sensitive information may still require human oversight. These limitations highlight the fact that, while generative AI can transform certain work activities and affect skills demand, full automation often depends on additional investments in complementary technologies.

Table 4 shows the occupations with the lowest average automation risk from generative AI. In line with our previous results, occupations requiring intensive use of social skills and manual work activities show lower automation risk. The relatively low automation risk for retail salespersons, for example, is because skills and work activities such as “negotiating”, “selling or influencing others,” and “persuading” are among their most essential requirements. This finding is consistent with the existing literature on automation, which shows that non-routine manual and interpersonal communication activities have lower automation risk (Lesonsky, 2023). This pattern is particularly evident in occupations requiring direct interpersonal interaction, such as hairstylists and barbers, home child care providers, and personal service workers.

Generative AI automation risk and employment shares

The implication that, over the medium term, generative AI is more likely to change the composition of work activities and skills within occupations rather than rendering entire occupations obsolete, is illustrated by figure 1. It plots the level of automation risk for occupations with the highest employment share. Importantly, the automation risk scores included here are based on skills and work activities essential to the occupation.

Combined, the occupations included in figure 1 accounted for approximately 50 per cent of total employment in Canada in 2021. Overall, automation risk scores among these occupations range in ranking from 2.77, for cashiers and other sales support occupations, to 3.3 for longshore workers and material handlers. As such, taking an average risk score of 3 to indicate moderate automatability, we note that the roles in which employment is most concentrated face a moderate level of automation risk.

Many occupations involve a mix of tasks, some of which have a higher risk of automation than others. For example, we have already seen that skills and work activities related to monitoring tend to have higher automation risk. This means that the moderate automatability (i.e., 2.77-3.3) observed in occupations employing the largest share of workers points to a partial rather than total impact. These roles consist of enough low-risk work activities and skills to maintain an overall automation risk that is moderate rather than high. In other words, the most common jobs are not highly automatable as a whole because they involve both high- and low-risk work activities and tasks.

That said, some occupations within this group may appear more or less automatable depending on the lens through which automation risk is assessed. For example, while our findings suggest that cashiers have a moderate generative AI automation risk score (2.77), other research (e.g., Oschinski & Nguyen, 2022) has classified cashiers as highly automatable. This difference reflects the distinction between generative AI automation and broader automation trends. While self-checkout kiosks and cashier-less stores may reduce the need for traditional cashier roles, our analysis focuses specifically on how generative AI, rather than retail automation in general, impacts the skills and tasks within occupations. In this sense, cashiers remain a role where AI can assist with certain cognitive tasks (e.g., handling customer inquiries via AI chatbots or generating reports), but physical transaction processing and customer interaction remain largely human-driven.

More broadly, this pattern is consistent across many high-employment occupations. Rather than fully replacing jobs, generative AI is likely to automate specific work activities, shifting skill requirements, while leaving others unchanged. Since the occupations in figure 1 account for half of total employment, moderate automation scores imply a shift in skill demand rather than the complete automation of occupations. In other words, generative AI is more likely to transform certain tasks while leaving others to humans, making the complete elimination of high-employment roles unlikely.

In summary, advances in generative AI are more likely to alter the composition of skills and work activities for most workers than render occupations obsolete. While nearly all occupations listed in figure 1 are likely to be impacted by generative AI, the impact will primarily involve a shift in tasks performed by humans.

Industry risk of automation

Having calculated the risk of automation for the skills and work activities related to certain occupations, we may now analyze the risk at industry level in Canada. To do this, we use detailed employment data by industry from Statistics Canada and calculate the share of high-risk occupations by industry.^[4] High-risk occupations here are defined as the top 25 per cent of occupations with the highest risk ratings.

Figure 2 shows industries by total level of employment and each industry’s share of high-risk occupations. The top 5 industries with the highest share of at-risk occupations include transportation and warehousing (56.4 per cent); manufacturing (51.9 per cent); construction (50 per cent); mining, quarrying, and oil and gas extraction (47.7 per cent); and agriculture, forestry, fishing and hunting (36 per cent). Our data suggest that occupations involving routine, standardized tasks such as data entry, basic customer inquiries and administrative work are at the highest risk of automation from generative AI technologies across these industries. In general, these roles involve a high degree of repetitive, predictable tasks. In fields like transportation, warehousing, manufacturing and construction, generative AI can optimize workflows, analyze data, generate schedules and support customer service. In agriculture, generative AI can further take over work activities involving crop and livestock monitoring, supply chain management and market analysis (Rane et al., 2024). In mining, key applications of generative AI include enhancing prospecting and deposit analysis, optimizing mining methods, improving worker safety and environmental monitoring, and increasing operational efficiency throughout the supply chain (Corrigan & Ikonnikova, 2024).

The industries with the lowest shares of high-risk occupations include educational services (3.1 per cent), finance and insurance (5.8 per cent), arts, entertainment and recreation (5.9 per cent), health care and social assistance (6.9 per cent), and professional, scientific and technical services (7 per cent). Industries with a low share of high-risk occupations typically involve roles that emphasize human-centric skills, complex decision-making and adaptability in unstructured environments — qualities that are less easily replaced by automation. While generative AI might assist in those occupations, it cannot easily replace nuanced communication, empathy, human judgment or ethical decision-making. Furthermore, creative and unstructured work, common in arts and entertainment, resists automation because it hinges on individual creativity, which AI struggles to codify. Finally, roles requiring high expertise, such as university researchers, demand years of specialized training and ongoing education to evaluate complex information, making them less vulnerable to AI-driven substitution and more likely to see AI as a complementary tool.

Our assessment indicates that industries vary considerably with regard to their share of at-risk occupations. It should be noted here, however, that the actual impact of generative AI on specific industries depends on both the speed and the nature of technology adoption.

Existing research suggests the pace of AI adoption may differ across regions and industries based on factors like firm size, R&D investment, available talent and quality of the IT infrastructure (Ali et al., 2024). Larger firms and tech-intensive industries may be able to integrate AI solutions quicker than smaller firms, or those in more traditional sectors (Bonney et al., 2024).

Additionally, the types of generative AI technologies implemented can significantly influence their workforce impacts. Firms have a choice in how they deploy these tools — whether to primarily automate and replace human labour, or to augment and enhance worker capabilities (Acemoglu & Johnson, 2023). Industries that strategically leverage AI to complement their human workforce may be able to mitigate job displacement risks.

Further, although some of Canada’s largest employers, such as health care and social assistance, and professional, scientific, and technical services, have relatively low shares of occupations at high risk from automation by generative AI, the total number of employees affected could still be considerable. Due to the large workforce in these sectors, more than 10,000 workers in each are vulnerable to potential job transformation from AI-driven automation.

Finally, the potential risk of automation from generative AI at the industry level will likely vary across regions. This is because different regions have distinct industry patterns, with some areas displaying a larger concentration of sectors with higher shares of at-risk occupations. As such, regions with a greater concentration of industries like manufacturing, transportation and warehousing could face more significant workforce disruptions, while those with a higher share of lower-risk sectors, such as health care and education, may experience less immediate impacts. We discuss regional labour market vulnerabilities in more detail next.

Assessing Regional Labour Market Vulnerabilities — Automation Risk Scores and Online Job Posting Data

The impact of generative AI on the labour market may differ within Canada according to differences in economic structure, workforce composition and industry presence. For policymakers, it is important to know which regions could be most adversely affected in order to react with appropriate policy interventions. Consequently, this section considers the impact of generative AI on the regional demand for occupational labour in Canada. To accomplish this, we leverage data from online job postings as a measure of demand. As such, combining our measure of automation risk with the online job posting data allows us to discern differences in demand for occupations that are the most and least vulnerable to advances in generative AI by geographical location. By investigating this relationship, we aim to determine regional labour market trends in the context of technological change, and assess any patterns that may arise.

To achieve this, we introduce data into our analysis from the Labour Market Information Council (LMIC). Updated weekly, the LMIC’s Canadian Job Trends Dashboard includes labour market information based on online job posting trends across Canada. The dashboard tracks movements in employers’ demand for work skills, knowledge requirements and occupational vacancies. Although many employers actively recruit online, it is important to note that job posting is not an all-encompassing metric for all job vacancies. In particular, research suggests that online job postings might oversample high-skilled occupations (Carnevale et al., 2014).

To determine labour market trends, it is important that we first define a method of establishing what high demand looks like for an occupation versus low demand. In other words, we assess the relative demand for an occupation in a province compared to its demand at the national level. To do this, we use a normalized score.

Normalization adjusts variables to a standard scale, allowing for fair regional comparison. An example of this is the use of per capita values to better compare data across areas with varying population levels. In our context, we calculate the relative proportion of postings for a specific occupation in a province, compared to the national level, to assess proportionate demand.

Specifically, we use the measure of the location quotient (LQ), following Alabdulkareem et al. (2018). We generate the LQ by province to indicate whether a province has a relatively higher or lower demand for a particular occupation. The score can be interpreted as follows:^[5]

1. LQ > 1: the province has a higher share of job postings for a particular occupation than the national average
2. LQ = 1: the province’s share of job postings for the occupation is in line with the national average
3. LQ < 1: the province has a lower share of job postings for a particular occupation than the national average

These values simply allow us to determine some relative method of explaining which occupations have a higher or lower demand. To reiterate, this is important because we are interested in identifying which in-demand occupations exhibit a high automation risk. It assists in informing policy about which occupations are primed for job-training investment, namely those exhibiting relatively higher demand but low levels of automation risk, and where they are in demand. For example, if the occupation “Electrician” is in high demand in Alberta and has a low level of automation risk, this would indicate there is a growing need for electricians in Alberta with a low likelihood of this demand declining due to technological change in the near future.

However, generative AI’s impact on an occupation is not solely determined by the share of tasks the tool can handle, but also by the types of tasks, their importance, and the context in which they are required. In contrast to automation risk, this concept – referred to as complementarity – is meant to capture the degree to which generative AI can enhance or supplement human tasks rather than entirely replace and automate them. Pizzinelli et al. (2023) built an index of AI complementarity at the occupation level, based on data on occupations and their work context (conditions, characteristics and relationships) from O*NET. We incorporate this dataset into our analysis, which was generously provided by the authors, by adapting it to align with the OaSIS framework.

Combining automation risk scores with complementarity scores allows us to determine which occupations are likely to experience the most fundamental disruption from generative AI. In other words, considering occupations that are in high demand, but that face high automation risk and low complementarity, is one way to identify those that may be positioned to face significant job losses. This subgroup of occupations is particularly vulnerable because these roles currently require filling (i.e., they are in high demand) but firms can choose between investing in modernization or human labour (i.e., due to high automation risk). This means that it may be more beneficial for a firm to invest in modernization than in human labour (i.e., due to low complementarity), leading to high displacement. Displacement does, however, depend on sector-based incentives. For example, it may be more difficult for a smaller firm to adopt automation due to financial or resource constraints, despite the automation risk. Conversely, larger firms with more capital may find it easier to invest in technology, resulting in greater displacement of workers in these high-risk, low-complementarity roles.

A Brookings Institution report by Kinder et al. (2024) draws attention to a significant shift in how generative AI impacts the labour market. Unlike previous waves of automation, which predominantly affected manual, routine tasks, generative AI is poised to disrupt routine cognitive tasks in office-oriented and information-based roles. In contrast, occupations that were once most vulnerable to automation — typically lower-skilled, routine manual jobs —are likely to be more insulated from displacement in this context. This is due to the slower adoption of advanced AI technologies in these industries, which often lack the resources or need for such investments. Conversely, sectors that rely on routine cognitive tasks, such as administrative work or customer service, may see greater disruption, but workers in these roles are also more likely to benefit from AI-driven productivity enhancements. The extent of job displacement will depend on factors such as firm size and sector characteristics. Larger firms and industries with more cognitive, routine tasks are better positioned to adopt generative AI, while smaller firms in sectors that rely on manual or complex tasks may face slower adoption. Thus, the impact of generative AI on job displacement will vary across industries and will be influenced by each sector’s capacity to invest in new technologies. As a result, the impact of generative AI on job displacement will not be uniform across industries and will be shaped by each sector’s unique needs and resources.

For occupations currently in high demand but with a high level of automation risk and low complementarity with generative AI, a sound course of action will be (with the caveats mentioned previously) to adopt automation in the field and manage transitions for workers expecting to be displaced in preparation for a different or altered line of work. Following Bender and Li (2002), who define additional intervals of a mathematically similar measure, we define the following categories of relative demand for more detailed insights:

Category 1: LQ > 2: very high overrepresentation
Category 2: 2 > LQ > 1: high overrepresentation
Category 3: LQ = 1: average representation
Category 4: 1 > LQ > 0.5: low representation
Category 5: If 0.5 > LQ > 0: very low representation

From a policy perspective, we are most interested in those occupations that are significantly overrepresented and, more specifically, highly represented in each province and with respect to their automation risk and complementarity scores (Categories 1 and 2).

It is also interesting to consider to what extent the location quotient is a reflection of regional economic specialization. For example, our data find that British Columbia has a very high over-representation for “Contractors and supervisors, carpentry trades” and Nova Scotia in “Conservation and fishery officers.”

To determine a macro-level understanding of regional differences in labour market automation risk, we aggregate the Category 1 data detailed above. That is, we generate the average level of generative AI automation risk for the highest in-demand occupations (those with an LQ greater than 2) in each province.

Using just Category 1 data allows us to gain further detail on specialization. Recall that, if the LQ is greater than 2, it demonstrates a significantly higher concentration of job postings for a specific occupation in a province than the national average.

Comparing the expected level of susceptibility to generative AI for those occupations most in demand in each province is a useful policy tool, as it may indicate priorities for strategic workforce planning from a macro perspective, serving as a high-level indicator.

Figure 3 shows the average level of generative AI automation risk for the most in-demand occupations by province.

We note that focusing on in-demand occupations does not account for automation risk in occupations where firms may already be choosing between investing in modernization or closing. Then, for the most in-demand occupations, we see that Ontario exhibits the highest average level of generative AI automation risk at 3.62, followed by Manitoba. In contrast, Prince Edward Island and Newfoundland and Labrador have the lowest average levels. Thus, occupations with the highest labour demand in Ontario and Manitoba have, on average, higher automation risk due to advances in generative AI compared to those in Prince Edward Island and Newfoundland and Labrador.

The implications of this insight are manifold. For example, workers in Ontario and Manitoba may face higher risks of job displacement, which could lead to increased unemployment or exacerbate economic inequality if workers struggle to transition to new roles or technology. On the other hand, Prince Edward Island and Newfoundland and Labrador may experience higher resiliency in their most specialized fields, indicating a lesser level of urgency in terms of upskilling or retraining policy efforts.

As noted above, the ultimate impact of generative AI on a specific occupation is going to depend on the level of automation risk as well as its level of complementarity with this technology. Automation risk refers to the likelihood that the activities required by an occupation can be reasonably performed by generative AI tools. Complementarity refers to the tasks within an occupation that can be augmented or enhanced by AI, rather than fully automated. This is determined by factors such as whether the tasks require human communication, complex decision-making, physical presence, or other uniquely human capabilities. For example, workers in occupations with high automation risk and high complementarity, like lawyers, are less likely to be at risk of displacement and more likely to see productivity gains — since generative AI is expected to be able to augment or enhance rather than replace their tasks.

As such, we compare the most in-demand occupations by province in terms of their complementarity and automation risk. Figures 4 and 5 plot regional comparisons of the distribution of the most in-demand occupations with high automation risk and high complementarity vs. low complementarity.

We define relatively high automation risk as those occupations with an automation risk score greater than or equal to 3.02. We follow Pizzinelli et al. (2023) and define relatively high complementarity as being greater than 0.58.

As mentioned previously, occupations exhibiting high automation risk alongside high complementarity indicate a likelihood of benefiting positively from advances in generative AI since there exists a high degree to which generative AI may enhance or supplement, rather than replace, tasks. Similarly, occupations with high automation risk and low complementarity are more likely to face negative repercussions, such as job replacement.

Given this, we can identify regions that are best positioned to benefit from advances in generative AI. These are areas with in-demand occupations that have both high automation risk and high complementarity. Occupations meeting these criteria have a higher potential for augmentation and benefit from generative AI; therefore, the imminent changes caused by advancements in generative AI are more likely to be positive for these regions.

Figure 4 highlights how well positioned each province is to benefit from generative AI in its most in-demand and at-risk roles. We focus on occupations that are both highly exposed to AI-driven automation and likely to benefit from it — those with a high automation risk and high complementarity score. These occupations are matched with a dataset of the most in-demand occupations by province. We then calculate the average complementarity score for each province. The resulting map shows where AI is more likely to enhance rather than replace in-demand jobs, offering key insights for workforce planning and investment in skills development.

Figure 5 is similarly constructed but instead focuses on occupations highly exposed to AI-driven automation and unlikely to benefit from it ­— those with a high automation risk and low complementarity score.

Figure 4 shows that Nova Scotia, Alberta, British Columbia, and Ontario’s in-demand occupations with high automation risk show relatively higher complementarity scores. In contrast, some provinces, such as New Brunswick, are not represented on the map, as they do not have occupations that are in high demand and that exhibit high automation risk with a high level of complementarity.

This highlights a critical key takeaway: not all provinces are poised to benefit equally from advances in generative AI. For policymakers, understanding which provinces are less likely to benefit — such as Newfoundland and Labrador or Manitoba — can inform targeted regional interventions to improve labour market robustness.

Moreover, our data reveal clear regional differences in terms of vulnerability. A region is considered vulnerable if it has more in-demand occupations with high generative AI automation risk and low complementarity. Occupations meeting these criteria, as shown in figure 5, are more likely to be negatively impacted by advances in generative AI, as the technology is less likely to complement the required skills and work activities.

We observe that more populous provinces like Ontario and Quebec have lower average complementarity across in-demand, high-automation risk, lowcomplementarity occupations than provinces like Newfoundland and Labrador, Nova Scotia, or Alberta. Economic specializations may play a role in these findings; these provinces often employ trades requiring skills and work activities in natural resource sectors, such as fishing, oil, or mining, where certain skills might offer slightly greater complementarity with technology despite high automation risk.

In contrast to figure 4, all provinces are represented in figure 5. This indicates another key takeaway: all Canadian provinces have occupations that are in high demand and likely to see a negative impact from advancements in generative AI. This is significant from a policy perspective, as having high-demand occupations across all provinces vulnerable to generative AI advancements could pose widespread challenges.

If widely held jobs are vulnerable to AI, a significant portion of the workforce could face displacement or the need for rapid reskilling. This could lead to economic instability, including higher unemployment rates and income inequality, which are traditionally addressed through government policy and programs.

At the same time, effective reskilling programs and workforce transitions are not just about mitigating risks — they are also key to ensuring that workers and businesses fully capture the productivity gains made possible by generative AI, in particular for occupations that have high automation risk and high complementarity. By equipping workers with the skills needed for AI-augmented roles, governments and employers can help unlock higher efficiency, innovation, and economic growth.

Governments have a critical role in funding and designing reskilling programs to help workers transition to new roles or adapt their skills to AI-augmented tasks. Without intervention, the workforce might not be able to meet the demands of the evolving labour market.

Overall, these maps highlight a critical asymmetry in the geography of generative AI disruption and opportunity. While workers in some provinces are better positioned (high risk, high complementarity) to benefit in advances in generative AI through the augmentation of existing roles, others face a disproportionate risk of replacement (high risk, low complementarity).^[6] The absence of proactive intervention may deepen existing regional inequalities. With regionally tailored support, generative AI can be harnessed to promote inclusive transformation across regions.

Regional Employment Vulnerability by Industry

While the preceding analysis focused on in-demand occupations and their automation risk, policymakers must also consider the general distribution of high-risk employment across industries and regions. Understanding these patterns is critical for designing effective workforce transition policies, particularly in sectors with a high share of vulnerable jobs. As noted earlier, industries such as transportation, manufacturing, construction, mining and agriculture face a high share of at-risk occupations due to generative AI, particularly those involving routine, standardized tasks. These roles are most vulnerable to automation, as generative AI can optimize workflows, support customer service, and handle tasks like data entry, scheduling and monitoring, leading to efficiency gains across these sectors.

Given regional variations in the types of occupations within industries, the distribution of high-risk employment is likely to differ across regions as well. Accordingly, figure 6 presents the percentage of employment within each industry and province/territory that is categorized as high-risk due to generative AI. As mentioned throughout the report, “automation risk” refers purely to the technical feasibility of a generative AI system replacing or significantly transforming a specific occupational skill or work activity. As such, it does not account for regulatory barriers or the fact that the implementation of AI systems in some cases might require substantial investments and thus take some time to materialize.

As shown in figure 6, the share of high-risk employment varies considerably both by industry and by region. Across the country, industries are made up of slightly different mixes of occupations, reflecting firm-specific differences such as production methods, size, and the availability of labour and resources. We find that some regions have a larger share of people employed in occupations more likely to be impacted by generative AI, even within the same industries.

In construction, between 47 per cent and 67 per cent of occupations across all provinces and territories exhibit high risk, with the highest vulnerability in Prince Edward Island (62 per cent) and Nunavut (67 per cent). Similarly, manufacturing shows high vulnerability with 50-67 per cent of occupations at high risk across regions, peaking at 67 per cent in the Northwest Territories and Yukon. This could leave them more susceptible to larger vocational displacement, though it will depend on the technological, financial and social feasibility of automating certain tasks.

Resource-based industries also demonstrate substantial vulnerability. In mining, quarrying, and oil and gas extraction, the share of high-risk employment ranges from 41 per cent in Alberta to almost 67 per cent in Nunavut. While generative AI can optimize certain tasks within these sectors, the manual aspects — such as equipment operation and fieldwork — will likely require significant investments in specialized equipment and technology. These investments may only be feasible where considerable capital has already been committed. Transportation and warehousing also shows high risk shares, while agriculture, forestry, fishing and hunting exhibits notable variations, from 27 per cent in Saskatchewan to 56 per cent in Nunavut.

In contrast, certain industries show consistently lower shares of high-risk employment. Educational services demonstrate the lowest vulnerability across all provinces, followed by arts, entertainment and recreation, and finance and insurance. Health care and social assistance also shows relatively low risk shares.

These findings complement our earlier analysis of in-demand occupations in important ways. While Ontario exhibits high average automation risk from generative AI for its most in-demand occupations, we also see that several industries in Ontario have moderate shares of high-risk employment compared to other provinces. For example, Ontario’s construction industry shows a 48 per cent high-risk share, which is lower than most other provinces. Conversely, while our LQ analysis indicated lower average risk in Prince Edward Island’s most in-demand occupations, the province shows elevated high-risk shares in several key industries, including construction and transportation.

Regional patterns also emerge that may require policy attention. Nunavut consistently shows very high shares of high-risk employment across multiple industries. The Northwest

Territories similarly shows elevated vulnerability across several sectors. Further studies are

needed to understand the economic factors behind these differences in the occupational structures of industries across regions.

These patterns suggest that certain regions may face more widespread workforce disruption from generative AI than others, across both current and indemand employment.

From a policy perspective, regions with high shares of high-risk occupations across multiple industries may need more comprehensive transition strategies, even in cases where the most in-demand occupations in these regions show lower average risk. This dual perspective — considering both occupational demand patterns and the current distribution of high-risk employment — provides a more complete foundation for targeted policy intervention.

What Does the Evidence Reveal So Far?

Concern about artificial intelligence’s impact on work has intensified with the emergence of generative AI — systems capable of creating content across text, code and other media. While previous waves of automation primarily affected routine manual and cognitive tasks, generative AI’s ability to perform complex cognitive and creative tasks has raised new questions about its implications for Canada’s workforce. This study provides empirical evidence of how generative AI could affect Canadian workers over the next five years, examining its impact across skills, work activities, occupations, industries and regions.

Our analysis of generative AI’s automation potential reveals three significant patterns in the Canadian labour market. First, the impact varies substantially across different types of skills and work activities. Clerical activities show the highest automation risk (4.29 out of 5), followed by operation monitoring (3.96) and data analysis (3.92). In contrast, skills involving human interaction and judgment — such as instructing (2.04), social perceptiveness (2.08) and coaching (2.17) — demonstrate markedly lower automation risk. This pattern suggests that, while generative AI may significantly transform information processing and monitoring tasks, it is less likely to replace activities requiring interpersonal skills and judgment-based functions.

Second, the evidence indicates that generative AI is more likely to transform the composition of work within occupations rather than eliminate entire job categories. Among occupations representing about 50 per cent of total Canadian employment, automation risk scores fall within a moderate range (2.77-3.3), suggesting partial rather than complete automation. In other words, most occupations will evolve rather than disappear, with workers needing to adapt to changing task compositions. However, the extent of this transformation will depend on factors beyond technical feasibility, including employer adoption strategies, worker reskilling efforts, and policy interventions. Governments will play a critical role in supporting workforce transitions through education and training investments, while businesses will need to implement AI in ways that enhance, rather than replace, human work.

Third, our analysis reveals significant regional and industry variations in automation risk. Transportation and warehousing (56.4 per cent), manufacturing (51.9 per cent), and construction (50 per cent) show the highest shares of high-risk occupations, while educational services (3.1 per cent) and finance and insurance (5.8 per cent) demonstrate lower vulnerability. Regional analysis indicates that Ontario and Manitoba have higher concentrations of at-risk occupations in their most in-demand jobs, while Prince Edward Island and Newfoundland and Labrador show greater resilience. These patterns suggest that the impact of generative AI will not be uniform across Canada’s economy, necessitating targeted policy responses.

The evidence also highlights important implementation challenges that will influence the pace and impact of generative AI adoption. Canada currently lags other G7 countries in AI adoption, with only 3.1 per cent of companies having implemented AI technologies by 2022. Two critical barriers emerge from the data: insufficient AI-enabling infrastructure and a shortage of AI-ready talent. Canada’s recent drop to 23rd place globally in AI infrastructure underscores these challenges. This suggests that, while generative AI offers significant potential for workplace transformation, actual changes may occur more gradually than technical feasibility alone would indicate.

Looking ahead, our analysis indicates that the ultimate impact of generative AI on Canada’s workforce will depend on several interrelated factors. While traditional AI adoption is constrained by infrastructure limitations, generative AI presents a unique dynamic. Unlike traditional AI systems that require significant organizational investment, generative AI tools are increasingly accessible to individual workers through consumer-facing applications. However, this accessibility creates both opportunities and risks. Without proper AI literacy and digital skills, workers may use these tools ineffectively or inappropriately, potentially reducing rather than enhancing productivity. This is echoed in a recent study by the Conference Board of Canada on the use of generative AI. According to the authors, generative AI has the potential to provide a significant boost to Canada’s economy — adding around 2 per cent to GDP — if deployed correctly (The Conference Board of Canada, 2024). Crucially, the report notes that AI talent and an AI-ready workforce are key factors in this process. Yet, a lack of an AI-ready workforce appears to be one of the biggest bottlenecks (Pamma, 2024).

The key challenge for policymakers and business leaders will therefore be the following:

1. Accelerate AI Infrastructure Investment

The federal government should continue investing in AI infrastructure, including AI compute capacity, data centres and broadband access as part of the AI Compute Access Fund and the Canadian AI Sovereign Compute Strategy (ISED, 2025a, 2025b). This will enable AI researchers, startups and businesses to have the necessary resources to innovate and scale AI solutions, ensuring equitable access across all regions.

Canada has made strides with the $2 billion AI investment in Budget 2024 but significant gaps remain in AI infrastructure (Finance Canada, 2024) These investments should be scaled to ensure equitable access for small businesses and ­underrepresented regions, supporting the overall AI ecosystem. High-performance computing is critical to ensuring that AI innovation is both scalable and sustainable. However, without widespread broadband access, regions with limited internet connectivity may be left behind in the AI revolution, as access to AI tools and cloud-based computing requires reliable, high-speed internet.

2. Foster a National AI Literacy Initiative

The federal government should work with the provinces and territories to implement a comprehensive AI literacy program across secondary, post-secondary and adult learning levels. This program should focus on digital literacy and complementary skills — such as critical thinking, problem-solving and leadership skills — that are vital in an AI-augmented workforce and exhibit low automation risk. Importantly, AI literacy should also emphasize human oversight and ethical engagement with AI tools.

Digital and AI literacy are crucial to preparing workers across all sectors for AI integration (Oschinski et al., 2024). Recent Canadian government AI consultations highlight the skills gap in AI knowledge and the lack of readiness among workers to engage with AI tools effectively (Government of Canada, 2024).

3. Strengthen Public-Private Partnerships for Workforce Reskilling
   The government should facilitate work-based learning programs that provide hands-on experience in AI. These programs should be collaborative in nature, with input from businesses, industry associations and educational institutions to create industry-relevant AI curricula. Apprenticeships and internships have proven to be effective at connecting education to real-world industry needs, particularly in AI fields. Moreover, they are effective pathways for underrepresented groups to gain access to emerging jobs (Koslosky & Feldgoise, 2025). By providing workers with both technical skills and practical experience, Canada can enhance the talent pipeline and ensure workers are ready to step into AI roles directly. This approach also empowers workers to anticipate how generative AI is reshaping their industries and to proactively identify opportunities where their skills can be applied. In this context, expanding the programs like the National Research Council’s AI Assist Program, ISED’s AI Compute Access Fund and the Regional Artificial Intelligence Initiative (ISED 2024 and 2025c), could help small and medium-sized businesses to adopt AI tools while simultaneously upskilling their workforce in AI technologies.

4. Support Region-Specific Workforce Development Strategies
   To address the regional disparities in AI adoption, the government should implement regional workforce strategies tailored to the specific economic contexts of each province/territory. These strategies should focus on reskilling workers in high-risk sectors and regions with high concentrations of vulnerable occupations.

As our analysis has shown, the impact of AI will not be uniform across Canada. Different regions have varying levels of AI infrastructure access and exposure to automation risks. Developing region-specific programs is key to ensuring equitable workforce development. In this context, expanding the Sectoral Workforce Solutions Program (ESDC, 2022) funding to target specific regions with higher automation risks could be part of a broader regional development strategy.

5. Strengthen AI Talent Development and Retention
   Canada should invest in AI talent development through research grants, AI-­related apprenticeships, and collaborations with businesses. This includes attracting and retaining AI researchers and practitioners, especially in emerging fields of AI application, by offering competitive incentives and training programs.

To strengthen its position in AI, Canada must expand its AI talent pipeline and ensure that research institutes and startups can attract and retain skilled professionals. AI-related apprenticeships, industry collaborations and effective workforce training programs can contribute to building a sustainable talent pool, ensuring that Canada can compete effectively on the global AI stage (Koslosky and Feldgoise, 2025; Oschinski et al., 2024).

While this study provides an important baseline assessment of generative AI’s potential impact, further research is needed to fully understand how automation risk translates into real-world workforce shifts. Future studies could examine employer adoption strategies, firm-level AI investment trends, and the effectiveness of reskilling programs in preparing workers for AI-augmented roles. Additionally, longitudinal research tracking workforce adaptation over time will be crucial for evaluating how generative AI reshapes employment patterns in practice.

Monitoring these changes will be crucial as generative AI continues to evolve. Ongoing assessment of how Canadian jobs actually transform as this technology is deployed will be essential for informing policy responses and workforce development strategies.

---

^[1]  These databases can be accessed using the following links: OaSIS: https://noc.esdc.gc.ca/Oasis/Oasis
Welcome; O*NET: https://www.onetcenter.org/database.html; ISCO: https://data.europa.eu/data/datasets/european-skills-competences-qualifications-and-occupations?locale=en.

^[2]  Note: Excluding weights from 0 to 2 does not impact our final results.

^[3]  For the complete list of skills and work activities with their associated level of automation risk, see appendix E.

^[4]  Note: We use occupation by industry data at the 4-digit NOC level (Statistics Canada, table 98-10-0594-01). This requires us to aggregate our occupational risk scores from the 5-digit NOC level to the 4-digit NOC level.

^[5]  To calculate this normalized score, the total number of postings for an occupation in a province was divided by the total number of provincial postings. This result was then divided by the result of the total number of postings for an occupation in Canada, divided by the total number of national postings.

^[6]     In some regions, like Ontario and British Columbia, there’s a significant presence of in-demand at-risk occupations with both high and low complementarity scores relative to other provinces.

---

Appendices

Appendix A

As OpenAI develops its models, it retires older versions. At the time of analysis, GPT-4 was running on version GPT-4-0613, the model used in this research. A comparison between each of the model versions used in this paper is shown in table A1. This bolsters the thoroughness of the methodology described in the paper, as it takes into account another source of variation in GPT’s ability to provide consistent automation risk scores.

Appendix B

Below is a list of prompt parameters and hyperparameters. In writing a script to use the GPT API, there are certain technical specifications that users may input to alter the consistency, length or any of the other response traits that arise. The hyperparameters most relevant to our study — and which were used in some of our prompts — are listed below to provide a complete view of the thoroughness of our approach.

• Temperature: Controls the randomness of the output. Setting this close to 1.0 may introduce some randomness; a value closer to 0 will produce deterministic rankings.
• Top_P: 0.1: How many options GPT considers before choosing the next word in its response. If set to 0.1, the model only looks at words that make up the top 10 per cent of the most likely options. If set closer to 0, the choices will be narrower. Only used in prompts that use post-processing explanations.
• Batch Size: Number of data points used in each iteration during training (impacts model convergence).
• Template-Based Prompt: Suggested method of creating prompt structures (as seen in appendix C) to improve consistency in outputs.
• Frequency Penalty: Used to penalize the frequency of words in the model’s output. Only used in prompts that use post-processing explanations.
• Presence Penalty: Discourages the use of words from the input prompt in the response. Can create more/less diverse ranking. Only used in prompts that use post-processing explanations.
• Prompt Ranking Criteria: Use of defined ranking criteria (e.g., 1 = some explanations, 5 = some explanation) to improve consistency in outputs.

Appendix C

Below are the different prompts that were fed to the ChatGPT Application Programming Interface (API) in order to retrieve automation risk scores for skills and work activities. The purpose of using different prompt variations in this process was to be able to average out automation risk scores and ensure consistency in the API outputs.

Version 1

Please rate the automatability of the skill in the context of generative AI development over the medium term (next 5 years) on a scale of 1 to 5, where:

1 = Not automatable
2 = Slightly automatable
3 = Moderately automatable
4 = Highly automatable
5 = Fully automatable

Please provide a single numerical rating based on this scale!

Version 2

As an AI expert how would you rate the automatability of the following skill over the medium term (next 5 years). Please provide a short explanation and rate automatability on a scale of 1 (=not automatable) to 5 (=fully automatable).

Version 3

Same prompt as version 2. In this version, skills and work activities include descriptions from OaSIS.

Version 4

Assume you are an expert in generative AI: As such rate how the following skill can be automated by generative AI. Please use a scale from 1 to 5 whereby 1 = ‘not automatable’ and 5 = ‘fully automatable’. As an answer, please just give 1 single number!

Version 5

Assume you are an expert in generative AI: As such rate how the following skill can be automated by generative AI. Please use a scale from 1 to 5 whereby 1 = ‘not automatable’ and 5 = ‘fully automatable’. As an answer, please give 1 single number and a short explanation for your rating!

Version 6

Please rate the following occupational skill for its susceptibility to advances in generative AI over the next 5 years on a scale of 1 to 5, where:

1 =  Not automatable: This skill is unlikely to be automated by generative AI in the next 5 years.
2 =  Slightly automatable: This skill may see limited automation by generative AI in the next 5 years.
3 =  Moderately automatable: This skill has a moderate chance of being automated by generative AI in the next 5 years.
4 =  Highly automatable: This skill is likely to be automated by generative AI to a significant extent in the next 5 years.
5 =  Fully automatable: This skill is highly likely to be completely automated by generative AI in the next 5 years.

Please provide a single numerical rating for each skill based on this scale. Also include a brief and concise explanation for the rating.

Version 7

Same prompt as version 6. In this version, the frequency penalty is set to 0.2 and presence penalty is set to 0.5. In version 6, both penalty values were set to 0.

Version 8

Please assess the following occupational skill for its susceptibility to automation due to developments in generative AI over the next 5 years on a scale of 1 to 5, where:

1 =  Not automatable: This skill is unlikely to be automated by generative AI in the next 5 years.
2 =  Slightly automatable: This skill may see limited automation by generative AI in the next 5 years.
3 =  Moderately automatable: This skill has a moderate chance of being automated by generative AI in the next 5 years.
4 =  Highly automatable: This skill is likely to be automated by generative AI to a significant extent in the next 5 years.
5 =  Fully automatable: This skill is highly likely to be completely automated by generative AI in the next 5 years.

Please provide a single numerical rating for each skill based on this scale. Also include a brief and concise explanation for the rating.

Version 9

Same prompt as version 8. In this version, the frequency penalty is set to 0.2 and presence penalty is set to 0.5. In version 8, both penalty values were set to 0.

Version 10

Please rate the following skill for its susceptibility to advances in generative AI over the medium term (next 5 years) on a scale of 1 to 5, where:

1 =  Not automatable: This skill is unlikely to be automated by generative AI in the next 5 years.
2 =  Slightly automatable: This skill may see limited automation by generative AI in the next 5 years.
3 =  Moderately automatable: This skill has a moderate chance of being automated by generative AI in the next 5 years.
4 =  Highly automatable: This skill is likely to be automated by generative AI to a significant extent in the next 5 years.
5 =  Fully automatable: This skill is highly likely to be completely automated by generative AI in the next 5 years.

Please provide a single numerical rating for each skill based on this scale.

Version 11

As an expert in labour and technology how would you rate the automatability of the following skill over the medium term (= the next 5 years). Please rate the automatability on a scale of 1 (=not automatable) to 5 (=fully automatable).

Version 12

As an expert on skills and emerging technologies how would you rate the automatability of the following skill over the medium term (= the next 3-5 years). Please rate the automatability on a scale of 1 (=not automatable) to 5 (=fully automatable).

Appendix D

 

To ensure consistency across prompts and the scores generated by GPT, a prompt structure was used as a tool to generate the prompt variations seen in appendix C. This structure is a loose outline and was modified slightly across prompts that asked for explanations or for the API to assume the role of an expert.

• Verb 1: rate, assess
• Noun 1: skill, work activity, occupational skill
• Verb 2: automatability, susceptibility to, automation
• Noun 2: in the context of, advances in, due to developments in
• Period: over the medium term (next 5 years), over the next 5 years, over the medium term (the next 3-5 years)
• Verb 3: return, give
• Return: single number, single numerical rating

Each bolded area below indicates where one of the above variables would be inserted:

Please [Verb 1] the following [Noun 1] for its [Verb 2] [Noun 2] generative AI [Period], where:

1 = …
2 = …
3 = …
4 = …
5 = …

Please [Verb 3] a [Return] based on the defined scale.

---

References

Acemoglu, D. (2024a, May 21). Don’t believe the AI hype. Project Syndicate.

Acemoglu, D., & Johnson, S. (2023). Power and progress. PublicAffairs.

Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labour. Journal of Economic Perspectives, 33(2), 3-30.

Acemoglu, D., & Restrepo, P. (2022). Tasks, automation, and the rise in U.S. wage inequality. Econometrica, 90(5), 1973-2016.

Agrawal, A., Gans, J. S., & Goldfarb, A. (2019). Artificial intelligence: The ambiguous labor market impact of automating prediction. Journal of Economic Perspectives, 33(2), 31-50.

Alabdulkareem, A., Frank, M. R., Sun, L., AlShebli, B., Hidalgo, C., & Rahwan, I. (2018). Unpacking the polarization of workplace skills. Science Advances, 4(7), https://doi.org/10.1126/sciadv.aao6030

Ali, H., ul Mustafa, A., & Aysan, A. F. (2024). Global adoption of generative AI: What matters most? Journal of Economy and Technology.

Bender, S., & Li, K.-W. (2002). The changing trade and revealed comparative advantages of Asian and Latin American manufacture exports. http://www.econ.yale.edu/~egcenter/ http://papers.ssrn.com/abstract=303259.

Bick, A., Blandin, A., & Deming, D. J. (2024). The rapid adoption of generative AI. Working Paper 32966. National Bureau of Economic Research.

Billy-Ochieng’, R., Anusha, A., & Garcia, D. (2024, May 28). Artificial intelligence technologies can help address Canada’s productivity slump. TD Economics.

Bonney, K., Breaux, C., Buffington, C., Dinlersoz, E., Foster, L. S., Goldschlag, N., Haltiwanger, J. C., Kroff, Z., & Savage, K. (2024). Tracking firm use of AI in real time: A snapshot from the Business Trends and Outlook Survey. Working Paper 32319. National Bureau of Economic Research.

Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). What can machines learn and what does it mean for occupations and the economy? AEA Papers and Proceedings, 108, 43-47.
https://doi.org/10.1257/pandp.20181019.

Burt, M. (2023). ChatGPT: Organizational and Labour Implications. (2023, November 16). The Conference Board of Canada. https://www.conferenceboard.ca/product/chatgpt_november2023/

Caranci, B., & Marple, J. (2024, September 12). From bad to worse: Canada’s productivity slowdown is everyone’s problem. TD Economics.

Carnevale, A. P., Jayasundera, T., & Repnikov, D. (2014). Understanding online job ads data.
https://cew.georgetown.edu/wp-content/uploads/2014/11/OCLM.Tech_.Web_.pdf.

Conference Board of Canada. (2024). Real talk: How generative AI could close Canada’s productivity gap and reshape the workplace — Lessons from the innovation economy. Conference Board of Canada.

Corrigan, C. C., & Ikonnikova, S. A. (2024). A review of the use of AI in the mining industry: Insights and ethical considerations for multi-objective optimization. The Extractive Industries and Society, 17, Article 101440.

Council of Economic Advisers. (2024, July 10). Potential labor market impacts of artificial intelligence: An empirical analysis. The White House.

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An early look at the labour market impact potential of large language models. http://arxiv.org/abs/2303.10130.

Employment and Social Development Canada (ESDC). (2022). About the Sectoral Workforce Solutions Program. Government of Canada. https://www.canada.ca/en/employment-social-development/programs/sectoral-workforce-solutions-program.html.

Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195-2217.

Finance Canada. (2024). Budget 2024: Fairness for every generation. Government of Canada. https://budget.canada.ca/2024/home-accueil-en.html.

Frank, K., & Frenette, M. (2021). Are new technologies changing the nature of work? The evidence so far. IRPP Study 81. Institute for Research on Public Policy.

Frey, C. B., & Osborne, M. (2013). The future of employment. Oxford Martin School Working Paper.

Frey, C. B., & Osborne, M. A. (2017). The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change, 114, 254-280.
https://doi.org/10.1016/j.techfore.2016.08.019.

Gmyrek, P., Berg, J., & Bescond, D. (2023). Generative AI and jobs: A global analysis of potential effects on job quantity and quality. ILO Working Paper 96. International Labour Organization.
https://www.ilo.org/wcmsp5/groups/public/—dgreports/—inst/documents/publication/wcms_890761.pdf.

Government of Canada. (2024, November 22). What we heard report: Consultations on AI Compute. https://ised-isde.canada.ca/site/ised/en/what-we-heard-report-consultations-ai-compute.

Innovation, Science and Economic Development Canada (ISED). (2024, October 22). Federal government launches programs to help small and medium-sized enterprises adopt and adapt artificial intelligence solutions [News release]. Government of Canada. https://www.canada.ca/en/innovation-science-economic-development/news/2024/10/federal-government-launches-programs-to-help-small-and-medium-sized-enterprises-adopt-and-adapt-artificial-intelligence-solutions.html.

Innovation, Science and Economic Development Canada (ISED). (2025a). AI Compute Access Fund. Government of Canada. https://ised-isde.canada.ca/site/ised/en/canadian-sovereign-ai-compute-strategy/ai-compute-access-fund.

Innovation, Science and Economic Development Canada (ISED). (2025b). Canadian Sovereign AI Compute Strategy. Government of Canada. https://ised-isde.canada.ca/site/ised/en/canadian-sovereign-ai-compute-strategy.

Jiang, J., Zhou, K., Dong, Z., Ye, K., Zhao, W. X., & Wen, J. R. (2023). StructGPT: A general framework for large language model to reason over structured data. arXiv preprint. arXiv:2305.09645.

Kinder, M., Muro, M., Liu, S., & de Souza Briggs, X. (2024, October 10). Generative AI, the American worker, and the future of work. Brookings. https://www.brookings.edu/articles/generative-ai-the-american-worker-and-the-future-of-work/.

Koslosky, L., & Feldgoise, J. (2025). The state of AI-related apprenticeships. CSET Data Brief, February 2025. https://cset.georgetown.edu/publication/the-state-of-ai-related-apprenticeships/.

Lesonsky, R. (2023, April 17). The future of retail: What the stats say about retailers in 2023. Forbes Newsletter.

Mehdi, T., & Morissette, R. (2024, September 3). Experimental estimates of potential artificial intelligence occupational exposure in Canada. Statistics Canada.

Moll, B., Rachel, L., & Restrepo, P. (2022). Uneven growth: Automation’s impact on income and wealth inequality. Econometrica, 90(6), 2645-2683. https://doi.org/10.3982/ecta19417.

Nuño, B. S. A. (2023, June 8). The Gutenberg Moment in AI and it’s shadow, the end of digital presumption of veracity. Blog post. https://world.hey.com/brunosan/the-gutenberg-moment-in-ai-and-it-s-shadow-the-end-of-digital-presumption-of-veracity-02893edd.

OpenAI. (n.d.-a). OpenAI platform documentation. https://platform.openai.com/docs/models/.

OpenAI. (n.d.-b). OpenAI research: Aligning language models to follow instructions.
https://openai.com/research/instruction-following.

Oschinski, M., Crawford, A., & Wu, M. (2024). AI and the future of workforce training. CSET Issue Brief, December 2024. https://cset.georgetown.edu/publication/ai-and-the-future-of-workforce-training/

Oschinski, M., and Nguyen, T. (2022). Finding the Right Job: A Skills-Based Approach to Career Planning. IRPP Study 86. Montreal: Institute for Research on Public Policy. DOI:
https://doi.org/10.26070/5paz-x420

Oxford University Press. (2023). Artificial intelligence. In Oxford English Dictionary.

Pamma, A. (2024, January 23). AI adoption in Canada rising, but upskilling remains top barrier: IBM. IT World Canada. https://www.itworldcanada.com/article/ai-adoption-in-canada-rising-but-upskilling-remains-top-barrier-ibm/556849.

Pizzinelli, C., Panton, A. J., Tavares, M. M. M., Cazzaniga, M., & Li, L. (2023). Labor market exposure to AI: Cross-country differences and distributional implications. International Monetary Fund.

Rane, J., Kaya, O., Mallick, S. K., & Rane, N. L. (2024). Smart farming using artificial intelligence, machine learning, deep learning, and ChatGPT: Applications, opportunities, challenges, and future directions. In J. Rane, O. Kaya, S. K. Mallick, & N. L. Rane (Eds.), Generative artificial intelligence in agriculture, education, and business (pp. 218-272). DeepScience.

Trajtenberg, M. (2018). AI as the next GPT: A political-economy perspective. Working Paper No. 24245. National Bureau of Economic Research.

This study was published as part of the Empowering Canada’s workforce research program under the direction of research director Ricardo Chejfec, supported by vice president of research Rachel Samson and former research director Natalia Mishagina. The manuscript was copy-edited by Justin Yule, proofreading was by Zofia Laubitz, editorial co-ordination was by Étienne Tremblay, production was by Chantal Létourneau and art direction was by Anne Tremblay.

Matthias Oschinski is a Senior Fellow at Georgetown University’s Center for Security and Emerging Technology (CSET) and the founder of Belongnomics. At CSET he leads the institute’s workforce line of research. His research primarily focuses on the impact of emerging technologies on labor and skills, as well as inclusive innovation.

Ruhani Walia is an incoming researcher in the Model Development and Research Division at the Bank of Canada. Her work has involved empirical economics and applied data science, focusing on the implications of emerging technologies. She has contributed to research at institutions including the Stanford Digital Economy Lab, the Los Angeles Behavioural Economics Lab, and the University of Toronto.

To cite this document:

Oschinski, M. & Walia, R. (2025). Harnessing Generative AI: Navigating the Transformative Impact on Canada’s Labour Market. IRPP Study No. 97. Montreal: Institute for Research on Public Policy.

---

The opinions expressed in this study are those of the authors and do not necessarily reflect the views of the IRPP or its Board of Directors.

IRPP Study is a refereed monographic series that is published irregularly throughout the year. Each study is subject to rigorous internal and external peer review for academic soundness and policy relevance.

If you have questions about our publications, please contact irpp@nullirpp.org. If you would like to subscribe to our newsletter, IRPP News, please go to our website, at irpp.org.

Photo: Istock.com

ISSN 1920-9436 (Online)

L’IA générative pourrait modifier les emplois au Canada, pas seulement les remplacer

Montréal — Une nouvelle étude de l’IRPP explore l’impact potentiel de l’intelligence artificielle générative sur la main-d’œuvre canadienne au cours des cinq prochaines années. Certaines tâches peuvent être à risque sans qu’un poste en entier soit appelé à disparaître. En fait, avec le bon soutien et de la panification, de nombreux emplois pourraient bénéficier de l’IA.

À l’aide de ChatGPT et de la base de données du Système d’information sur les professions et les compétences du gouvernement fédéral, les auteurs de l’étude, Matthias Oschinski et Ruhani Walia, ont évalué la capacité technique de l’IA générative à effectuer différentes tâches. Leurs conclusions suggèrent que ce n’est pas parce que l’IA peut accomplir une tâche qu’elle va nécessairement remplacer un emploi en entier.

« L’impact de l’IA sur le travail dépend de bien plus que la technologie elle-même. Les entreprises ont également besoin d’infrastructures adaptées, de capitaux, d’autorisations légales et d’une capacité à aller de l’avant au sein de l’organisation. Cela signifie que de nombreux emplois ne sont menacés que si ces autres éléments sont réunis », explique M. Oschinski.

L’étude a identifié trois grandes tendances :

• Ce sont certaines tâches qui sont à risque, pas des emplois en entier : Les tâches de bureau et de traitement de données présentent le plus grand risque d’automatisation, mais il est peu probable que l’IA se substitue à la totalité de ses emplois. Elle risque plutôt de changer comment le travail est fait.
• Certaines compétences sont moins menacées : Les emplois qui reposent sur les relations interpersonnelles, comme l’enseignement, les soins ou les postes de direction sont beaucoup moins à risque d’être automatisés et pourraient même bénéficier des outils d’IA.
• L’impact potentiel varie par région et par industrie : Certaine régions, comme les territoires nordiques et certaines industries, comme le transport et l’entreposage, pourraient connaître des perturbations. D’autres, comme l’éducation, sont plus à l’abri.

Les auteurs invitent les gouvernements et les employeurs à adopter une approche proactive. Pour tirer le meilleur parti de l’IA – et pour protéger les emplois – ils recommandent des investissements ciblés dans l’acquisition de compétences, en particulier pour les capacités que l’IA ne peut pas reproduire, comme la résolution de problèmes, la collaboration et les tâches de direction d’employés. Ils recommandent également aux gouvernements de tenir compte des différences régionales dans les stratégies de main-d’œuvre et d’investir dans les données et l’infrastructure nécessaires pour tirer le meilleur parti de cette technologie en évolution rapide.

« L’IA générative pourrait être un outil puissant pour améliorer la productivité du Canada. Mais elle ne s’implantera pas toute seule. Nous avons besoin d’une action coordonnée pour former la main-d’œuvre adéquate et veiller à ce que les avantages soient partagés », ajoute Mme Walia.