Why Mapping Education to Work Is Harder Than It Looks
Students often ask, "What jobs could this program lead to?" Workforce planners flip the question. They ask, "What programs are feeding talent into this job?"
Questions like this matter:
- to students choosing what to study,
- to institutions reviewing their programs,
- to policymakers linking education to employment,
- and to employers examining talent pipelines.
But they’re harder questions to answer than they sound.
Canada already has the information on both sides of the question. The Classification of Instructional Programs (CIP) organizes post-secondary programs by field of study, while the National Occupational Classification (NOC) organizes occupations by the work involved and the training, education, or experience needed. Both are useful on their own. What’s missing is a bridge between them—one clear, current framework connecting specific programs to the occupations they could lead to.
LMIC is fixing that problem. We are working on closing that gap and exploring if large language models (LLMs) can make national CIP-to-NOC mapping practical at scale.
Why would a national program-to-occupation mapping be useful?
Program-to-occupation mapping could support decisions across education and the labour market. It could help:
- students and families explore the range of work connected to a field of study
- post-secondary institutions understand where graduates can apply what they learn
- policymakers identify education and training programs that could respond to labour market needs
- employers better understand which programs can produce workers with the knowledge and skills that are relevant to their needs
For example, if a region is experiencing a shortage of software developers, mapping can help identify the education and training programs that contribute to the talent pipeline. A post-secondary institution reviewing a computer science program can use the same mapping to understand the types of occupations connected to that program, and how that reflects current labour market demand. A student or employee can use it to find occupations they could transition to with minimal upskilling.
The mapping wouldn’t replace career guidance, local knowledge, or data about graduate outcomes. It would be a shared infrastructure to help people combine sources to ask better questions that reflect their specific needs.
Existing work shows that the connection is useful. BC Stats publishes data that relate fields of study to graduates' occupations. In the U.S., federal statistical agencies publish a crosswalk between instructional programs and occupations. Users can apply the U.S. crosswalk to education planning, labour market analysis, and career exploration.
These examples do not provide a ready-made Canadian framework. But they show that the need is real and the idea is feasible.
Why is it difficult to map education programs to occupations?
Mapping education programs to occupations is difficult because there is no single way to define what a meaningful relationship between them should represent.
Some connections are direct. A licence or certification may require a particular type of education. But many other pathways depend on work experience, further training, personal circumstances, and opportunities that arise over the course of a career.
Consider software development again: people may enter the field after studying computer science, software engineering, mathematics, statistics, or a related subject. However, a mathematics program may support pathways into analysis, finance, research, technology, education, or further study. In comparison, occupations like optometry or dentistry may require specific credentials or licences.
The relationships between NOCs and CIPs form a network, not a list of one-to-one matches. People also change direction as they gain experience or training-so while a person’s education shapes their options, it doesn’t determine their future.
That means there are several ways to describe connections. A program could be relevant to an occupation, provide typical preparation for it, or meet a formal entry requirement—but these use cases represent different types of connections. Without a clear definition of the relationship, two teams of researchers could examine the same CIP-NOC pair and reach different conclusions.
This tells us that a useful mapping should show likely connections and plausible pathways without prescribing a person's choices or suggesting that a credential guarantees a particular job.
A national mapping requires more than one million judgments
At their most detailed levels, the 2021 editions of the CIP and the NOC contain 2,119 classes and 516-unit groups. With this granularity, multiplying every distinct CIP class and NOC unit group would result in 1,093,404 possible CIP-NOC pairs.
Even if someone could assess one pair every minute, an initial review would take more than 18,000 hours. That would leave no time for research, difficult cases, quality checks, or revisions.
In addition, the work does not end after the first pass. Given that Statistics Canada updates classifications to reflect changes in programs and occupations, a useful mapping must preserve earlier decisions, account for revised codes and descriptions, and remain consistent over time.
A comprehensive mapping, then, is not simply a large spreadsheet project. It is an ongoing classification and evaluation problem.
Why can’t simple rules solve the mapping problem?
Keyword matching looks for shared words in titles and descriptions. It can identify obvious similarities. However, shared words do not prove a meaningful relationship. Closely related programs and occupations may use different language. Similar wording can also hide important differences in content or requirements.
Adding fixed keyword matching rules to the process does not remove the need for judgment. A rule can work for a familiar case and fail when a connection is indirect, multidisciplinary, or context-dependent.
Experts can make these distinctions and judgments. The challenge is asking them to make these across more than one million possible pairs, and then to repeat the work as the classifications change.
Could large language models help solve the scale problem?
LLMs could help address the scale problem by comparing context across large volumes of text. This makes them promising tools for assessing program-to-occupation relationships when keyword matching is insufficient and manual review cannot cover every possible comparison.
An LLM-assisted system could help examine possible relationships faster and more consistently than a fully manual process. It could also direct expert attention toward uncertain or difficult cases, where human judgment adds the most value.
That does not mean we should ask a general-purpose chatbot for a list and treat the response as fact. LLMs can be inconsistent, sensitive to wording, and confidently wrong. A model’s answer is not evidence simply because it sounds persuasive.
The opportunity lies in combining the reasoning capacity of LLMs with a consistent process that human experts can review. That could make a Canada-wide assessment practical at a cost and speed that manual evaluation cannot match.
The hard work sits around the model
Access to an LLM is only one part of the solution. A credible mapping also depends on clear definitions, dependable source information, consistent treatment of ambiguous cases, expert checks, and a clear way to communicate uncertainty.
Governance matters, too. The resulting data cannot be allowed to become career instructions or hiring requirements. How limits are defined and communicated will affect whether mapping can support real decisions responsibly.
Scale adds another challenge. Comparing a handful of programs and occupations can demonstrate an idea, but applying the same standard across more than one million pairs requires reliable infrastructure, quality controls, and ongoing maintenance as classifications, evidence, or models change.
That work draws on different kinds of expertise. Technical systems need to be built and maintained, program-to-occupation relationships need to be defined and reviewed, and uncertain cases need human judgment. A credible mapping is a shared resource that a team maintains over time, not a one-time AI exercise.
The significant financial and technical resources needed make this task one that requires organizational-level support. Otherwise, a project of this scale could quickly drain an organization of the limited resources it has to tackle issues that it may be better positioned to address.
What is LMIC doing?
LMIC is testing whether an LLM-assisted approach, paired with expert human judgment and a consistent framework, could support a national resource that is credible, explainable, and maintainable.
Our role in Canada’s LMI ecosystem gives us a national perspective on the need for consistent, accessible information and on the different ways in which governments, post-secondary institutions, researchers, and other organizations may use it.
We are not announcing a finished product. The work is evolving, and technology is only one part of the approach. LLMs can extend expert capacity, but cannot replace human expertise or clear limits.
Building a national mapping will take more than technology
Connecting Canada’s education and occupational classification systems could produce better LMI for education planning, career exploration, policy analysis, and workforce development. LLMs may make it possible to tackle the scale of that challenge in a new way, but the value of a national mapping will ultimately depend on its credibility and whether it is responsive to the people who use it.
As this work develops, we want to hear from organizations that work with education-to-employment pathways, graduate outcomes, and workforce planning related to LMI. How would you use a national program-to-occupation mapping? What would it need to do well to be useful in your work? Connect with us at research@lmic-cimt.ca.