← aitransitionworkforce.com · Executive summary (PDF) · Français
A six-part summary of the proposal. Open any drawer to read the support behind a headline.
AI is moving into Canadian workplaces faster than the safety net can adapt. About 13.1 million Canadian jobs, roughly two in three, are in occupations exposed to it. Canada's labour data shows strain in hiring, participation and youth unemployment, but it cannot yet show whether AI is the cause.
The AI Workforce Transition Fund (AWTF) would raise roughly $250 to $330 billion over ten years from corporate levies on AI-driven displacement, productivity gains, profits and AI vendor revenue (Base case, illustrative, not yet costed by the Parliamentary Budget Officer). About $204 billion of that does not depend on the Productivity Dividend, the least certain pillar. A federal bridge of up to $65 billion is held as a contingency. It plans for a defined number of displaced workers and lets that number move with five dials. Each worker gets income support, retraining, health support and start-up help over about five years, and those who cannot return to work get a permanent floor. An independent panel verifies who qualifies.
AI is set to change Canada's job market sharply within years. The data so far is mixed, and Canada cannot yet measure AI's role.
About 13.1 million Canadian jobs, roughly two in three, sit in occupations with significant exposure to AI. The framework's working hypothesis is that, initially, the loss arrives as roles that are never filled, which headline statistics cannot see, and that layoffs follow later. Waiting for the numbers to confirm the problem means acting after it has peaked.
Administrative and clerical workers (about 3.3 million, 16% of the workforce) are the most exposed group. Statistics Canada ranks clerical and administrative occupations first for generative AI exposure. They are concentrated in insurance, banking, government, professional services and retail, where document AI, scheduling tools and language models are being deployed fastest. Part 2 lists all six groups behind the 13.1 million.
Women are more exposed to AI than men, mainly through clerical and administrative work. Women hold 63.4% of high-exposure jobs (Future Skills Centre, 2026), and, across many countries rather than Canada alone, 29% of female-dominated occupations are exposed against 16% of male-dominated ones (ILO, 2026). Job losses do not yet follow that pattern. Statistics Canada finds no persistent decline in exposed jobs for either women or men through December 2025, and its October 2026 follow-up finds no broad evidence of substantial falls in hiring through August 2026. The Bank of Canada, however, finds the unemployment-risk gap between fully exposed and unexposed workers widening from 1.9 to 2.8 percentage points, driven by weaker job-finding rather than layoffs. The earliest US losses are among workers aged 22 to 25, including in male-heavy software roles. The first wave is likely to fall on tech, manufacturing and operations roles, and the clerical wave that follows is likely to fall more on women. Young and entry-level workers lose out in both. No Canadian study measures AI exposure for immigrants or Indigenous workers, so the framework asks the AI Impact Assessment Panel to track displacement by group.
A role is eliminated when an employee resigns or retires and AI absorbs the work instead of a new hire. No notice is filed and no headcount cut is announced. The Labour Force Survey records the same employment it would show if the job had simply stayed vacant. Canada has no direct way to measure it: no statistic counts roles that were never filled because AI absorbed the work. What the framework can observe are the signs it would leave. Youth unemployment is 13.0% (September 2026), and unemployment among workers aged 55 and older rose 0.2 percentage points to 5.3% in the month (CP24, citing the September Labour Force Survey). Both have other explanations, including the cyclical slowdown and more people looking for work, so they are consistent with the hypothesis but do not prove it. Statistics Canada's research so far does not show it. Its January 2026 note found no faster decline in exposed occupations through December 2025, and its October 28, 2026 study found hiring fell across all occupation groups, fastest in the least exposed. The Bank of Canada's August 2026 analysis points the other way, with weaker job-finding in exposed occupations, and calls its results early signals. The evidence so far is mixed. The framework expects layoffs to follow as adoption deepens. The Displacement Levy covers that later, visible stage, and the Profitability Proxy covers the silent stage in the meantime. Working hypothesis, not yet observed
What would count against it. The hypothesis predicts that exposed, low-complementarity occupations show weaker hiring and job-finding than low-exposure ones, and that the gap in unemployment risk between exposed and unexposed workers keeps widening. If, over the next six monthly updates of the Canary dashboards and the Statistics Canada hiring series, hiring in exposed occupations holds up as well as in low-exposure ones and the gap stops widening, the framework would treat the hypothesis as unsupported and revisit the design for the silent stage. The Canary dashboards are updated monthly, and the next update follows the November 6, 2026 releases.
The entry level shows it first. Data entry, scheduling and basic document work once gave new graduates and recent immigrants their first experience. Employers now ask for two to three years of experience for jobs once classed as entry level. Canada's job vacancy rate fell from 3.1% (Q3 2024) to 2.8% (Q3 2025), consistent with employers who are neither firing nor hiring.
The framework uses US outcomes as a leading indicator for Canada, with a stated lag of 12 to 18 months. That lag is the framework's own working assumption and an adjustable parameter, not a measured constant. TD Economics (January 14, 2026) found that AI adoption is broader in the US and wrote that "the U.S. offers a glimpse of the future as the rate of adoption expands in Canada." TD also found that Canadian employment in high-exposure, low-complementarity occupations has so far been more resilient than in the US, and that Canada's cooling labour market is mainly cyclical. The Bank of Canada (May 2026) ranks clerical work as facing the highest automation risk but calls the current effect transformation rather than elimination. The framework treats that as a question of timing.
Recovery headlines are when the appetite for structural reform fades. The June 2026 jobs gain of 18,000 came entirely from accommodation and food services, an industry with many low-exposure jobs such as chefs and bartenders, while manufacturing, the public sector and workers over 55 lost jobs. September's fall is a stronger version of the same pattern. The cost of acting early is a levy design that may prove larger than needed. The cost of acting late is a wave of displacement with no institution ready to catch it.
Up to 13.1 million jobs are exposed. The AWTF plans for between about 3.7 million (Base case) and 7.9 million (Stress case) displaced workers over ten years.
In the Base case about 1.2 million workers are displaced by Year 5 and 3.7 million by Year 10. Five dials change that number, so the fund's size follows from stated assumptions.
Illustrative. Calculated with the cohort Pillar 1 method, Pillar 3 on the prior year's gain, the levy starting in Year 1, and a placeholder assumption that 35% of Pillar 1 payers also owe Pillar 2. Not costed by the Parliamentary Budget Officer. Scenarios are not probabilities.
Why Part 2 uses 13.1 million. It is a different measure from Statistics Canada's figure, and the two are not directly comparable. Statistics Canada's ~60% is its own exposure estimate, split roughly equally between roles AI may compete with and roles it may complement. The 13.1 million is the framework's at-risk table: six occupational groups totalling about 64% of the model's ~20.6 million workforce base (the Labour Force Survey shows about 21.1 million employed in May 2026). Both are about exposure, not job loss.
| Occupational group | Est. workers | Share |
|---|---|---|
| Administrative and clerical | ~3.3M | ~16% |
| Retail and hospitality | ~3.2M | ~16% |
| Transportation and logistics | ~2.7M | ~13% |
| Mid-skill manufacturing | ~1.6M | ~8% |
| Customer service and call centre | ~1.5M | ~7% |
| Junior knowledge workers | ~0.8M | ~4% |
| Total at significant risk | ~13.1M | ~64% |
A note on the retail and hospitality group. This group is defined by sector, so it mixes jobs of different exposure. Statistics Canada classes retail salespeople as highly exposed and cashiers, bartenders and chefs as low-exposure. The framework keeps the 13.1 million until occupation-level data can size the exposed part of the group clearly. To be rebuilt from occupation data when available
Anthropic's paper has cognitive employment 3.9% lower in its "substantial" scenario and 21.5% lower in its "extreme" one. Cognitive workers are 62.4% of employment, so that is about 2.4% and 13.4% of the workforce (our arithmetic). Checked The AWTF Base reaches 6.0% by Year 5, about 2.5 times the substantial level and under half the extreme level. Elevated (8.7%) and Stress (12.7%) approach the extreme level. The Base case is a deliberately above-substantial planning case. Two cautions: the AWTF counts physical sectors that Anthropic excludes, and Canadian adoption lags the United States.
Four corporate levies pay for a ring-fenced fund. A federal bridge of up to $65B stands ready as a contingency and is not part of the headline.
The levies track what each firm gains or causes: workers it displaces, productivity it captures, profit where attribution is unclear, and revenue from selling AI into Canada.
| Pillar | Ten-year total (working) | How it is built |
|---|---|---|
| 1. Displacement Levy | ~$159B | Each displaced cohort pays a fraction of salary for three years, then stops |
| 2. Productivity Dividend | ~$128B, less credit | Half of a firm's Pillar 1 payment is deducted from its own Pillar 2 bill, capped at that bill |
| 3. Profitability Proxy | ~$23B | Over ten years, charged on the prior year's gain |
| 4. Vendor Access Levy | ~$22B | Over seven years, on a $12B Year-1 revenue base |
| Levy base | $253B to $332B | Lower figure if every Pillar 1 payer also owes Pillar 2 |
How the numbers are built. Each displaced cohort pays Pillar 1 for three years and then stops. The levy starts in model Year 1, the year the law takes effect. The law itself would need to address displacements and gains from the previous full year, so the framework does not claim to be free of retroactive effect. The working figures have not yet been adjusted for that prior year. These are ESGAI working figures, not costed by the Parliamentary Budget Officer.
Method: the cohort method, the 50% credit (half of what a firm pays in Pillar 1 reduces its own Pillar 2 bill, capped at that bill), and Pillar 1 as the largest pillar. Headline: "roughly $250 to $330 billion in the Base case" is a range, not a guaranteed minimum. Pillars 1, 3 and 4 give about $204 billion, and reaching $250 billion needs almost all of Pillar 2 even at the highest overlap, while recent evidence does not show the productivity gains Pillar 2 assumes. Stated assumption: 35% of Pillar 1 payers also owe Pillar 2. This is a placeholder pending Parliamentary Budget Officer costing. At 0% overlap the levy base is about $332B, at 70% about $277B, and at full overlap about $253B.
Pillar 2 is the least certain pillar. Recent studies have not shown productivity gains in line with Pillar 2 expectations. Gains are positive but small and do not yet appear in aggregate data. The framework expects continued observation of productivity per employee to show over time whether the gains arrive. Until then the Pillar 2 figure of about $128B is an illustration of what verified gains would yield, not a forecast.
What Pillar 1 does not charge. Pillar 1 is triggered by verified displacement, so a role left unfilled after an AI-enabled resignation or retirement does not trigger it. Pillar 3, charged on profit gains, is the framework's answer for that silent stage. The six-month test in Part 1 will show whether that is enough. The closing section sets out requests to government for a better way to measure it.
Refusing to disclose does not lower the bill. A firm that declines to report its AI-attributable gains is charged the Pillar 2 rate on its whole profit gain. Pillar 3's 18% applies only to firms that disclose and where the AI Impact Assessment Panel finds attribution cannot be established. The working figures do not yet model this rule. Not yet costed
Who pays. The levies apply to firms with more than $500 million in Canadian revenue. Firms that retrain their own staff early earn voucher credits against their contribution. The working figures do not yet reflect either the threshold or the credits. The number of firms in scope and the cost of the credits are not yet estimated
A draw authority for Years 1 and 2, when setup costs are highest. It is repaid from levy surpluses so the fund stands alone from Year 4. In the Base case the levy covers the modelled requirement and the bridge is a contingency. In the Stress case the bridge does not cover the gap. The bridge is not drawn in the Base case. If it is drawn, it is repaid from levy surpluses, and repayment is not part of the $189B Base-case requirement. The headroom described above is larger than the full $65B.
Why the Base case raises more than it spends. In the Base case the levy raises about $304B against a programme requirement of $189B, which leaves about $115B of headroom over ten years. Most of it comes from Pillar 2, the least certain pillar: Pillars 1, 3 and 4 alone raise about $204B, which still covers the Base requirement. The headroom is the margin for error if Pillar 2 delivers less than assumed or if displacement runs above the Base case, and it is what would repay the bridge if it were ever drawn. Levy rates are reviewed against actual displacement from Year 5 and at each regular legislative review, which is where a lasting surplus would be returned or the rates lowered.
Every verified displaced worker moves through three stages, from stable income to a new job, a business, or a permanent floor.
The AI Impact Assessment Panel verifies each case. Support is generous at first and tied to activity as it goes.
| Stream | What it does | Est. annual cost |
|---|---|---|
| 1. The AI Workforce Transition Fund | Collects and holds the levies; Panel verifies cases | Admin $300M to $600M |
| 2. UBI, UBS, UHI pathway | Income, services, re-entry | $8B to $12B + $3B to $5B |
| 2b. Stage 3B permanent floor | $33,000 to $36,000 a year for workers whose occupation no longer exists | ~$1.3B by Year 5, ~$5.6B by Year 10 |
| 3. National Retraining Program | Catalogue of validated modules of 18 months or less; 70% placement within 6 months or lose accreditation; SME vouchers | $2B to $4B |
| 4. Future Skills Lab | 10 to 15 university research centres, early-warning system, quarterly reports to Parliament | $500M to $1B |
| 5. Mental Health and Purpose | Up to 26 therapy sessions a year with no diagnosis needed, peer networks, family support | $1B to $2B |
| 6. Entrepreneur Acceleration | $1,500 monthly top-up for 24 months, $10,000 to $25,000 micro-grants, regional hubs | $800M to $1.5B |
| 7. Human Reserved Jobs | Regulatory designation run by the Panel. Outline stage, with open federal-provincial and legal questions. | No levy draw |
Two cost lines to note. The cost table carries a line for the permanent floor, which reaches about $1.3B a year by Year 5 and $5.6B by Year 10. The $8B to $12B income line assumes an average support period of 5 to 7 months, in line with the Base dial's 6 months. Twelve months is the maximum, not the average. Over the ten years the floor costs about $21.6B in the Base case (5% of each displaced cohort, at $34,500 a head, no indexing), and it is already counted in the $189B Base-case programme requirement, so it is paid from the levy income in Part 3. At the Elevated and Stress settings the floor share rises to 8% and 12%. Funding for the floor beyond Year 10 is left to the regular reviews described in Part 3. Working estimate, not costed by the Parliamentary Budget Officer
Six objections, each with a short answer. The full replies are one click away.
The levy hits firms with over $500 million in Canadian revenue, whose profits depend on the Canadian market. Firms that retrain their own staff early earn voucher credits against their contribution.
Pillar 1 is charged on verified displacement, Pillar 2 on AI-attributable gains, and Pillar 3 on profit gains where attribution cannot be established. None of them is charged on total revenue, so a firm's bill follows what AI does to its workforce and profits. Only Pillar 4 is charged on revenue. No revenue-share figure is given until Parliamentary Budget Officer costing can support one.
The Future Skills Lab updates the retraining catalogue every six months from real employer demand. Doing nothing costs far more.
Retraining providers must place 70% of graduates within six months or lose accreditation. Results are audited yearly and scores are published.
Support is built as a ladder. Stage 2 is tied to activity, and re-entry pays more than staying on support. Unconditional income lasts 12 months at most (about 6 on average in the Base case), and only workers the Panel has verified as AI-displaced qualify.
The Mincome experiment in Dauphin, Manitoba (1970s) is the Canadian precedent, and it did reduce work somewhat. Hum and Simpson's study of the Winnipeg site reports hours down about 1% for men, 3% for married women and 4% for unmarried women. Calnitsky and Latner (Social Problems) find that labour market participation in Dauphin fell 11.3 percentage points relative to the Manitoba control group, concentrated among young and single-headed households, and conclude that a moderate guaranteed income does not cause a collapse in the formal labour market. Riddell and Riddell (2020) are cited as finding larger hours reductions among married women with children. Separately, Evelyn Forget, University of Manitoba, found that hospitalizations declined relative to the control group and that teenagers stayed in school longer (CBC); economist David Green disputes the causal reading of the hospitalization result (as cited on Progress.org). Mincome was an experiment, not a law, and the framework's unconditional income is limited to 12 months, so the results are a guide, not a forecast. Work-hours figures read through a text extract of Calnitsky and Latner; confirm against the full paper. Riddell and Riddell not checked. A secondary source gives 5% for unmarried women.
It adds to them for one case they were not built for: structural AI displacement. Provincial cost-sharing would avoid overlap. Stage 1 pays more than the Employment Insurance maximum for a limited period, and a two-tier safety net is a fair concern. The payment level is a design parameter that the Parliamentary Budget Officer costing should test.
Employment Insurance assumes the same job returns. Provincial welfare can trap people in poverty rather than lead them to new work.
This is the biggest external risk, and it falls on one pillar. Pillars 1 to 3 treat every firm the same whatever its home country.
On the working figures, Pillars 1 to 3 supply about $231B to $310B, 91% to 93% of the levy base (our arithmetic). In the Base case, where the requirement is $189B, the fund still works without Pillar 4. Trade-law advice is a condition for tabling it.
Verification is the hardest problem in the framework, and the page does not hide it. The AI Impact Assessment Panel verifies each displacement case, the first request in the closing section is for employers to report AI-attributed discharges, and Pillar 3, charged on profit gains where attribution cannot be established, catches gains that go unreported.
Pillar 1 (about $159B) is charged only on verified displacement, so it depends on this evidence. The working figures assume full reporting and do not model firms classifying exits as something else. Until a reporting requirement exists, the Pillar 1 figure illustrates what verified cases would yield, not a forecast. The Parliamentary Budget Officer costing should test how firms would respond. Not yet modelled
The $304B is a gross working figure. It does not yet reflect the $500 million revenue threshold, the retraining credits, or displaced workers whose employers fall outside the levy. The real base will be lower, and the Parliamentary Budget Officer costing is the way to size it.
The $204B from Pillars 1, 3 and 4 is also a figure before the threshold and the credits, so the page does not claim that the Base case still covers its $189B requirement once they are applied. Who ultimately bears the levies (firms, shareholders, workers or customers) is a question for the same costing. Not yet modelled
September reversed it: Canada lost 68,000 jobs, 48,000 of them young workers. The June gain had come entirely from accommodation and food services, an industry with many low-exposure jobs such as chefs and bartenders.
In June the 18,000 gain came entirely from accommodation and food services, while manufacturing, the public sector and workers over 55 lost jobs (Statistics Canada, July 10, 2026). September's fall makes a recovery reading harder to hold. Part of the education and public-sector losses is policy-driven, from international student permit cuts, so the objection is answered by the pattern across months, not by one month.
What experts say and what the employment data shows, collected in one place.
Three groups: the two Canary dashboards, research and official data, and public statements by leaders. Each entry says whether it is a direct quote or a paraphrase.
Both dashboards sort sectors by Statistics Canada's exposure index: HELC (high exposure, low complementarity, so replacement risk) and HEHC (high exposure, high complementarity, so augmentation). A HELC sector that falls by 10k (Canada) or 75k (USA) triggers a displacement signal.
Read with care. The dashboard says September's education and public-sector losses are partly policy-driven and should not be read as AI displacement alone. The Canadian sector table still shows August data.
"Even under the best circumstances, the transition to this new AI era will be one of the most turbulent times in human history."Bill Gates, "A turbulent AI era and critical choices to make", Gates Notes, August 2026. Quote checked against press reproductions of the essay (see sources).
Canada cannot yet see AI-related job loss, so the first request is to measure it.
No law requires an employer to say that it discharged an employee, or retired a role, because of AI. Official statistics therefore cannot attribute any job loss to AI, and the levy cannot be verified without that evidence.