Five AI Moonshots for Canada

In September 2025, the federal government convened an AI Strategy Task Force for a 30-day sprint to shape Canada's AI approach. Several members of the Build Canada network served on the task force and submitted memos addressing themes central to Build Canada: reducing friction for entrepreneurs, reforming government incentives, building the conditions for Canadian companies to compete globally, and creating moonshots that would transform the economy. Following submissions, the government requested a 30-day hold on publication, which has now passed. As such, Ajay is sharing his memo to contribute to the public conversation on how Canada can lead in AI.

Summary

Setting a target to become the "Best AI country in the world" is meaningless. AI is a tool — successful adoption will be defined not by quantity of use but by applying that tool to problems with measurable outcomes. The questions around the roll out of AI are not whether Canada leads in AI per se, but whether Canadians get faster healthcare, better education, and stronger defense.

The way to approach this is by setting aspirational, measurable, high impact goals i.e. moonshots. Moonshots work because they force system-wide change. A single bold target—like cutting healthcare wait times by 90%—requires redesigning every part of the system with the latest technology to reduce redundancy and increase productivity. Governments are uniquely positioned to lead these efforts because they can coordinate across agencies, provinces, and sectors in ways the private sector cannot.

Ajay’s moonshots have five qualities: they are ambitious, benefit many Canadians, require AI to succeed, build infrastructure other projects can use, and cannot be delivered by the private sector alone.

If Canada were to achieve the five moonshots, we would become a world leader in these services and transform our infrastructure for adopting AI:

• Healthcare: Cut referral-to-treatment time by 90% using AI for triage, scheduling, and early detection.
• Education: Reduce Grade 3 reading failure by 90% through adaptive screening, personalized tutoring, and teacher support tools.
• Natural Resources: Shrink wildfire and oil spill response time by 90% using sensor networks, drone verification, and AI dispatch.
• Defence: Reduce time to detect and classify airspace or maritime incursions by 90% through multi-sensor fusion and automated tracking.
• Homelessness: Cut the time from first contact to stable housing by 90% by integrating shelter data, benefits systems, and housing matching.

A Modest Proposal for Canada: Five AI Moonshots 

I was once leading a strategy meeting with the CEO and senior leadership team of a large airline. During the opening segment, one of the executives said “We’d like to be the best AI airline in the world.” I replied “I have no idea what that means. If you said that you’d like to be the airline with the most on time arrivals, or the least lost luggage, or the highest cargo yield, or the safest, or the most profitable, then I could suggest ways to use AI to help you meet any one of those objectives.” AI is essentially computational statistics that does prediction. Whether we use AI to read healthcare records and medical images to make a diagnosis (health condition prediction), operate a robot (control system prediction), or fill out a form (next best word prediction) - there is no ghost in the machine, it’s all statistics¹.

I’d like Canada to be the best AI country in the world. But I have no idea what that means. However, I do have a sense of what it means to have excellent healthcare, highly effective education, well-stewarded natural resources, a world-class defense infrastructure, and a well-functioning system to support citizens and residents when they are at their lowest. And I know that AI can be used as part of the solution to deliver those things at a higher quality and lower cost than was ever possible before. Furthermore, in building the systems required to deliver the capabilities I describe below (like a 10x reduction in healthcare wait times “Median days referral-to-definitive-treatment”) we will, in the process, build the infrastructure to subsequently support many other solutions with a very high social and economic return on investment (ROI) because the foundational components will have already been built so the marginal cost of each new solution will be relatively low.

I emphasize that AI is only part of the solution because to really move the needle in domains like healthcare, education, and defense, we require system-level rather than point-level solutions². A point solution is a bolt-on AI model that improves one narrow decision inside an existing workflow while leaving upstream and downstream processes unchanged (e.g., fraud detection at a bank - a point solution replaces old predictive analytics with higher fidelity AI-based prediction of fraud without changing any of the workflows at the bank); it’s easier to pilot but its impact is capped by surrounding bottlenecks and policies. A system solution reorganizes the entire decision flow to capitalize on cheap, accurate prediction—redesigning roles, handoffs, incentives, data rights, IT, and governance so that when the model outputs a prediction, the rest of the system can act on it immediately. That requires complementary investments (data pipelines, integration, automation, safeguards, measurement), takes longer to stand up, and faces more coordination risk than a point solution—but it has the potential to unlock a much greater gain because the prediction is embedded in a re-engineered system rather than taped onto an old one³. Governments have a comparative advantage in overseeing system-level solutions because they are especially well placed to manage coordination risk.

I propose five moonshots as the centerpiece of Canada’s AI strategy. Each moonshot is defined by a singular objective like reducing healthcare wait times by 10X. A singular objective can drive profound system-level redesign. For example, Estonia’s “once-only” design rule meant agencies couldn’t ask people to re-enter data the state already held, so services were forced to pull verified information automatically via the shared data infrastructure (“X-Road”) instead of pushing paperwork back onto citizens. The result was shockingly faster delivery (from taxes to prescriptions), fewer errors and fraud from re-keyed data, and staff time refocused on edge cases rather than routine verification. It also enabled proactive benefits—e.g., a single life event like childbirth triggers bundled notifications and enrollment—dramatically cutting administrative burden and boosting take-up. In short, “once-only” turned digital government from nicer websites into end-to-end automation with much higher performance, higher productivity, higher satisfaction, lower costs, and better integrity. 

These moonshots have five key attributes. 1) They are ambitious. If we achieve the specified targets, then Canada would become the best in the world in the specified domains. 2) They benefit a large number and a broad range of Canadians. 3) They utilize the power of AI; these objectives would not be possible without employing AI. 4) They drive the creation of core infrastructure that can then be leveraged by many other private- and public-sector initiatives that could have a very high social and economic return on investment⁴. 5) They could not be delivered by the private sector as they rely on the comparative advantage of the government because they require significant coordination across many spheres of influence⁵. I put significant thought into selecting these five domains. At the same time, there may be domains and/or performance metrics that better meet these five key objectives. The proposed moonshots could likely be significantly improved with the benefit of feedback and iteration with experts in the domains they address. So, these moonshots should be taken as directional rather than as prescriptive government funding. In every case, I asked myself if this is such a great idea, then why won't the private sector fund it? Why should tax payers pay for this? Sometimes, there was a good answer - the product was a public good or the service generated significant spillovers such that the public benefit was much greater than the private benefit so private markets would underinvest. Overall, my view was that many proposals I received were either not sufficiently transformational for Canada given the magnitude of the moment or they did not sufficiently answer the question, “if this is such a great idea, then why won’t the private sector pay for it?”

Moonshot #1: Drastically reduce health-care wait times and increase access 

The first moonshot is focused on healthcare. It is a results-oriented push to expand timely access and is measured by a single north star: median days referral-to-definitive-treatment (RTT). We aim to cut RTT by 10× (e.g., 120 → 12 days) across multiple high-volume pathways (e.g., oncology starts, elective cardiology such as PCI stent procedures and catheter ablation for atrial fibrillation, hips/knees, cataracts) in two provinces, integrating provincial eReferral and EMRs, diagnostics, and OR scheduling into a coherent, instrumented flow. This mission is achievable only through the aggressive use of AI—for AI triage and imaging, eReferral routing and load-balancing, OR/block optimization, virtual pre-op, and AI documentation—and the equally essential human and institutional inputs: clinical teams, operational redesign, data-sharing agreements, privacy/security compliance, and bilingual patient communications. Much higher fidelity early detection that drives prevention that reduces treatment loads will be a critical part of this solution. AI is a critical accelerator, necessary but not sufficient, driving prevention (early detection, risk prediction, adherence and self-management) and treatment (triage, routing, scheduling) in equal measure, within a coordinated system that keeps people healthier, reduces demand on specialists and ORs, and—when treatment is needed—moves patients to definitive care faster while preserving safety and quality.

The table below sketches ten illustrative AI applications that could help compress RTT end-to-end; they’re examples, not prescriptions. The team leading this mission is expected to determine which specific AI tools are most critical in their operating context. Just as important: AI is only one input in a broader system that also relies on excellent clinical practice, streamlined operations, interoperable data, regulatory and privacy compliance, and disciplined execution. 

Moonshot #2: Radically improve education by drastically reducing Grade 3 reading non-proficiency rate 

Canada’s education moonshot is a results-oriented push to lift early literacy—anchored on a single north star: Grade 3 reading non-proficiency rate (lower is better). The objective is a 10× reduction (e.g., 30% → 3%) across participating districts in at least two provinces/territories, with sustained gains and transparent reporting. It is achievable only through the aggressive use of AI—adaptive K-3 screening, precision tutoring minutes, teacher copilots that lighten cognitive load, and attendance nudging—and the equally essential human and institutional inputs: expert teachers, evidence-based instruction, interoperable data, privacy/consent frameworks, and school/board operating discipline. AI is the critical accelerator—necessary but not sufficient—within a coordinated Canadian system that delivers more effective instruction to every learner, earlier and more often. 

The table below sketches illustrative AI components that can compress time-to-proficiency and lower non-proficiency rates; they are examples, not prescriptions. The delivery team is expected to determine which specific AI tools are most critical in their context, and to remember that AI is only one input in a broader system that also relies on people, data, and operations.

Moonshot #3: Radically enhance natural resource management by drastically reducing median detection-to-containment time (D2C) for wildfires and oil spills 

Canada’s natural-resources moonshot is a results-oriented push to cut the time from event onset to verified containment—anchored on the north star median D2C (lower is better). The objective is a 10× reduction (e.g., 10h → 1h) across at least two resource domains (wildfire ignition and oil/chemical spills) in different Canadian regions, with sustained performance and transparent reporting. By compressing D2C, we reduce loss of life and property and mitigate risks to the environment—protecting communities, critical infrastructure, and ecosystems. This mission is achievable only through the aggressive use of AI—multi-sensor anomaly detection, lightning-to-ignition nowcasts, drone tasking, crew pre-positioning, playbook generation, and incident-command copilots—and the equally essential human and institutional inputs: detection networks and aircraft, trained crews and incident commanders, interoperable data, permitting and regulatory readiness, and bilingual operations. AI is a critical accelerator—necessary but not sufficient—within a coordinated Canadian system that verifies faster, moves assets earlier, and starts containment sooner.

The table below sketches illustrative AI components that can compress D2C end-to-end; they are examples, not prescriptions. The delivery team is expected to determine which specific AI tools are most critical in their operating context, and to remember that AI is only one input in a broader system that also relies on people, data, and operations.

Moonshot #4: Radically enhance defence by drastically reducing time-to-detect-and-classify (TTDC) incursions (air/maritime/space) at a specified probability of detection 

Canada’s defence moonshot is a results-oriented push to cut the time from first sensor contact to correct classification—anchored on the north star coverage-weighted TTDC (lower is better). The objective is a 10× reduction (e.g., 20 min → 2 min) across defined Canadian approaches and domains (air, maritime, space) at a specified probability of detection (Pd) and controlled false-alarm rate (FAR). Achieving this requires the aggressive use of AI—multi-sensor fusion, on-edge recognition, “dark target” detection, cross-sensor/interceptor tasking, and watchfloor copilots—and equally essential institutional inputs: interoperable sensing networks (including NORAD partner feeds), secure data pathways, cleared personnel, robust testing/evaluation, and bilingual command-and-control. As with the other moonshots, AI is the critical accelerator—necessary but not sufficient—within a coordinated Canadian system that verifies faster, cues assets earlier, and classifies with confidence so action can be taken sooner. 

The table below sketches illustrative AI components that can compress TTDC end-to-end; they are examples, not prescriptions. The delivery team is expected to determine which specific AI tools are most critical in their operating context and to remember that AI is only one input in a broader system that also relies on people, data, and disciplined operations.

Moonshot #5: Drastically Reduce the Harmful Impact of Homelessness in Canada 

Canada’s fifth moonshot on homelessness is a results-oriented push to restore dignity and stability—measured by a single north star: median days-from-first-contact-to-stable-housing (DFCSH). The mission is to cut DFCSH by 10× (e.g., 300 → 30 days) with 180-day retention, operating across three cities (Toronto, Vancouver, Montreal) and integrating by-name lists, shelters, provincial benefits, ID services, and inspection/landlord workflows⁶.  This mission is achievable only through the aggressive use of AI—for predictive triage, AI-assisted unit/benefit matching, process mining, and retention risk nudging—and the equally essential human and institutional inputs: skilled and empathetic frontline teams, reliable data, provincial/municipal agreements (HIFIS interoperability, ID/benefits), and strong landlord partnerships. In parallel, mission success will likely require also employing AI for prevention—spotting and addressing antecedents such as untreated mental illness, addictions, family violence, and income shocks—so fewer people become unhoused in the first place. AI is a critical accelerator, a necessary but not sufficient input for mission success, along with other key inputs in a coordinated national system designed to move people into a home faster and keep them housed.

The table below sketches ten illustrative AI applications that could help slash median days-from-first-contact-to-stable-housing; they’re examples, not prescriptions. The team that leads this mission is expected to determine which specific AI tools are most critical for success in their operating context. Just as important: AI is only one input in a broader system that also relies on skilled and empathetic frontline staff, strong landlord partnerships, available housing, interoperable data (e.g., by-name lists/HIFIS), streamlined benefits and ID workflows, and rigorous operational execution. 

People are understandably worried that the increased use of AI will eliminate jobs. Many feel that way because of our lack of understanding, imagination, and ambition. It’s true that if our ambition was to maintain our current access level to healthcare, then we could potentially use AI to enhance productivity and reduce jobs. But that is not the ambition represented in this proposal. The ambition is to 10X healthcare, 10X education, 10X natural resource management, 10x defense, and 10X our responsiveness to the unhoused. Without AI, we could not afford to pursue these objectives. With AI, we have a fighting chance. We will likely need to increase jobs in order to achieve these objectives. However, we will not need to 10X labor to realize a 10X increase in performance. In other words, AI will be used as a tool to increase the productivity of people, increasing the returns to labor and thus potentially creating more jobs, not less⁷. 

Implementation 

As this 30-day memo process is the first step in a process, I will leave implementation details for a subsequent step. However, I will offer three guiding principles for implementation: 1) Engage the creativity and leadership of the private, public, and non-for-profit communities in Canada. Specify the requirements for each moonshot in a Request for Proposal (RFP) with a meaningful budget and unleash the creativity of Canadians to propose different approaches to deliver the solution. Choose one or two or three at the outset based solely on their likelihood of success with zero weighting on any other social distributive objectives. 2) Commit a budget to each moonshot that recognizes the importance of the moment. Recall that economies that are significantly smaller than Canada are competing more aggressively (e.g., UAE economy is about a quarter the size of Canada (~$0.5T), Kingdom of Saudi Arabia is about half the size of Canada (~$1.0T) yet last year, in 2024, both committed 50x more to AI than Canada (both committed ~$100B compared to $2B)). For example, if we budgeted $5B for this moonshot strategy, $1B per moonshot, then although the amount seems large relative to normal expenditures, it would represent only 20% more than last year’s expenditure by a single AI startup (Anthropic, ~$4B expenditures in FY2024). We risk falling behind if we don’t recognize that we’re in a unique moment in history. Disburse funds on a performance/milestone basis. Pay for performance, not effort. 3) Ensure lessons learned and operating procedures developed for each of the moonshots are well documented and fully shared at no cost so that we can implement them broadly across the country as fast as possible. Report progress against the key metric regularly and publicly. 

Increasing Returns to Experimentation 

Moonshot success will require a liberal use of regulatory holidays and sandboxes. Given the objective is five national moonshots, the scarce input is not ideas but evidence about what actually drives outcomes. In these mission areas, the social value of information is high: targeted experiments convert uncertainty into guidance for where to deploy capital, talent, and compute at scale to achieve both productivity and the designed social outcomes. The organizing principle is simple: run short, decision-oriented trials where policy choices are imminent, publish what we learn, and use the results to prioritize the next tranche of investment. 

A powerful general-purpose technology like frontier AI creates a supply shock of cheap, fast, cognition-like machine intelligence that leaves us unsure which redesigned workflows deliver the highest ROI and safety in practice. This elevates the option value of learning: local, well-identified trials reveal which tasks should be reclassified as AI tasks, what complements (data access, evaluation pipelines, liability frameworks) are required, and which deployments produce measurable gains. Without this evidence, the private sector won’t finance complements, regulators won’t clarify rules, and customers won’t pay. 

Markets alone will under-invest in experimentation because information from pilots spills over to competitors and regulatory uncertainty amplifies the risk, especially in sensitive domains. Policy should therefore raise the return to evidence generation: create regulatory sandboxes for high-information pilots, offer narrowly scoped regulatory holidays for parallel, low-risk trials, tie public support to transparent reporting of performance and failures, and retire exemptions as results arrive. The goal isn’t a free pass—it’s faster feedback loops that move validated solutions from local pilots to national scale. 

Sovereign AI 

The issue of AI sovereignty has been raised by our government on several occasions, so I briefly address it here. Sovereignty is the nation-state’s ultimate power and authority to govern people, resources, and activities within its territory free from external coercion. Our AI strategy must enable and encourage the rapid adoption of AI to drive high-impact productivity gains without increasing risks to our sovereignty. So, AI sovereignty is not about building everything at home; it’s about preserving Canadian optionality so we can deploy critical AI systems on our terms while compounding productivity gains. That means three things: first, turn AI into a sustained engine of national wealth by redesigning workflows so cheaper prediction translates into measurable productivity growth; second, guarantee assured domestic access to critical inference so essential workloads in health, defence, finance, and public services keep running even if foreign providers falter; third, secure bargaining power by becoming a top-tier producer of at least one indispensable AI-system component (e.g., energy for data centres, a critical mineral, or advanced chip-packaging). We should track progress with a concise cabinet dashboard—productivity lift, Canadian-controlled backup inference capacity, and our global share of a chosen supply-chain component—while sequencing policy over near-, mid-, and long-term horizons to lock in resilience, prosperity, and security. I submitted a separate document on this topic for Minister Solomon on July 1 2025. 

My Background 

I offer a summary of my background to put my recommendations in context. I am an economist. I have studied the economics of innovation for the past twenty five years and the economics of AI specifically for over a decade. In addition to writing scholarly journal articles on AI, I coauthored two best-selling books on the economics of AI, “Prediction Machines” and “Power & Prediction” both published by Harvard Business Press, and co-edited four scholarly books on the economics of AI, “The Economics of AI: An Agenda,” “The Economics of AI: Healthcare Challenges,” “The Political Economy of Artificial Intelligence,” and “The Transformative Economics of Artificial Intelligence,” all published by the University of Chicago Press. I am the founder of what is, to my knowledge, the largest startup program in the world focused on the commercialization of science that was the first in the world to begin a program dedicated to the commercialization of machine learning-based startups in 2015: the Creative Destruction Lab. This program admits approximately 800 startups a year at sites around the world, and graduates have created over US$50B in equity value. I’m a cofounder of Intrepid Growth Partners, which allocates capital to scale growth-stage AI-application-based companies, and I am a cofounder of Sanctuary, a company racing to create the world’s first human-like intelligence in general-purpose robots. I have led AI strategy sessions for over two dozen Fortune 500 companies over the past five years. I have contributed to policy-related research on AI in a range of institutional settings: as a Research Associate at the National Bureau of Economic Research in Cambridge, MA; an Academic Advisory Council Member at the Center on Regulation and Markets at the Brookings Institution in Washington, DC; an Advisory Board Member for Carnegie

Mellon University’s Block Center for Technology and Society in Pittsburgh, PA; a Research Fellow at the Stanford Digital Economy Lab in Palo Alto, CA; and a Faculty Affiliate at the Vector Institute for Artificial Intelligence in Toronto, ON.

If these ideas resonate and you want help to build Canada sign up to join the movement today to be the first to hear about events and volunteer opportunities.

Footnotes

¹ The language of predictive or discriminative versus generative AI can be the source of some confusion. It is useful to first distinguish between predictive and generative tasks. For example, a predictive task might be predicting the binding efficacy of some molecule with a target protein; a generative task might be to generate a molecule that binds with that protein, which would typically involve sampling from an appropriate conditional distribution -- itself a form of prediction. A second useful distinction is between discriminative and generative models. Typically, discriminative models are used for predictive tasks and generative models for generative tasks. However, predictive tasks could be pursued directly using a discriminative model or indirectly using a generative model and applying Bayes Rule. Similarly, generative tasks could be pursued directly using a generative model or indirectly using a discriminative model (e.g., by ranking molecules for testing based on the outputs of the discriminative model). I use the term prediction as a shorthand for both types of models as applied to both types of tasks.

² Joshua Gans, Avi Goldfarb, and I wrote a book that explains the key attributes of system versus point solutions: “Power & Prediction: The Disruptive Economics of Artificial Intelligence” (Harvard Business Press, 2022).

³ Consider navigational AIs like Waze or Google Maps. When these were first invented, they could have been brought to market by licensing the technology to taxi companies. That would have increased the productivity of taxi drivers. That type of implementation would be a point solution. It would have increased the productivity of taxi drivers without requiring any significant process changes to the taxi industry. In the City of London, for example, taxi drivers have to go to school for three years to learn “The Knowledge” where they study maps and routes in order to be granted a license. Even London taxi drivers, especially those who are new on the job, could be more productive with some help from a navigational AI that predicts the optimal route given real time traffic data. This would enhance driver productivity without requiring process innovation. 

Alternatively, with a fully automated navigational AI, people who have never set foot in London can now fly into Heathrow, rent a car, and navigate the City as well as a professional driver. That enabled the creation of companies like Uber and Lyft. None of the ride-sharing companies could exist without navigational AIs. Before Uber, there were approximately 200,000 professional limousine and taxi drivers in the US. Today, about four million people drive for Uber in the US. If the average value of a car they drive is $25,000, then this unlocked $100 billion in underutilized capital assets that were otherwise sitting in people’s garages or car parks. These companies totally transformed the transportation industry. Uber, Lyft, and the others like them represent a system-level solution in contrast to a point solution. Many process innovations are required to make these solutions work.

⁴ Mission success will require the creation of shared national core infrastructure—a standardized data and identity fabric that links program data across jurisdictions (e.g., eReferral/EMR, HIFIS and by-name lists, school assessments, environmental and sensor feeds) using privacy-preserving record linkage, explicit consent/audit trails, open schemas, and public APIs. This will have to be paired with a secure compute and AI layer (cloud + edge) featuring certified model registries, evaluation/monitoring pipelines, red-teaming and incident response, plus governance toolkits (regulatory sandboxes, templated DPAs, procurement patterns) and bilingual UX components and training curricula. Finally, it will require field infrastructure—standardized sensors and comms (EO/SAR, drones, radios, diagnostics), reference workflows, and open playbooks—so future public and private initiatives can build faster on proven rails at far lower marginal cost.

⁵ When the task force membership was announced, I was inundated with recommendations, suggestions,and pitches by well-meaning Canadians, many of whom had great ideas and some had even already builtgreat products. Most made a case for why their project should receive government funding. In every case,I asked myself if this is such a great idea, then why won't the private sector fund it? Why should taxpayers pay for this? Sometimes, there was a good answer - the product was a public good or the servicegenerated significant spillovers such that the public benefit was much greater than the private benefit soprivate markets would underinvest. Overall, my view was that many proposals I received were either notsufficiently transformational for Canada given the magnitude of the moment or they did not sufficientlyanswer the question, “if this is such a great idea, then why won’t the private sector pay for it?”

⁶ Those experiencing prolonged instability spent a median of 304 nights in shelter from 2010 to 2021." Prolonged instability is defined as accessing shelters at least once per year over three consecutive years,ad indicating extended time before achieving stable housing. https://housing-infrastructure.canada.ca/homelessness-sans-abri/reports-rapports/chronic-homelessness 2017-2021-litinerance-chronique-eng.html

‍

⁷ See my paper with Joshua Gans and Avi Goldfarb, “The Economics of Bicycles for the Mind,” National Bureau of Economic Research, Working Paper # 34034, 2025.