The Amplified Economy

A Measurement and Learning Framework for Quality-Adjusted AI Adoption, Productivity and Human Capability

Policy framework | Package 0.5 Discussion Draft | September 2026

Proposed first action: Amplification Initiative — Newfoundland and Labrador Full Learning Pilot

Executive proposition

Canada should aim to become unusually capable at converting artificial intelligence potential into measurable gains in productivity, innovation, public value and human capability.

Canada’s 2026 National Artificial Intelligence Strategy, AI for All, establishes adoption, skills, safety, sovereign infrastructure, commercialization and public-service transformation as national priorities. It targets business AI adoption of 60 per cent by 2034. The Amplified Economy complements that strategy by asking a different question: which forms of adoption produce quality-adjusted value, where do constraints move, and what should be replicated, redesigned or stopped?[12]

This framework supplies a shared measurement and learning layer for that purpose. It would support bounded experiments in real workplaces, measure economic and human results, identify where constraints move after AI is introduced, replicate credible findings and diffuse practices that survive independent scrutiny. Public support would be staged against evidence. Weak cases would stop early; promising cases would earn replication and deployment support. Its economy-wide value would depend on whether those local findings can reduce duplicated experimentation, improve investment decisions and spread quality-adjusted gains across organizations and sectors.

The Amplified Economy does not claim that human–AI collaboration, workflow redesign or experimental adoption are individually new. Its proposed contribution is to integrate established research, emerging evidence and testable hypotheses into a public measurement and learning system that compares configurations in real workplaces, identifies where constraints move, replicates credible findings and determines what should be stopped, redesigned or scaled.

The policy is deliberately ambitious about the opportunity and conservative about what is known. Evidence shows that AI can produce large gains in some tasks and settings, especially where workers have suitable knowledge, workflows are redesigned and outputs can be checked. The same evidence shows null effects, hidden quality losses and slower performance in other settings. Perceived time savings can diverge from measured results. Task-level improvement does not automatically become firm-level productivity, better jobs or national advantage.

The proposed first action is an 18-month Full Learning Pilot in Newfoundland and Labrador. It would run a capacity-gated portfolio of up to 36 workplace experiments, with independent evaluation and an expected gross budget of approximately $3.68 million, within an authorization ceiling of $4.00 million. Newfoundland and Labrador would serve as the first provincial testbed for the national framework—not as the limit of the policy.

The immediate policy decision is to:

1. endorse the Amplified Economy as a complementary national measurement and learning approach for quality-adjusted AI adoption, productivity and human capability;
2. authorize federal–provincial design work for a Canadian amplification learning system and common measurement framework;
3. advance the Newfoundland and Labrador Full Learning Pilot as the preferred first implementation, subject to host, procurement and funding validation;
4. establish independent evaluation, worker participation, responsible-use controls and public learning as conditions of support; and
5. require a national scale, narrow or stop decision after the first portfolio produces sufficient evidence.

1. Canada’s economic and technological moment

Canada enters the next phase of AI adoption under pressure to increase productivity, strengthen business investment, diversify trade and improve the capacity of public services. The federal Spring Economic Update reported real GDP growth of 1.7 per cent in 2025 and the first increase in real GDP per capita after declines in 2023 and 2024, while also noting that trade uncertainty continued to weigh on exports and business investment. The economic outlook remains exposed to tariffs, geopolitical risk and a reconfigured global trading system.^[1]

At the same time, governments and technology firms are committing extraordinary capital to the supply of AI capability. The European Union launched an initiative intended to mobilize €200 billion in AI investment; the Stargate consortium announced an intention to invest US$500 billion in American AI infrastructure over four years; and UN Trade and Development projects that the global AI market could reach US$4.8 trillion by 2033.^[9][10][11] These figures combine intended investment, mobilized capital and projected market size—not money already spent—but they show the scale and direction of the global commitment.

Workplace adoption is also accelerating. Statistics Canada reported that the share of Canadian businesses using AI to produce goods or deliver services rose from 6.1 per cent in the second quarter of 2024 to 12.2 per cent in the second quarter of 2025 and 19.2 per cent in the second quarter of 2026. Adoption remains highly uneven: it is much higher in information, finance and professional services than in construction and agriculture.^[2]

Yet investment in technical capability does not automatically produce economic value. The capacity of organizations to redesign work, allocate judgment, verify outputs, redeploy saved time and absorb new capabilities is developing much more slowly. AI capability is advancing at model speed; organizational capability is advancing at institutional speed.

The result is an AI capability-capture gap: a widening difference between what increasingly capable systems can theoretically do and what organizations can reliably convert into productivity, innovation, public value and stronger human capability. The opportunity is not untouched—large gains already exist in some settings—but it is only partially and unevenly captured.

The central policy question is therefore no longer whether AI will be used. It is whether Canada can become unusually effective at closing this capability-capture gap faster, more deliberately and more broadly than it otherwise would.

The capability-capture gap creates a distinctive policy opportunity. Five features define where that opportunity lies and how Canada could pursue it.

1.1 The capability-capture gap is an economic opportunity

The technical frontier is moving quickly, but most organizations still apply AI to fragments of existing work. They often add a tool without redesigning the surrounding workflow, clarifying decision rights, improving data, creating verification capacity or changing how saved time is used. Capability can therefore rise at the task level while organizational output remains unchanged.

Canada is unlikely to win a spending race for the world’s largest models or compute clusters. It can compete in the underdeveloped demand-side discipline of applied discovery: learning how to identify, test, measure and diffuse high-value human–AI configurations, and developing the interaction tools and organizational methods needed to make those configurations usable at scale.

1.2 AI can reduce the cost of trying

AI can make research, drafting, coding, modelling, translation, design, analysis and prototyping cheaper and faster. This may allow workers and small organizations to test ideas that were previously uneconomic. The result could be more than efficiency: new products, improved services, broader participation in innovation and the ability to cross practical skill boundaries.

Lowering the cost of an attempt does not guarantee useful innovation. It can also produce more low-quality output, duplicated effort and review burden. The opportunity arises only when experimentation is paired with selection, measurement, stopping rules and learning.

1.3 Bottleneck migration creates a constraint-discovery opportunity

When AI accelerates one activity, the constraint may move to data access, management attention, expert review, physical testing, approvals, regulation, commercialization, procurement or capital. A policy focused only on the AI tool may record an apparent gain while missing the new binding constraint that prevents value from being realized.

The opportunity is to become better at finding and relieving moving constraints. Detecting the next bottleneck allows complementary investment to target real outcomes instead of scaling activity without value. Constraint discovery can therefore become a productive capability in its own right.

1.4 Scarce human judgment creates an interaction-design opportunity

AI can generate options and outputs faster than people can frame, inspect, compare and absorb them. This creates attention debt and verification debt: more material exists, but the capacity to decide what deserves attention and whether it is correct does not keep pace. The productivity of human judgment—not the volume of AI output—may become the decisive constraint.

When an AI-assisted project advances faster than participants can absorb its reasoning history, another burden emerges: velocity-induced context debt. Decisions, rejected paths, unresolved disagreements and evidentiary judgments accumulate faster than collaborators can reconstruct them. This can make a nominally productive project difficult to govern, transfer or scale.

The opportunity is to design workflows and tools that reserve human attention for the decisions, uncertainties and exceptions where it produces the most value. As generation becomes abundant, effective attention allocation, verification and context management may become differentiating capabilities—and areas in which Canadian organizations can develop reusable methods and products.

1.5 Institutional learning offers Canada a plausible competitive path

Canada’s potential advantage is not that it is inherently better at AI adoption. It is that the country could organize its distributed institutions as a learning system. Federalism, diverse regional economies, public services, research organizations and sector networks create varied settings in which configurations can be tested. The same dispersion can produce fragmentation; it becomes an advantage only if common methods allow local results to be compared, replicated and translated.

Canada does not need to lead every layer of model development to capture substantial value. A country that learns more quickly which configurations work, what complementary changes they require and how to diffuse them can capture value from globally available technologies. The proposed advantage is therefore institutional: a higher rate of applied learning, adaptation and absorption.

Practices will spread internationally, so a permanent national moat should not be assumed. The realistic objective is to improve Canada’s rate of learning, adaptation and diffusion while retaining more of the resulting commercial, worker and public value.

1.6 From workplace learning to economy-wide value

The five opportunities above form a conditional economic pathway. Economy-wide value will not arise simply because individual workers complete tasks faster. It would arise if organizations become better at identifying configurations that produce real gains, stopping weak deployments, relieving complementary constraints and converting released capacity into higher-value output, improved services, new products, shorter queues or lower cost.

The national mechanism adds value when validated learning travels. Shared measurement reduces repeated trial and error; replication identifies which results survive changes in worker, organization, sector or region; and diffusion lowers the cost of adoption for organizations that could not independently fund rigorous experimentation. Recording null and negative findings can also improve allocative efficiency by preventing public and private investment from scaling activity that does not create quality-adjusted value.

Over time, these channels could contribute to higher productivity, greater innovation and service capacity, stronger workforce capability and a broader distribution of AI-related gains. They could also strengthen Canada’s absorptive capacity: the ability of firms, public institutions and workers to recognize useful technical capability, adapt it to local conditions and continue improving as models change. These are plausible economic effects, not promised outcomes.

The Newfoundland and Labrador Full Learning Pilot cannot establish an economy-wide effect. It can test the early links in the chain: whether high-value configurations can be discovered and measured, whether constraints can be identified and relieved, whether findings replicate, whether shared evidence changes decisions and whether the learning mechanism produces value beyond ordinary adoption support. Broader economic claims should depend on later replication, diffusion and longitudinal evidence.

2. The implementation paradox

AI adoption can rise without a corresponding increase in measured organizational productivity. This is not necessarily evidence that the technology has little value. It may indicate that current implementation methods are aimed at the wrong unit of change or that benefits are being lost between the task and the organization.[6][8]

2.1 Tool deployment is not workflow redesign

Licences, training sessions and approved-use policies can increase access. They do not, by themselves, determine which work should remain human-led, which decisions require expert review, how data should enter a system, who is accountable for outputs, or how newly available capacity should be redeployed.

An AI assistant added to an unchanged workflow may save minutes for an individual while leaving queues, handoffs, approvals and incentives untouched. It may even increase downstream work by producing more drafts, code, analyses or proposals that require review.[3][4]

2.2 Output volume is easily mistaken for productivity

More emails, documents, code, designs or analyses are not automatically valuable. Useful measurement must adjust for quality, rework, errors, risk, timeliness and whether the output changes an operational result. It must also test additionality: whether the result would have happened without the intervention.

2.3 Perception is an unreliable proxy

Workers can reasonably feel faster because an AI system reduces effort or makes a task more pleasant. Yet controlled measurement can find smaller gains, no gain or even slower completion. Conversely, a learning period can feel slower while building a workflow that later produces value. Self-report is important evidence about experience and adoption, but it cannot substitute for operational measures.[5]

2.4 Task gains may be trapped inside the organization

Time saved does not become productivity unless the organization can redirect it to additional output, higher quality, shorter queues, new revenue, better service or reduced cost. Without demand, authority, complementary investment and changed performance systems, saved time may remain latent.[8]

2.5 The reliable frontier is jagged

AI capability is not uniformly distributed across tasks. A system can perform impressively on one part of a job and fail on an adjacent part that appears similar. Workers and managers may not know where that boundary lies. Effective implementation therefore requires empirical mapping of the frontier, safeguards outside it and frequent re-testing as systems change.[7]

The policy problem can be stated plainly:

Canada lacks a common, practical mechanism for discovering where human–AI configurations create measurable value, determining what complementary changes make that value deployable, and diffusing credible findings across organizations and jurisdictions.

3. Configuration, outcomes and amplification

Human–AI work should be described on separate dimensions. Configuration records how functions are allocated across human-only, human-led, collaborative, AI-led and automated modes. Outcome records what changed using the Type I–IV taxonomy in section 8.2. Amplification is an evidence-dependent magnitude designation: a material and disproportionate shift in the quality-adjusted production frontier relative to added human, technical and organizational resources. It can apply to efficiency, capacity, capability or creation. A multiple increase in quality-adjusted productivity can qualify without a new capability or creation effect.

Automation and augmentation remain useful descriptions of recurring working arrangements, but they are not outcome categories. Automation is appropriate where work is routine, risks are controlled and human participation adds little value; augmentation will be common where AI assists an existing role. Amplification is the underexplored policy opportunity and a stronger evidentiary designation. The search is for configurations that produce disproportionate quality-adjusted gains, including multiples of productivity in familiar work or previously impractical capability or creation. Automation and augmentation can contribute to such results; workflow redesign is a candidate mechanism rather than a prerequisite or proof. Novelty alone does not establish amplification. The policy challenge is to identify these situations in real work, determine which constraints suppress the gains and test whether both the result and its enabling conditions can be reproduced.

The framework rejects a universal rule that people must remain in every operational step. It instead adopts a Human–AI Configuration Principle: choose among human-only, AI-assisted, collaborative, AI-led and automated modes according to task characteristics, relative capability, risk, accountability and evidence.

Within that principle, the framework adopts a human-participation presumption as a tie-breaker under uncertainty: where expected value, risk and cost are materially equivalent, prefer the configuration that preserves meaningful human contribution and uses AI additively to extend human strengths. This is a policy preference, not an established economic result or a prohibition on automation. The separate, testable hypothesis is that purposeful participation may support capability, agency, well-being, adaptability and long-run productivity. Evidence of those effects informs the comparison; effort is not valuable for its own sake.

4. The hypotheses to test

The concepts below are working hypotheses, not settled doctrine. They are grouped by the role they play in the policy and pilot rather than by their broader theoretical importance. The core hypotheses require credible, decision-relevant evidence sufficient to determine whether the proposed framework warrants continuation; full replication and broader scale claims remain Stage 2 questions. The operational hypotheses shape how experiments should be designed, observed and adapted. The exploratory hypotheses address potentially important effects that may require longer observation, specialized measures or larger samples, and will be tested opportunistically where a credible design is available. This hierarchy concentrates evidentiary effort on the questions that must be answered while preserving structured opportunities to learn from the others.

Core hypotheses — require credible, decision-relevant evidence

4.1 Frontier capability is configuration-specific

Large gains are likely to arise in particular combinations of worker, task, workflow, AI system, data, incentives and organizational support. The correct unit of analysis is therefore the worker–task–workflow–AI configuration, not the tool alone.

The phrase frontier worker can be useful shorthand for a person who discovers or operates an unusually effective configuration. It should not imply a fixed elite or an innate category of employee. The same person may be at the frontier for one task and outside it for another. Frontier capability is relational, teachable in some cases and dependent on context.

Test: Can high-performing configurations be identified, explained and reproduced by other workers or organizations?

4.2 Workflow redesign creates more value than insertion alone

AI is frequently inserted into an existing process without changing roles, handoffs, data, quality controls or decision authority. The stronger hypothesis is that measurable value often requires redesign around the new capability.

Test: Compare tool insertion with configuration redesign where feasible. Measure quality-adjusted output, cycle time, rework, error, capacity and participant burden.

Operational hypotheses — shape experiment design

4.3 Middle-out workflow testing can reveal the appropriate division of work

Organizations often begin either from the top down—deciding in advance which tasks should be automated—or from the bottom up—allowing individuals to attach tools to existing tasks. Both approaches can embed untested assumptions about what people and AI should do.

The stronger hypothesis is that a hybrid workflow can be used as a discovery environment. A team begins with meaningful human and AI participation, decomposes the workflow into functions such as framing, generation, execution, evaluation, verification and authority, and then shifts those functions toward human-only, collaborative, AI-led or automated modes according to measured results. This middle-out configuration search does not assume that the hybrid is optimal. It is a practical way to discover unknown human capabilities, unknown AI capabilities and productive combinations that a fixed task allocation could miss.

Test: Compare multiple configurations against human-only and AI-dominant reference conditions where feasible. Record which allocation of functions produces the best quality-adjusted result, what previously unrecognized capabilities emerge, and whether the configuration can be explained and transferred.

4.4 AI expands the feasible search space

By reducing the cost of research, iteration and prototyping, AI may allow organizations to test more possible solutions and make previously marginal ideas viable. The value is not the number of attempts; it is the value discovered per dollar and per hour of scarce attention.

Test: Does the intervention increase the number and diversity of credible options, reduce time or cost to a validated result, or produce a solution that would otherwise not have been attempted?

4.5 Bottlenecks migrate

AI-enabled acceleration in one stage may expose a new constraint elsewhere. Discovering the new bottleneck is itself a useful finding because it directs complementary investment.

A practical response worth testing is to reinvest some AI-released upstream capacity in stronger triage, validation and preparation, so scarce downstream reviewers see fewer routine items and can concentrate on consequential or ambiguous decisions.

Test: Map the workflow before and after intervention. Track queue times, review demand, approval latency, data preparation, physical throughput, procurement, regulatory work and commercialization barriers.

Exploratory hypotheses — tested opportunistically

4.6 Human attention, verification and context alignment determine realized value

As generation becomes cheaper, framing, selection, review and accountability become more valuable. Poor attention allocation can erase productivity gains or create unsafe reliance. When project velocity exceeds the rate at which participants can build and maintain an adequate mental model, velocity-induced context debt may also slow onboarding, weaken challenge and concentrate effective control in the people who hold the history.

Test: Measure review time, error detection, correction yield, escalation, information overload, decision latency, onboarding time, ability to explain the project state and the share of generated work that is actually used.

4.7 Cooperation plus calibrated challenge improves judgment

AI can be configured to cooperate with a user while also testing assumptions, presenting alternatives and identifying weak evidence. Constant agreement creates over-reliance; constant contradiction creates friction without value. The hypothesis is that selective, evidence-grounded challenge improves decisions when it is material and well-timed.

Timing may be as important as the presence of challenge: supportive cooperation may expand the search space during brainstorming, while challenge becomes more valuable as work moves toward evaluation, commitment or high-consequence execution. Candidate triggers include a user-declared transition, material uncertainty, unsupported confidence, unresolved assumptions, rising consequence or diminishing novelty.

Test: Compare cooperative-only and cooperation-plus-challenge workflows. Track decision quality, missed risks, useful corrections, false reversals, trust calibration and time cost.

4.8 Active human contribution can produce earned amplification

The aim is not to preserve effort for its own sake. It is to preserve the high-value cognitive acts through which people develop understanding and remain capable of judgment: framing the problem, setting standards, evaluating evidence, integrating context and accepting responsibility. Clerical friction should be removed; productive difficulty should remain where it builds or exercises essential capability. This concept can be illustrated as the difference between using AI to make hard work easier, or instead using AI to make hard work more powerful.

Test: Compare configurations with different levels and types of human participation. Assess immediate output, retained understanding, ability to explain and adapt the work, error recognition, skill development and performance when the system is unavailable.

4.9 Organizational capability can compound

An organization that runs disciplined experiments may improve its data, management practice, workflow design, verification and adoption judgment. Those complements can make subsequent experiments faster and more productive.

Test: Track whether later experiments improve in selection quality, setup time, evidence quality, implementation speed and realized value. Treat any claim of a durable national advantage as unproven until replication and longitudinal evidence exist.

5. National policy objective and operating principles

5.1 Objective

Build a Canadian system that can discover, measure, support, scale and diffuse high-value human–AI configurations while protecting safety, rights, worker agency and public trust.

5.2 Operating principles

1. AI informs; people remain accountable. Decision rights must be explicit even when a system performs substantial work.
2. Configuration before technology. Start with a consequential problem and the surrounding workflow, not a preferred product.
3. Measure value, not activity. Licences, prompts, outputs and participation are inputs—not outcomes.
4. Quality-adjust every claim. Time or cost savings count only with evidence on errors, rework, safety and usefulness.
5. Explore cheaply, measure rigorously, stop quickly, scale aggressively. Uncertainty should determine the size and reversibility of an intervention, not force either paralysis or premature commitment.
6. Treat null and negative results as public assets. Credible evidence that a configuration does not work can prevent repeated waste or harm.
7. Preserve meaningful human capability. Remove low-value friction while testing for dependency, deskilling and loss of judgment. When credible configurations otherwise perform equivalently, prefer the one that preserves purposeful human contribution; automate when evidence shows a clear advantage in value, safety, reliability, accessibility or risk.
8. Measure who captures the gain. Productivity that worsens job quality, concentrates benefits or transfers unpaid review burdens is not a complete success.
9. Use independent challenge. Delivery teams should not have sole authority to validate their own claims.
10. Diffuse only what survives replication. Promising local results should be tested across people, settings or institutions before broad promotion.

6. The national amplification learning system

The proposed system is a recurring learning cycle rather than a one-time grant program.

Explore

Recruit consequential problems from businesses, workers, public services, Indigenous organizations, nonprofits and sector institutions. Favour questions with a credible operational owner, measurable baseline and plausible path to use.

Discover

Test multiple human–AI configurations. Identify frontier capability empirically, including where AI adds no value or performs worse than existing practice.

Amplify

Redesign roles, workflows, data, review and authority around promising findings. Support the complementary changes needed to convert a task result into operational value.

Measure

Assess economic, human and strategic returns against a baseline and comparator. Separate perceived gains from measured results and record uncertainty.

Remove bottlenecks

Identify the constraint that now limits value. Use bounded support, regulatory coordination, procurement, specialist review or capital where the expected learning or value justifies it.

Scale

Move validated configurations to an operational, commercial or public owner. Public support should decline as private value becomes clearer, except where the public return justifies continued investment.

Diffuse

Publish sanitized methods, failure modes, measurement templates and replication results. Report recipient readiness, adaptation and assistance costs alongside each transfer result, distinguishing independent reuse from transfer requiring coaching. These costs should inform expansion decisions. Protect privacy, security, Indigenous data governance, commercial confidentiality and intellectual property.

Compound

Use portfolio evidence to improve the next round of selection, measurement, workforce development, procurement and policy. Periodically retire practices invalidated by newer systems or evidence.

7. Policy instruments and additionality

The Amplified Economy should complement—not duplicate—existing Canadian investments in research, compute, commercialization and adoption. Its distinctive function is to provide a shared discovery, measurement and diffusion apparatus.

7.0 Relationship to AI for All and existing programs

AI for All already establishes national priorities for adoption, skills, trusted use, sovereign infrastructure, commercialization and public-service transformation. It targets an increase in business AI adoption from 12 per cent to 60 per cent by 2034 and commits $500 million to expand the Regional Artificial Intelligence Initiative.[12] The Amplified Economy does not duplicate those commitments. It adds a common, cross-sector mechanism for determining which adoption produces quality-adjusted value, how constraints migrate, which results replicate and which uses should stop.

The Regional Artificial Intelligence Initiative, delivered in Atlantic Canada through ACOA, supports AI adoption and commercialization. NRC IRAP AI Assist supports innovative small and medium-sized enterprises developing or incorporating AI, while Scale AI supports applied AI and commercialization in supply chains. The proposed framework is additional because it funds and governs comparative workplace experiments, independently evaluates economic and human outcomes, records null and negative findings and produces reusable evidence across programs and jurisdictions.

In Newfoundland and Labrador, workforce, entrepreneurship and university organizations could help recruit participating workplaces, study effects on workers or provide research support. Their roles would be assessed during pilot scoping. This framework does not assume that any organization has agreed to host or deliver the pilot.

7.1 Frontier discovery portfolios

Fund provincial, sectoral and public-service portfolios of bounded workplace experiments. Admission should depend on problem quality, evidence feasibility, responsible-use readiness and the likelihood that the result will inform a real decision.

7.2 A common measurement standard

Create a practical national protocol for baselines, comparators, quality adjustment, worker outcomes, additionality, uncertainty and knowledge disposition. The protocol should support multiple levels of rigour rather than require an academic trial for every case.

7.3 Replication and transfer capacity

Reserve funding and delivery capacity to repeat promising configurations in a different team, organization, sector or province. A result that cannot travel may still have local value, but it should not be treated as national evidence.

7.4 Constraint-removal support

Provide small, staged amounts for data preparation, integration, specialist review, security, accessibility, physical testing, regulatory navigation, procurement and other bottlenecks revealed by evidence. Support should solve a defined constraint, not subsidize indefinite implementation.

7.5 Applied human–AI tools and interaction infrastructure

Where commercially available products cannot adequately implement or test a promising human–AI configuration, the initiative may support bounded development of prototypes, interaction components and evaluation infrastructure. Eligible work may include cooperation-and-challenge modes, attention and verification interfaces, provenance and decision traces, workflow orchestration, context reconstruction and reusable evaluation harnesses.

Development should be hypothesis-driven and subordinate to measured worker, workflow and organizational outcomes. Experiments should normally begin with existing products or configurable tools; a prototype should be developed only when interaction design appears to be the binding constraint. Continuation should depend on comparative evidence—not the novelty or polish of the software.

Validated components may subsequently be standardized, procured, licensed or commercialized through Canadian firms and research institutions. Public-good methods and interoperability requirements should remain portable, while intellectual-property treatment should preserve incentives for Canadian commercialization and prevent vendor capture.

Conflict-of-interest and procurement rule: no tool, vendor or prototype owner may influence case selection, evaluation or procurement decisions involving its own product. Any pilot-developed tool must compete through an open, documented process against credible general-purpose and specialist alternatives. Evaluator and procurement authority must remain independent of development interests.

7.6 A Canadian amplification clearinghouse

Maintain a trusted repository of experiment designs, sanitized results, null findings, replication records, measurement tools and implementation patterns. It should make evidence reusable without exposing participant data or trade secrets.

7.7 Workforce and management capability

Develop training around problem framing, workflow mapping, evidence design, verification, human–AI configuration and accountable decision-making. Tool proficiency alone is insufficient. Managers need the ability to redeploy capacity, redesign work and judge where automation, augmentation or amplification is appropriate.

7.8 Public procurement and testbeds

Use public institutions as responsible testbeds and, where warranted, early customers for validated Canadian solutions. Procurement should reward measurable outcomes, interoperability, auditability and knowledge transfer—not merely the inclusion of AI.

7.9 Commercialization, trade and diffusion

Connect validated configurations and Canadian suppliers to commercialization programs, sector networks, export support and internal trade initiatives. The aim is to turn credible operational evidence into adoption, revenue and broader capacity while avoiding vendor capture.

8. Measurement and evidence architecture

8.1 Three return lenses

8.2 Canonical Type I–IV outcome taxonomy

• Type I — Efficiency: the same valuable result with fewer resources or less time.
• Type II — Capacity: more valuable work with the same constrained resources.
• Type III — Capability: a person or organization can perform work it could not reliably perform before.
• Type IV — Creation: a new product, service, process, market or public-value outcome becomes viable.

Portfolios should report the distribution across these categories rather than collapsing them into a single average. Classify outcome type separately from amplification magnitude: Type I or Type II results can qualify when quality-adjusted gains are material and disproportionate to added resources; Type III or Type IV results do not qualify solely because something new becomes possible. Report combined effects where appropriate without counting the same benefit twice.

8.3 Evidence proportional to the decision

Every experiment requires a baseline, a defined outcome, quality measures and a credible counterfactual. The method may range from structured before-and-after analysis to phased rollout, matched comparison, blinded review or randomized testing. Rigour should increase with the consequence and cost of the next decision.

Each experiment should pre-specify a Minimum Worthwhile Effect where feasible: the smallest improvement that would justify implementation after considering cost, risk and participant burden. This implementation threshold is distinct from the stronger amplification designation.

For an amplification claim, the measurement record should specify:

• the baseline and comparator, workflow boundary, observation period and quality standard;
• the outcome type or types, absolute change and, where meaningful, quality-adjusted productivity multiple;
• the human, technical and organizational resource basis, including setup, implementation, oversight, verification, rework, integration, compute, maintenance and other material lifecycle costs;
• a pre-specified, context-appropriate criterion for a material and disproportionate gain, with uncertainty reported;
• the level of demonstrated benefit—individual, team or workflow, and organization or service—and evidence of downstream absorption or bottleneck migration; and
• repeatability within the original setting and, separately, evidence of replication by other workers or organizations.

A verified individual or workflow gain can qualify at that level without demonstrating an equivalent organizational gain. Review or rework transferred to others must be counted: a local speedup erased across the relevant workflow is not realized amplification at that boundary. Conversely, a demonstrated local gain should be recorded even where a downstream constraint prevents its full organizational capture. For capability or creation outcomes without a meaningful numeric baseline, document feasibility, value and resource comparisons rather than inventing a multiplier.

8.4 Operational and knowledge dispositions

At each gate, the program makes two separate decisions.

Operational disposition:

• STOP — the premise, value, safety or feasibility no longer justifies investment.
• REDESIGN — the question remains valuable, but the configuration or evidence plan requires bounded revision.
• REPLICATE — the result is credible enough to test in another setting.
• SCALE–HANDOFF — evidence supports transfer to an operational, commercial or public owner.

Knowledge disposition:

• DIFFUSE — the finding and method are sufficiently credible and shareable.
• DO NOT DIFFUSE — the result is unreliable, unsafe or too context-specific.
• DIFFUSE WITH RESTRICTIONS — learning is useful but privacy, security, rights, commercial sensitivity or professional risk limits how it should travel.

8.5 Portfolio economics

Outcomes are likely to be skewed: many experiments will produce modest, null or negative findings, while a small number may create substantial value. The program should therefore be assessed as a portfolio.

Key measures include:

• cost per credible, decision-relevant finding;
• value discovered per public dollar;
• resources preserved through early stopping;
• share of positive findings that replicate;
• time from discovery to operational use;
• distribution of benefits and harms;
• participant and administrative burden; and
• improvement in later experiment performance.

The target is not a high percentage of projects labelled successful. Provider performance and compensation should reward sound investigation, usable evidence and timely decisions, including justified STOP and REDESIGN findings. They should not depend on positive findings, claimed savings or progression to scale. Evaluation authority must remain independent of delivery interests.

9. Governance and federal–provincial delivery

9.1 Federal role

The federal government should establish the common policy objective, measurement standard, co-investment rules, national clearinghouse and cross-jurisdiction diffusion mechanisms. It should connect findings to existing innovation, research, skills, procurement, commercialization and trade programs.

9.2 Provincial and territorial role

Provinces and territories should select portfolios that reflect regional industries, public-service pressures and institutional capacity. They should sponsor local delivery, connect participants to relevant authorities and help move credible findings into operational use.

9.3 Local and sector role

Municipalities, Indigenous governments and organizations, sector bodies, unions, post-secondary institutions, innovation organizations and community partners can identify problems, contribute specialist knowledge and support diffusion. Participation must respect jurisdiction, rights, consent and data governance.

9.4 Arm’s-length delivery

Individual portfolios should normally be operated by an existing arm’s-length host with credible finance, procurement, privacy, security, human resources, records and audit systems. Government should set policy, fund and oversee the portfolio without selecting individual cases or directing findings.

9.5 Independent evaluation

Evaluation should be separately protected from delivery authority. Evaluators need access to confidential evidence under appropriate controls, the ability to challenge additionality and causal claims, and the right to record unresolved disagreement. Program teams may correct factual errors but should not approve evaluative conclusions.

9.6 Workplace delivery and research-partner studies

The delivery host should own operational configuration search, bottleneck-migration tests and bounded replication. Research questions involving delayed recall, late-joiner context debt, challenge-mode comparisons or prototype comparisons should normally sit with a university research partner under separate research funding, protocol, consent and ethics review where required. Participation in the operational pilot must not automatically constitute participation in research.

9.7 Worker voice and distribution

Workers affected by an experiment should be involved in design and measurement. Existing labour, privacy, safety, accessibility, professional and human-rights obligations continue to apply. The framework should track how gains are divided among workers, organizations, customers, taxpayers and communities.

10. Proposed first action: Newfoundland and Labrador Full Learning Pilot

10.1 Strategic role

Newfoundland and Labrador is proposed as the first testbed because it combines manageable scale with meaningful diversity across small and medium-sized businesses, public administration, health, education, professional services, ocean industries, natural resources, tourism and rural or remote service delivery.

The province is large enough to test different configurations and small enough to connect government, industry, labour, post-secondary institutions and delivery organizations around a common evidence architecture. It also provides a demanding test of whether value can be created outside the largest technology and professional-services markets.

The pilot’s provincial purpose is to discover useful configurations and build local operating capability. Its national purpose is to test the policy mechanism, measurement framework, replication method and governance model before wider commitment.

10.2 Preferred design

The experiment count is a ceiling, not a quota. Fewer well-supported experiments are preferable to a larger portfolio with weak baselines, overloaded staff or unreliable conclusions.

10.3 Participant and problem selection

The pilot should recruit consequential problems rather than requests for technology purchases. Eligible settings may include firms, public bodies, nonprofits, Indigenous organizations and cross-sector partnerships.

Admission should require:

• an accountable operational owner;
• a defined problem and affected workflow;
• a credible baseline or feasible way to establish one;
• access to the people, data and decisions required for testing;
• readiness to address privacy, security, safety, labour and professional obligations;
• willingness to report null or negative findings confidentially; and
• a plausible route to use, replication or transferable learning.

Portfolio selection should balance sectors, organization sizes, task types, risk levels and expected result types. It should reserve capacity for replication and redesign rather than fill every place with a new idea.

10.4 Experiment charter

Before material implementation, each case should establish:

1. the operational problem and decision owner;
2. the baseline, comparator, workflow boundary, observation period and evidence method;
3. the proposed human–AI configuration;
4. the configuration-search plan, including human-only and AI-dominant reference conditions where feasible;
5. the intended economic, human and strategic benefit;
6. the Minimum Worthwhile Effect where feasible and, if amplification is claimed, a separate magnitude criterion and resource basis;
7. quality, safety, rights and distribution measures;
8. data, privacy, security, accessibility, professional and labour requirements;
9. any prototype trigger and the evidence required before custom development;
10. resources, specialist support and participant obligations;
11. stop, redesign, replication and handoff criteria; and
12. what may be shared publicly in sanitized form.

10.5 Delivery sequence

Mobilization — approximately three months

Select and contract the host; appoint the Portfolio Lead; finalize employment and procurement paths; procure independent evaluation; establish information controls; adapt the national measurement protocol; and open structured recruitment.

Wave 1 — early calibration

Launch a limited first group across more than one setting. Test participant burden, staffing ratios, measurement feasibility, review latency, procurement, risk triage and decision authority. Use middle-out configuration search where appropriate: begin with a hybrid discovery workflow, compare it with human-only and AI-dominant conditions, and move functions toward the mode supported by evidence. Do not open later waves simply because the calendar advances.

Capacity gate

An independent review should determine whether the host can support additional cases without weakening evidence, participant support or oversight. It should assess case concurrency, measurement quality, unresolved incidents, recruitment pipeline, spending and evaluator concerns.

Later waves and replication

Admit additional cases within demonstrated capacity. Use deliberately uneven support. Reserve places and funds for replication, transfer and removal of evidence-identified constraints.

Synthesis and handoff

Complete experiment dispositions, independent evaluation and sanitized portfolio reporting. Transfer deployable cases to appropriate owners. Produce a national recommendation to scale, narrow, redesign or stop the mechanism.

10.6 Core team

• Portfolio Lead: owns admissions, risk tiering, resource allocation, gates and STOP, REDESIGN, REPLICATE and SCALE–HANDOFF decisions within delegated authority.
• Two Measurement Leads: design proportionate evidence, challenge claims, monitor data quality and synthesize portfolio findings.
• Experiment and Configuration Facilitator: maps work, supports redesign and helps teams test alternative configurations.
• Program Operations Coordinator: manages agreements, scheduling, procurement flow, records, participant support and reporting discipline.

Senior judgment is essential. The lead and measurement positions should not be downgraded to protect the nominal experiment count.

10.7 Host and accountability model

The preferred host is an existing public agency, Crown entity or government-controlled delivery organization with compatible authority and mature operating systems. An established innovation or economic-development nonprofit may be selected if it better satisfies the mandatory criteria under a strong agreement. A new entity should not be created for the first pilot.

Host selection should use pass/fail capability gates covering legal authority, staffing, compensation, procurement, finance, data governance, security, records, audit, staged funding, conflict of interest, multi-funder administration and evaluator independence. Organizational names should follow confidential capability assessment rather than appear in the policy framework.

Provincial public-service experiments must follow the Government of Newfoundland and Labrador Responsible Use of Artificial Intelligence Technology policy and AI Assessment Process, ATIPPA, 2015, PHIA where health information is involved, applicable collective agreements and early union engagement. Federal participants must assess the federal Directive on Automated Decision-Making and Algorithmic Impact Assessment. [VERIFY] Confirm any current provincial vendor AI-use disclosure requirement before external circulation.

10.8 Budget and fiscal controls

Participant support should remain a centrally managed portfolio reserve rather than be divided into equal grants. Payments should be staged and released only where the next step can change a decision or produce transferable learning.

The initial commitment should cover mobilization, the operating base and Wave 1. Later-wave, replication and contingency funds should remain centrally held until capacity and evidence gates are satisfied. The $4.00 million ceiling should not be interpreted as an instruction to spend.

[AUTHOR DECISION: identify the proposed lead funder, federal/provincial split and approval authority. The operating design can combine provincial, federal or regional-development funding behind one participant-facing program, subject to eligibility and host validation.]

10.9 Options analysis

Option 1 — Status quo / existing programs only. New pilot cost: $0. Existing programs continue to support adoption, training, implementation and commercialization, but no common portfolio establishes comparable baselines, quality-adjusted outcomes, constraint movement, stop decisions or cross-case replication. Principal risk: adoption activity rises without a reliable account of which uses created value or should stop.

Option 2 — Minimum Credible Pilot. Gross planning budget: $2.25 million; authorization ceiling: $2.40 million; estimated net public cost: $2.19 million; 16 experiments within a 12–18 range; 4.0 FTE; 15 months. It can test whether the operating method is feasible, whether configuration comparison changes decisions and whether credible STOP/REDESIGN/REPLICATE findings can be produced. It cannot credibly estimate prevalence, compare sectors, characterize rare frontier results or support broad diffusion claims. Principal risk: a narrower mechanism test is mistaken for a full learning portfolio.

Option 3 — Full Learning Pilot, preferred. Gross planning budget: $3.68 million; authorization ceiling: $4.00 million; estimated net public cost: $3.54 million; up to 36 capacity-gated experiments; 5.0 FTE; 18 months. It adds portfolio breadth, cross-domain comparison, limited replication, bottleneck and distributional learning, and a stronger chance of observing rare high-value configurations. It still will not be a statistically representative economic study. Principal risk: recruitment, evaluator capacity or governance burden forces the portfolio below its ceiling; explicit Wave 1 gates manage that risk.

Decision rule: choose the Full Learning Pilot when the objective is decision-relevant portfolio learning. Choose the Minimum Credible Pilot only when the smaller commitment is a binding condition and explicitly withdraw the claims it cannot support.

10.10 Pilot success and failure

The pilot succeeds if it produces credible decisions and transferable learning—not only positive technology stories. A strong result may include a replicated capability gain, a safe and valuable automation, evidence that workflow redesign matters, identification of a removable bottleneck, or a well-supported stop decision that prevents larger waste.

The pilot should be judged unsuccessful if it cannot establish baselines, becomes primarily a grant-distribution exercise, suppresses null results, overloads its delivery team, lacks worker trust, permits delivery interests to control evaluation, or produces claims that cannot survive replication.

11. National pathway after the first pilot

National expansion should be conditional rather than automatic.

Stage 1: First provincial portfolio

Test the operating mechanism, measures, governance and core hypotheses in Newfoundland and Labrador. Publish the independent portfolio evaluation and reusable tools.

Stage 2: Cross-jurisdiction validation

If the first stage is credible, repeat selected configurations and the portfolio method in at least two materially different settings. Prioritize variation in sector, labour market, organizational scale and public-service context.

Stage 3: Deployment and sector diffusion

Move replicated configurations into existing federal, provincial, sector and commercialization programs. Support procurement, training, constraint removal and, where evidence warrants, productization of validated interaction patterns through Canadian firms and research institutions. Require continued outcome reporting for material public investments.

Stage 4: Strategic scale

Expand only where evidence shows that the national learning system creates value beyond what organizations and existing programs would produce on their own. Focus public funding on shared evidence, spillovers, underserved organizations, high public value and barriers that private actors cannot efficiently solve.

12. Risks, limits and falsification

Technology outruns the evidence

Models and products may change before a study concludes. The program should record the capability tested, use shorter decision cycles where possible and distinguish tool-specific from configuration-level findings.

Selection produces only enthusiastic participants

Early adopters may not represent typical workplaces. Portfolios should include variation in digital maturity and explicitly test transfer to ordinary operating conditions.

Measurement changes behaviour or creates burden

Evidence collection can overwhelm participants. Use the least burdensome design capable of informing the next decision and automate collection only where it remains valid and transparent.

Challenge becomes theatre

Requiring AI or people to challenge every decision can waste attention and create false doubt. Measure whether challenges identify material issues, improve decisions and justify their cost.

Human participation becomes ritual

Keeping a person nominally in the loop does not ensure understanding or control. Test whether the person can explain, question, override and accept responsibility for the result.

Gains are captured narrowly

Organizations may realize benefits while workers absorb monitoring, workload, insecurity or deskilling. Distribution and job-quality measures must be decision variables, not after-the-fact commentary.

Public funding crowds out private investment

Support should decline as private value and implementation certainty rise. Public funds should focus on uncertainty, evidence, spillovers, public-value uses and transferable learning.

Positive cases do not replicate

The policy should narrow or stop if apparent frontier gains consistently fail replication, the measurement burden exceeds decision value, or existing programs can produce equivalent learning more efficiently.

No national additionality emerges

The national mechanism is not justified merely because individual organizations benefit. It must show additional value through shared standards, credible evidence, cross-jurisdiction replication, diffusion, reduced duplication or access for organizations that would otherwise be excluded.

13. Recommended decisions and immediate actions

Within the next policy-development cycle, governments should:

1. adopt the Amplified Economy national objective and Human–AI Configuration Principle;
2. establish a small federal–provincial design table with worker, employer, Indigenous, research and responsible-AI expertise;
3. finalize the common experiment charter, measurement protocol, risk tiers and knowledge-disposition rules;
4. authorize confidential host capability assessment in Newfoundland and Labrador;
5. obtain host-specific costing, employment validation and independent-evaluation procurement estimates;
6. negotiate a staged funding and accountability agreement for the Full Learning Pilot within the $4.00 million ceiling;
7. recruit the Portfolio Lead before broad participant intake;
8. prepare the national clearinghouse and privacy-preserving publication standard before Wave 1 begins;
9. maintain and update the companion research foundation, Human–AI Amplification — Theory, Evidence and Testable Design Hypotheses, mapping each concept to evidence, pilot measures and falsification criteria;
10. define the prototype-development rules, interoperability expectations and intellectual-property treatment for tools emerging from experiments; and
11. set the evidence threshold and decision date for cross-jurisdiction expansion.

Conclusion

Canada does not need another policy that treats access to AI as the outcome. Nor does it need to choose, in advance, between optimism and caution about technology whose effects vary by task and organization.

It needs a disciplined way to learn.

The Amplified Economy would make that learning a national productive asset. It would search for unusually valuable human–AI configurations, test whether they survive quality adjustment and replication, reveal the complementary changes that allow gains to reach the organization, and measure whether workers and the public share in the result.

The Newfoundland and Labrador Full Learning Pilot is the proposed first move: large enough to generate portfolio learning, bounded enough to stop, and structured to produce evidence that can inform national action. The strategic bet is not that every experiment will work. It is that Canada can become better at discovering what works, rejecting what does not, and converting credible findings into broader capability.

Selected evidence base

Document control and disclosures

Package 0.5 Discussion Draft. Last reviewed 21 September 2026. Next review: before external circulation and after any material host, funding, governance, scope, budget or capacity decision.

Authorship: Terry Hudson is the originator and principal author. This document was developed in a personal capacity. The author is employed by the Government of Newfoundland and Labrador; that employment is disclosed for transparency. The document was not commissioned, approved or endorsed by the Government of Newfoundland and Labrador, and the views expressed are the author’s own. No institutional partner or co-signatory is represented by this draft. Contact: terry@hed1.ca.

AI-use disclosure: This document was developed with substantial AI assistance under the author’s direction. The project’s originating thesis, problem definition and core hypotheses were developed by the author from professional experience, observation and reasoning. AI supported research and evidence discovery, structured challenge, elaboration and organization of ideas, comparison of alternatives, synthesis, drafting and revision. The author directed the inquiry, supplied the institutional context, judged which outputs to retain or reject, and made the substantive policy and design decisions. The author remains responsible for source verification, accuracy, judgment and final content. The development process informed hypotheses about human–AI work, but it is not evidence that those hypotheses are true.

Conflict of interest: [AUTHOR DECISION: insert the author’s and any affiliate’s relevant financial, commercial or institutional interests before external circulation.] No tool, vendor or proprietary product is endorsed, preselected or given preferential access.