How we calculate AI disruption scores
The data sources, the formula, the modifiers, and the limitations — everything the numbers can and cannot tell you.
Independent project — not a government service
This is a free tool built by Scott Hazlitt as a personal project. It is not affiliated with the Government of Manitoba, any provincial department, or any employer or industry association. The site does not ask for names or contact details. Risk scores are modelled estimates derived from peer-reviewed academic datasets — not predictions, and not professional employment, legal, or business advice. Calculator answers are processed in the browser, but shared result links can include assessment choices in the URL. See Terms of Use for full disclaimers.
What this tool does
The Manitoba AI Disruption Explorer shows how likely each Manitoba industry and occupation is to be affected by AI — based on published academic research, not opinions. It combines four datasets into a single score (0–100) for each sector and job type, then adjusts for your business size and current AI use.
The scores are pre-calculated and stored as static data. No AI is involved in generating them at the time you visit the site — the numbers come entirely from the published datasets listed below, processed offline.
Manitoba was chosen because its economy is a useful cross-section: large insurance and financial services in Winnipeg, significant aerospace manufacturing, a large public sector, and agriculture — each with a very different AI exposure profile.
Source review status
The June 2026 review separates current public claims from older model inputs. Scores remain static model estimates; source review does not convert them into forecasts or live labour-market metrics.
| Area | Status | Notes |
|---|---|---|
| Core scoring datasets | Reviewed June 4, 2026 | Foundational automation, AI exposure, language-AI, Census, and crosswalk sources remain reachable. Older vintages are intentional model inputs, not live metrics. |
| Canadian adoption and labour context | Reviewed June 4, 2026 | The national AI-use benchmark was refreshed to 19.2% for Q2 2026. Manitoba sector adoption still uses the Q2 2024-Q2 2025 CSBC table and should be read as source-vintage directional indicators, not live metrics. |
| Supplementary research sources | Reviewed June 4, 2026 | Anthropic, Remote Labor, MIT Iceberg, Future Skills Centre, Conference Board, Dais/FSC, Policy Exchange, and OpenAI references are used for context or labelled modifiers where noted; not all are included in the composite formula. |
| Link health | Checked June 4, 2026 | The latest local pass found 19 direct 200 responses across major research, legal, and data links. The OpenAI policy page required browser verification after a command-line fetch failure. |
| Remaining review gap | Not complete | This pass improves public provenance, but it is not a full claim-by-claim freshness audit of every statistic, projection, and methodology statement on the site. |
Claim review ledger
Status
claim-reviewed local release candidate, pending final owner sign-off
Reviewed
2026-06-04
Public claims
10 reviewed public claims
Public claims, methodology claims, source-vintage caveats, privacy disclosures, and mechanism examples used by the local production build.
Release owner: Scott Hazlitt. This ledger makes the review state visible, but final owner sign-off is still pending and the live deployment must be rechecked before any public 9.5/10 claim.
Owner sign-off protocol
Final owner sign-off stays pending until the release owner has checked every claim against the rendered route, accepted the source freshness or source-vintage caveat, and retained approval evidence for the release candidate.
Protocol
- 1. Review each claim row against the rendered public route where the claim appears.Evidence: For every CLM identifier, record the checked route, reviewer, date, and whether the rendered page still matches the claim and caveat.
- 2. Confirm that source references are still acceptable for a public release claim.Evidence: Record source freshness notes, broken-link exceptions, source-vintage caveats, and any sources that need replacement before approval.
- 3. Confirm that methodology and model-input caveats are clear enough for non-specialist readers.Evidence: Record any copy changes or explicit acceptance that the caveat is sufficient for public use.
- 4. Check the live release candidate after deployment before changing this ledger to approved.Evidence: Attach a passing 13-route release gate, the 9.5 preflight output, and route screenshots or reviewer notes for trust pages.
- 5. Record final owner approval only after every claim has a signed-off owner status.Evidence: Update every claim from pending owner sign-off to owner signed off, with an approval date and reviewer note retained outside this ledger.
Evidence required before approval
- Passing local and live release-readiness gates for all 13 public routes.
- Claim-by-claim route copy check for every CLM identifier.
- Source freshness notes or source-vintage acceptance for every cited evidence cluster.
- Explicit acceptance of methodology choices, caveats, and non-advice language.
- Completed screen-reader review and production field-performance evidence before any 9.5/10 public claim.
- Final owner approval note naming reviewer, date, and release candidate.
Owner sign-off execution matrix
Each public claim needs route-bound review evidence before the ledger can move from pending owner sign-off to approved. The matrix names the rendered copy to check, the source-freshness review, caveat acceptance, and evidence to record.
| Claim | Rendered review target | Source freshness review | Evidence to record | Failure severity guidance |
|---|---|---|---|---|
| CLM-001site-wide | Check the site-wide score disclaimer in the footer, calculator result summary, About limitations, and Terms non-advice language. | Confirm the non-predictive scoring scope still matches the current methodology and public disclaimers.Caveat acceptance: Confirm readers are told the scores are directional relative exposure signals, not job-loss, business-failure, legal, financial, employment, procurement, or investment advice. | Claim ID and route; Rendered route or site-wide copy checked; Reviewer and review date; Release candidate identifier; Source freshness or methodology acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord the rendered footer, About, calculator result, and Terms copy checked. | Blocker if the public copy implies prediction or professional advice; serious if the caveat is present but easy to miss; minor if wording needs cleanup; pass only with recorded approval evidence.Acceptance: Rendered public copy still limits the score to relative AI disruption exposure. |
| CLM-002/about | Check the About formula, source manifest, scoring weights, and component-weight explanation against the implemented model. | Confirm the formula still matches the current scoring-weights.json methodology and source list.Caveat acceptance: Confirm readers are told the weighting scheme is a methodology choice that should be sensitivity-tested before policy or procurement use. | Claim ID and route; Rendered About formula checked; Reviewer and review date; Release candidate identifier; Source freshness or methodology acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord the About formula text and scoring-weights source checked. | Blocker if rendered weights differ from implementation; serious if methodology caveats are missing; minor if evidence wording needs cleanup; pass only with recorded approval evidence.Acceptance: About formula text matches the implemented scoring weights. |
| CLM-003/calculator | Check calculator, methodology, and source-manifest copy describing the CSBC Q2 2024-Q2 2025 Manitoba sector proxy and Q2 2026 national benchmark. | Confirm the CSBC source vintage and national benchmark are still labelled correctly before approval.Caveat acceptance: Confirm the Manitoba sector values are not presented as a live adoption feed. | Claim ID and route; Rendered calculator/source-vintage copy checked; Reviewer and review date; Release candidate identifier; Source freshness or source-vintage acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord calculator and source-manifest text checked for CSBC vintage wording. | Blocker if source-vintage values are presented as live; serious if benchmark context is unclear; minor if wording needs cleanup; pass only with recorded approval evidence.Acceptance: Calculator and methodology copy label Manitoba sector values as source-vintage model inputs. |
| CLM-004/about | Check About data release sign-off, source manifest, and public data file inventory. | Confirm the manifest still reports 5 model data files, 114 model records, and 3 public review artifacts.Caveat acceptance: Confirm the inventory is labelled as local package evidence and not live deployment parity. | Claim ID and route; Rendered About data inventory checked; Reviewer and review date; Release candidate identifier; Source-manifest count acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord the About inventory and source-manifest counts checked. | Blocker if public counts are wrong; serious if live parity is implied; minor if table wording needs cleanup; pass only with recorded approval evidence.Acceptance: About inventory matches source-manifest counts and review-artifact list. |
| CLM-005/policy | Check policy source-review and bibliography sections for Canadian exposure and adoption statistics. | Confirm source vintages, including the 19.2% national Q2 2026 AI-use signal, remain labelled as Canadian context.Caveat acceptance: Confirm Canadian context sources are not presented as Manitoba-specific forecasts. | Claim ID and route; Rendered policy source-review copy checked; Reviewer and review date; Release candidate identifier; Source freshness or source-vintage acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord policy source-review rows and bibliography links checked. | Blocker if Canadian context is presented as Manitoba forecast; serious if source vintage is hidden; minor if citation wording needs cleanup; pass only with recorded approval evidence.Acceptance: Policy page labels Canadian context and source vintages clearly. |
| CLM-006/policy | Check policy retraining-capacity language and source-review caveats. | Confirm the Policy Exchange reference remains labelled as a UK recommendation.Caveat acceptance: Confirm no Manitoba retraining target is created without a local model. | Claim ID and route; Rendered policy retraining copy checked; Reviewer and review date; Release candidate identifier; Source freshness or context acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord the retraining-capacity policy copy checked. | Blocker if the UK recommendation becomes a Manitoba target; serious if caveat is too far from the claim; minor if wording needs cleanup; pass only with recorded approval evidence.Acceptance: UK retraining recommendation is not converted into a Manitoba target. |
| CLM-007/threat-model | Check threat-model cost-floor language, 280x claim, and evidence-review table. | Confirm the 280x cost signal remains limited to GPT-3.5-level inference-price comparison.Caveat acceptance: Confirm total AI operating costs are not overstated by excluding integration, support, compliance, labour transition, quality assurance, and ROI uncertainty. | Claim ID and route; Rendered threat-model cost claim checked; Reviewer and review date; Release candidate identifier; Source freshness or mechanism acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord the threat-model cost-floor copy and evidence row checked. | Blocker if 280x is framed as all-in operating cost; serious if cost caveats are missing; minor if wording needs cleanup; pass only with recorded approval evidence.Acceptance: 280x claim is constrained to GPT-3.5-level inference pricing. |
| CLM-008/threat-model | Check threat-model company and vendor examples plus Manitoba transferability caveats. | Confirm examples remain company-reported or third-party-reported mechanism examples.Caveat acceptance: Confirm examples are not presented as direct Manitoba benchmarks or forecasts. | Claim ID and route; Rendered threat-model example copy checked; Reviewer and review date; Release candidate identifier; Source freshness or mechanism acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord the threat-model example rows checked. | Blocker if examples are framed as Manitoba forecasts; serious if mechanism caveats are missing; minor if labels need cleanup; pass only with recorded approval evidence.Acceptance: Company and vendor examples remain labelled as mechanism examples. |
| CLM-009/privacy | Check Privacy, Terms, and calculator share-result URL warnings. | Confirm URL-based assessment-choice disclosure still matches calculator behaviour.Caveat acceptance: Confirm users are warned not to enter confidential business information and that shared URLs can expose choices through history, recipients, and hosting logs. | Claim ID and route; Rendered privacy/share URL copy checked; Reviewer and review date; Release candidate identifier; Privacy behaviour acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord Privacy, Terms, and calculator share-result text checked. | Blocker if URL-based choices are undisclosed; serious if confidentiality warning is missing; minor if wording needs cleanup; pass only with recorded approval evidence.Acceptance: Privacy, Terms, and calculator sharing copy consistently describe URL-based assessment choices. |
| CLM-010/privacy | Check Privacy Speed Insights disclosure and field-performance review artifact. | Confirm Speed Insights instrumentation status and field-performance pending state still match the deployed release candidate.Caveat acceptance: Confirm production Core Web Vitals are not described as measured until real traffic evidence exists. | Claim ID and route; Rendered privacy performance disclosure checked; Reviewer and review date; Release candidate identifier; Instrumentation and field-evidence acceptance; Caveat acceptance; Durable evidence artifact location plus reviewer summaryRecord Privacy Speed Insights disclosure and field-performance artifact checked. | Blocker if pending performance evidence is described as complete; serious if disclosure is missing; minor if wording needs cleanup; pass only with recorded approval evidence.Acceptance: Privacy page describes anonymous performance metrics without overclaiming measured field performance. |
Completed owner/source sign-off evidence
A completed final sign-off must record a claim ID, owner, reviewer, approval date, release candidate, source-freshness acceptance, caveat acceptance, and an durable evidence location for every public claim.
No completed owner/source sign-off evidence recorded yet.
| Claim | Evidence status | Review state | Owner approval | Caveat |
|---|---|---|---|---|
| CLM-001 / site-wideRisk scores measure relative AI disruption exposure, not certainty of job loss, business failure, or policy outcome. | methodology choice | reviewed with caveat | pending owner sign-offConfirm site-wide disclaimer and score explanation still make the non-predictive scope clear. | The score is a directional composite model and should not be used as legal, financial, employment, procurement, or investment advice. |
| CLM-002 / /aboutThe composite score blends automation probability, AI occupational exposure, language-model exposure, and sector adoption gap inputs. | methodology choice | reviewed with caveat | pending owner sign-offConfirm the About formula text still matches the implemented scoring weights. | The weighting scheme is a methodological decision and should be sensitivity-tested before policy or procurement use. |
| CLM-003 / /calculatorManitoba sector AI adoption inputs use a CSBC Q2 2024-Q2 2025 source-vintage proxy and are compared with the 19.2% national Q2 2026 benchmark. | source-vintage model input | reviewed with caveat | pending owner sign-offConfirm calculator, methodology, and source-manifest copy all label the CSBC table as source-vintage. | The Manitoba sector values are not a live adoption feed and should be read as directional model inputs until the sector table is refreshed. |
| CLM-004 / /aboutThe local public release separates 5 model data files covering 114 records from 3 public review artifacts. | current public source | reviewed | pending owner sign-offConfirm the source manifest record counts and review-artifact inventory match the current public data files. | The record count proves the local model data package shape and review-artifact inventory, not that the same package has been deployed to the live domain. |
| CLM-005 / /policyCanadian exposure and adoption statistics on the policy page are tied to their cited source vintages, including the 19.2% national Q2 2026 AI-use signal. | context-only reference | reviewed with caveat | pending owner sign-offConfirm the policy page still labels national and Canadian context sources with source vintages. | These sources provide Canadian context and should not be treated as Manitoba-specific forecasts unless the page explicitly says so. |
| CLM-006 / /policyRetraining capacity references are labelled as a UK recommendation, not converted into a Manitoba target. | context-only reference | reviewed with caveat | pending owner sign-offConfirm the policy page does not convert the UK retraining recommendation into a Manitoba target. | Any Manitoba retraining-capacity target would need a local model using Manitoba workforce counts, training capacity, funding, and employer demand. |
| CLM-007 / /threat-modelThe 280x cost signal is a GPT-3.5-level inference-price comparison, not a blanket statement about all AI operating costs. | caveated mechanism example | reviewed with caveat | pending owner sign-offConfirm threat-model copy limits the 280x claim to GPT-3.5-level inference pricing. | Inference-price declines do not include integration, support, compliance, labour transition, quality assurance, or uncertain ROI costs. |
| CLM-008 / /threat-modelAI-native company and vendor examples are mechanism examples, not direct Manitoba benchmarks or forecasts. | caveated mechanism example | reviewed with caveat | pending owner sign-offConfirm company and vendor examples remain labelled as mechanism examples. | Reported examples show plausible competitive mechanisms but require independent operating data before they should be treated as local sector benchmarks. |
| CLM-009 / /privacyAssessment answers are processed in the browser unless a user shares or revisits a result URL, which can include assessment choices in the URL. | privacy/legal disclosure | reviewed with caveat | pending owner sign-offConfirm calculator sharing, Privacy, and Terms copy consistently describe URL-based assessment choices. | URL-based sharing may expose choices to browser history, recipients, and standard hosting logs; users should not enter confidential business information. |
| CLM-010 / /privacyVercel Speed Insights is disclosed as anonymous performance-metric collection for Core Web Vitals and related route/device metrics. | privacy/legal disclosure | reviewed with caveat | pending owner sign-offConfirm Privacy copy and field-performance artifact describe Speed Insights without overclaiming measured performance. | The local build prepares instrumentation and an evidence checklist, but actual Core Web Vitals/RUM evidence requires deployment and real traffic after release. |
Data release sign-off
Status
source-reviewed local release candidate
Reviewed
2026-06-04
Model data files
5 model data files, 114 records
Review artifacts
3 review artifacts
Public model JSON data files, public methodology ledgers, public review artifacts, and source-vintage caveats used by the local production build.
Release owner: Scott Hazlitt. This sign-off is for the local release candidate and does not prove that the current public deployment has the same content.
| File | Records | Review status | Main limitation |
|---|---|---|---|
| industries.json | 20 | source-vintage directional model input | Sector employment, GDP share, and AI adoption values are static model inputs. |
| occupations.json | 68 | source-vintage composite model input | Occupation scores are relative exposure estimates, not forecasts of job loss. |
| scoring-weights.json | 1 | methodology reference | Weights are methodological choices and should be sensitivity-tested before policy use. |
| sector-playbooks.json | 20 | directional planning guidance | Playbook actions are strategic prompts, not sector-specific legal, HR, or procurement advice. |
| threat-scenarios.json | 5 | mechanism examples with source caveats | Threat scenarios are mechanism examples and are not Manitoba forecasts. |
| Artifact | Purpose | Review status | Main limitation |
|---|---|---|---|
| claim-ledger.json | Public accountability ledger for claims, evidence-status categories, review state, and caveats that need owner sign-off before a 9.5 release. | claim-reviewed local release candidate, pending final owner sign-off | This artifact documents local release-candidate review state and does not prove live deployment parity. |
| accessibility-review.json | Public accessibility status ledger for completed local automated and keyboard checks, pending assistive-technology checks, and certification caveats. | local accessibility review candidate, screen-reader review pending | This artifact is not a formal WCAG certification. |
| field-performance-review.json | Public field-performance evidence ledger for production Core Web Vitals requirements, Speed Insights instrumentation status, and the evidence still needed before a 9.5 claim. | field performance evidence pending production traffic | This artifact documents evidence requirements and does not prove deployed production traffic has met them. |
Data Sources
| Source | Description | Vintage |
|---|---|---|
| Frey & Osborne (2013) | 702 US occupations rated for automation probability using a Gaussian process classifier. The foundational dataset for occupation-level automation risk. | 2013 |
| Felten, Raj & Seamans — AI Occupation Exposure Index | AI Occupational Exposure index linking AI patent applications to O*NET occupational task descriptions. Captures AI-specific exposure, distinct from general automation. | 2021 |
| Eloundou et al. (2023) — GPTs are GPTs | Occupation-level language AI exposure dataset published by OpenAI researchers. Human-reviewed scores (0–1) covering 923 occupations, mapped to Canadian job codes. Measures direct language AI substitution plus half-weight tool-augmentation exposure. Used as the Language AI Impact component. | 2023 |
| Brookfield Institute NOC Crosswalk | Maps US SOC occupation codes to Canadian NOC 2021 codes. Required because Frey & Osborne and the AI exposure index use SOC; Canadian employment data uses NOC. | 2019–2021 |
| Statistics Canada — Census 2021 | Manitoba employment counts by NOC occupation. The primary source for how many Manitobans work in each occupation. | 2021 |
| Statistics Canada — Canadian Survey on Business Conditions (CSBC) | Provincial AI adoption rates by industry sector. Provides Manitoba-specific figures used for the Sector Adoption Gap component. The national benchmark was 12.2% in Q2 2025 and has since moved to 19.2% in Q2 2026; treat the Manitoba comparison as source-vintage, not a live metric. | Q2 2024–Q2 2025 |
| Manitoba Bureau of Statistics — GDP by Industry | Manitoba GDP shares by NAICS sector. Provides the economic weight context alongside employment counts. | 2022–23 |
| Anthropic Economic Index (2026) | Tracks actual Claude usage patterns across occupations, sampled from real interactions (Feb 5-12, 2026). Unlike the other sources which measure theoretical AI capability, this index measures what AI is doing in practice — based on 94GB of Claude conversation data categorized by the Clio privacy-preserving analysis tool. Displayed as supplementary context in occupation detail panels; not included in the composite score formula due to SOC major-group granularity (22 groups vs. 923 individual occupations in the other datasets). | Q1 2026 (quarterly updates) |
| Remote Labor Index — remotelabor.ai (2025) | Benchmarks actual autonomous AI agent performance on 240 real paid freelance projects across 23 Upwork skill categories, tested with 358 freelancers. Measures what advanced AI systems actually complete end-to-end today (0.83%–4.17%), providing real-world calibration for the theoretical exposure scores in this tool. Used as a contextual callout in calculator results to ground the gap between capability and deployment. Not incorporated into the composite score formula. | Feb–Mar 2025 |
| OpenAI — Industrial Policy for the Intelligence Age (2026) | Policy paper projecting that AI infrastructure buildout (data centres, power grids, cooling systems) will require approximately 20% more skilled trades workers — electricians, mechanics, ironworkers, carpenters, plumbers — than currently exist. Used as a counter-exposure signal on trades occupations in the explorer: low AI displacement risk combined with rising infrastructure demand. Not a research dataset; cited as a policy reference only. | April 2026 |
| MIT Project Iceberg (2024) | Maps 13,000+ deployed AI tools against 32,000+ skills across 923 O*NET occupations (151 million workers) to produce an Iceberg Index: the percentage of an occupation's wage value where AI has demonstrated capability. Key finding: visible tech-sector disruption is just 2.2% of the U.S. labour market's wage value, while hidden white-collar and administrative exposure is 11.7% — five times larger and geographically distributed across all states. Referenced as a methodological context source; not incorporated into composite scores because occupation-level Index values are not published as a downloadable dataset. | 2024 |
| Future Skills Centre — Canada's Workforce in Transition | Classifies 57.4% of Canadian jobs as highly AI-exposed, split between AI-competing roles (where AI automates core tasks) and AI-augmenting roles (where AI enhances human capabilities). Analyzes 19 million job postings to track shifting demand. AI-augmenting roles grew 2.9% in 2024, outpacing AI-competing roles at 1.6%. | Sept 2025 |
| Conference Board of Canada — Understanding the Influence of AI on Employment | Canadian task-level AI exposure index covering 501 NOC occupations and 304 NAICS industries. Uses a 3-phase framework: exposure, productivity gains, and automation likelihood. Projects a short-term employment dip of 535,000 jobs by 2030, followed by a long-term gain of 555,000 jobs by 2045 as productivity benefits materialize. | Jan 2026 |
| The Dais / FSC — Right Brain, Left Brain, AI Brain | Exposure-complementarity framework classifying 506 Canadian NOC occupations into four quadrants based on AI exposure and whether AI assists or replaces workers. 56% of Canadian workers are in occupations with higher AI exposure. Used as the basis for the AI-augmenting and AI-competing labels shown in occupation detail panels. | Jan 2025 |
| Policy Exchange — Government in the Age of Superintelligence | UK policy think-tank report examining workforce transformation, skills revaluation, and government preparedness for AI disruption. Projects large-scale labour market dislocation across white-collar sectors, recommends national retraining capacity of 250,000 workers annually, and argues that roles dismissed as 'low-skilled' are actually 'low-paid' and will command increasing premiums as cognitive work is automated. | 2025 |
Score Formula
Overall Risk Score =
(Automation Probability × 0.30)
+ (AI Occupation Exposure × 0.30)
+ (Language AI Impact × 0.25)
+ (Sector Adoption Gap × 0.15)
Adjusted Score =
Overall Risk Score × Business Size Factor × AI Adoption Factor
The Frey & Osborne (2013) automation probability for the closest matching occupation, converted to a 0–100 scale. Reflects the probability that an occupation’s tasks could be automated by computerisation over roughly a 10–20 year horizon.
The Felten-Raj-Seamans AI Occupational Exposure index, normalized to 0–100. Measures how much of an occupation’s task content corresponds to capabilities demonstrated in recent AI patent applications. More AI-specific than the automation probability measure.
Human-reviewed scores from the Eloundou et al. (2023) published dataset, normalized to 0–100. Measures the fraction of an occupation’s tasks that could be handled directly by language AI, plus half-weight credit for tasks where AI tools provide assistance. Covers 923 occupations; mapped to Canadian job codes via the Brookfield Institute crosswalk.
The gap between maximum possible AI adoption (100%) and the sector’s source-vintage adoption rate. A high gap means the sector has not yet adopted AI broadly, indicating that AI-driven disruption is ahead of, not behind, the current workforce.
Business size modifiers
AI adoption modifiers
Risk tier thresholds
Limitations
- Frey & Osborne predates LLMs. The 2013 paper was written before GPT, diffusion models, and modern generative AI existed. It likely underestimates disruption risk for knowledge-work occupations. We partially compensate with the Language AI Impact component, but the underlying dataset remains a product of its era.
- National AI adoption rates used as Manitoba proxy. No province-specific AI adoption survey exists. Manitoba may lag national averages in some sectors (particularly due to firm size distribution and distance from technology hubs) or lead in others (aerospace).
- Composite scores are relative, not predictive. A score of 72 does not mean 72% of jobs in that sector will be lost. It means that sector scores in the 72nd percentile of AI disruption exposure relative to others. Actual employment outcomes depend on labour market conditions, regulation, adoption rates, and adaptation.
- O*NET-to-NOC crosswalk covers 45 of 49 occupations directly. Four occupations (College Instructors, Biological Scientists, Equipment Assemblers, Shelf Stockers) use averaged scores from the nearest job category groupings rather than a single exact match. Their confidence rating is still marked “published” as the underlying scores are real data, but the crosswalk introduces slightly more noise for these occupations.
- SOC-to-NOC crosswalk introduces noise. Frey & Osborne and the AI exposure index use US Standard Occupational Classification codes. The Brookfield Institute crosswalk maps these to Canadian NOC codes, but the mapping is not always one-to-one. Some NOC occupations combine multiple SOC categories; scores for these are averaged.
- Small-sector estimates carry higher uncertainty. Sectors with fewer than 10,000 Manitoba employees (e.g., Mining & Oil, Corporate Management) have fewer reference occupations, making sector-level score aggregation noisier.
- Anthropic Economic Index data is at the SOC major-group level — all occupations within a group (e.g., all “Computer and Mathematical” occupations) receive the same usage intensity value. Shown as supplementary context only; not incorporated into composite scores.
- Remote Labor Index reflects a point-in-time snapshot. The 0.83%–4.17% autonomous completion range is from a Feb–Mar 2025 benchmark across 23 Upwork categories. AI agent capability is improving rapidly; this figure should be treated as a lower-bound calibration anchor, not a permanent ceiling.
- MIT Iceberg occupation scores are not yet incorporated. The Iceberg Index covers 923 O*NET occupations using 32,000+ skills mapped against 13,000+ deployed AI tools. Its key insight — that visible tech disruption (2.2% of wage value) is dwarfed by hidden white-collar exposure (11.7%) — is referenced in the methodology but individual occupation Index values are not published as a downloadable dataset, so they cannot be added to the composite score formula.
- AI-augmenting/AI-competing classification is a binary simplification. Derived from the FSC complementarity framework (building on IMF methodology by Pizzinelli et al.) applied to Canadian NOC codes. Some occupations near the threshold could reasonably be classified either way. The classification reflects the current generation of AI tools and may shift as capabilities evolve.
Methodology Decisions
Each component captures a different dimension of AI-related risk. The automation probability measures task routineness. The AI occupation exposure index measures demonstrated AI capability overlap. The language AI impact score captures the generative AI wave specifically. The adoption gap captures timing — a sector with high exposure but low source-vintage adoption is at acute near-term risk, not just theoretical risk.
Automation probability and AI occupation exposure are given equal weight as the two most established academic measures. Language AI impact is weighted slightly lower because some scores are estimated rather than directly published. The adoption gap is weighted lowest because it is a sector-level proxy rather than an occupation-specific measure.
Micro-businesses (<5 employees) tend to have less access to AI tools and less organizational capacity to manage AI transitions, increasing their effective risk. Large businesses have more resources to adapt and to absorb workforce changes, reducing their effective risk. These modifiers adjust the composite score to reflect business context without changing the underlying occupational data.
Businesses already using AI tools have de facto begun their transition. Their effective exposure to disruption from AI adoption is lower because they are driving the change rather than being surprised by it. The 30% reduction reflects this first-mover advantage, not immunity.
All four composite score components measure theoretical AI capability exposure — what AI could automate, based on task and skill overlap. Real-world deployment lags significantly behind theoretical capability. The Remote Labor Index (2025) found that advanced AI systems complete just 0.83%–4.17% of complex professional projects end-to-end without human intervention. The composite scores in this tool reflect the destination of the disruption curve, not its current position. The cost convergence charts in the calculator results are explicitly modelled as a 24-month trajectory, not an instantaneous shift.
MIT Project Iceberg (2024) found that visible tech-sector AI disruption represents just 2.2% of total U.S. labour market wage value — while hidden white-collar and administrative exposure is 11.7%, five times larger, spread across manufacturing, financial services, logistics, and healthcare administration in every state. This tool's occupation and industry scores capture both the visible layer (software, engineering, creative roles) and the larger hidden layer (administrative, coordination, office support). The Iceberg research was independently validated against the Anthropic Economic Index with 69% geographic agreement and 85% accuracy in predicting occupational transitions.
The Anthropic Economic Index (March 2026) measures actual Claude usage patterns across 22 SOC major occupational groups — a 2026-vintage real-world signal. Adding it to the composite formula at 15–25% weight would flatten within-group differentiation: Software Developers and Network Technicians would both become “Computer and Mathematical = same score,” which is methodologically dishonest for a precision tool. It is shown as supplementary context in occupation detail panels instead, with a note about the group-level granularity.
Flag a Data Issue
If you notice an occupation score that appears incorrect, a Manitoba employer missing from a sector, or a data source that has been updated since our vintage year, please open an issue on the project repository or contact the maintainer directly. We prioritize corrections that affect high-employment occupations or flagship Manitoba sectors.