Fructus

11 source-checked claims. Estimates and projections are labeled.

No matching outcomes.

Awaiting review

19 claims whose source or financial interpretation still needs checking.

1. 22 October 2015 Source published

   NCAR's DICast wind forecasts cut Xcel Energy's forecast error to 10.8%; Xcel estimates US$49 million saved

   NCAR's DICast cuts Xcel's wind forecast error to 10.8%

   US $49 million

   Cost savings

   Awaiting source reviewUSD

   1 claim source

   System

   NCAR wind power forecasting system with the DICast adaptive forecast blending engine

   Developer

   National Center for Atmospheric Research (Research Applications Laboratory)United States

   Deployer

   Xcel EnergyUnited States

   Claim geography

   United States

   Case evidence: independent reporting3 case publications

   Updated 21 October 2026

   Compare benefit

   NCAR's DICast wind forecasts cut Xcel Energy's forecast error to 10.8%; Xcel estimates US$49 million saved

   Source and detail for NCAR's DICast wind forecasts cut Xcel Energy's forecast error to 10.8%; Xcel estimates US$49 million saved

   saved through forecasting improvements, cumulative to 2015

   Reported amount

   US $49 million

   Currency

   The savings concern Xcel Energy customers in the United States and are reported in U.S. dollars.

   Beneficiary

   Xcel Energy customers

   Paid by

   Not established in the current source review.

   Unit

   cumulative savings across Xcel customers

   Period

   Cumulative to 2015; the account places forecasting work from 2009, but the amount-specific start is unstated.

   Amount basis

   Unresolved

   Accounting

   Appears to be gross avoided operating cost; implementation and running costs are unstated.

   Comparison

   Operational forecast-error baseline; no comparable monetary baseline is stated.

   AI's part

   The estimate concerns forecasting improvements as a whole; the incremental AI share is not established.

   Amount-specific start date

   Calculation method and baseline operating cost

   Implementation and ongoing costs

   AI component and its incremental contribution

   cumulative_not_annual

   whole_program_attribution

   ai_component_needs_review

   operator_estimate

   NCAR's DICast wind forecasts cut Xcel Energy's forecast error to 10.8%; Xcel estimates US$49 million saved

   Xcel estimates that improvements in forecasting have saved customers US $49 million so far.

   Do not add the separate 2010 saving or prospective annual 2017 figure to the cumulative amount.

   DIV-0001, AIBD-0168

2. Date unknown

   CMS's Fraud Prevention System identifies or prevents $654.8 million in 2015, an $11.5-to-$1 return on investment

   CMS's FPS identifies or prevents $654.8 million in 2015

   $654.8 million

   Avoided losses

   Mixed accounting statesUSD

   1 claim source

   System

   Fraud Prevention System (FPS)

   Developer

   Contractors engaged by CMS (the report lists 'Developing Contractor Costs' and 'Modeling Contractor Costs' without naming firms)

   Deployer

   Centers for Medicare and Medicaid Services (CMS)United States

   Claim geography

   United States

   Case evidence: party or catalog claim1 case publication

   Updated 21 October 2026

   Compare benefit

   CMS's Fraud Prevention System identifies or prevents $654.8 million in 2015, an $11.5-to-$1 return on investment

   Source and detail for CMS's Fraud Prevention System identifies or prevents $654.8 million in 2015, an $11.5-to-$1 return on investment

   inappropriate Medicare payments identified or prevented in calendar year 2015

   Reported amount

   $654.8 million

   Currency

   The amount concerns the U.S. Medicare program and is reported in U.S. dollars.

   Beneficiary

   Medicare program (CMS)

   Paid by

   Not established in the current source review.

   Unit

   program aggregate

   Period

   Calendar year 2015

   Amount basis

   Unresolved

   Accounting

   The benefit amount is distinct from the reported estimated $57.1 million cost; the headline amount is not a net benefit.

   Comparison

   The existing comparator is the prior program year, not a no-AI counterfactual.

   AI's part

   FPS-supported investigations and actions; aggregate includes several accounting categories.

   Separate realized recovery, prevented disbursement, projected future avoidance and referrals

   Full ROI cost boundary

   Contribution of AI versus investigation

   mixed_recovery_prevention_projection

   prior_year_not_counterfactual

   whole_program_attribution

   estimate_flag_too_coarse

   CMS's Fraud Prevention System identifies or prevents $654.8 million in 2015, an $11.5-to-$1 return on investment

   The FPS helped identify or prevent $654.8 million in inappropriate payments during calendar year 2015 through actions taken due to the FPS or through investigations expedited, augmented, or corroborated by the FPS (Table 1).

   Hold: the $654.8 million combines accounting states including prevention, projected savings, and amounts referred for recovery. Split components before curating a monetary benefit.

   DIV-0002, AIBD-0186

3. 15 September 2016 Source published

   Global Fishing Watch data lets Kiribati fine a vessel US$1 million for fishing in the Phoenix Islands Protected Area

   Global Fishing Watch helps Kiribati fine a vessel US$1 million

   ~AUD 1.43 million

   Transfers and awards

   Awaiting source reviewAUD

   Equivalent at October 2026 rates.

   1 claim source

   System

   Global Fishing Watch (AIS-based fishing detection using machine learning)

   Developer

   Oceana, SkyTruth and GoogleUnited States

   Deployer

   Government of KiribatiKiribati

   Claim geography

   Kiribati

   Case evidence: independent reporting4 case publications

   Updated 21 October 2026

   Compare benefit

   Global Fishing Watch data lets Kiribati fine a vessel US$1 million for fishing in the Phoenix Islands Protected Area

   Source and detail for Global Fishing Watch data lets Kiribati fine a vessel US$1 million for fishing in the Phoenix Islands Protected Area

   fine secured by Kiribati against the owner of Marshalls 203 for fishing illegally in the Phoenix Islands Protected Area, plus an additional $1 million paid as a grant under a goodwill arrangement

   Reported amount

   $1 million

   Currency

   Kiribati uses Australian dollars. The fine paid to its government is reported in U.S. dollars.

   Local equivalent

   ~AUD 1.43 million. Converted from USD at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   Government of Kiribati

   Paid by

   Not established in the current source review.

   Unit

   one enforcement settlement with two differently described payments

   Period

   Single case reported by 2016; exact payment date absent.

   Amount basis

   Unresolved

   Accounting

   Gross government receipts or commitments; enforcement costs and losses to the payer are not included.

   Comparison

   No monetary comparator.

   AI's part

   Tracking supported enforcement; classifier contribution and payment realization remain unclear.

   Proof and dates of payment

   Separate fine from grant

   Reconcile later reporting that describes a $2 million fine

   Counterfactual enforcement outcome

   transfer_not_total_social_value

   multiple_payment_components

   fine_grant_conflict

   ai_increment_unresolved

   Global Fishing Watch data lets Kiribati fine a vessel US$1 million for fishing in the Phoenix Islands Protected Area

   This helped the country secure a $1 million fine against the Central Pacific Fishing Company, the owner of Marshalls 203. The company also made a special “goodwill arrangement” with the government, agreeing to pay an additional $1 million in the form of a grant.

   The amount field is $1 million, while what/quote also describe a separate $1 million grant. Do not silently treat the sum as one fine.

   DIV-0003, AIBD-0196

4. Date unknown

   CGU, supported by ALICE, cuts R$2.38 million from a 2021 Hospital Federal de Ipanema glove tender and R$5.95 million from an undated INCA glove tender

   CGU, with ALICE support, avoids R$8.33m in overpriced purchases

   ≥ R$ 8,33 milhões

   Cost savings

   Awaiting source reviewBRL

   1 claim source

   System

   ALICE (Analisador de Licitações, Contratos e Editais), CGU automated audit-trail and AI alert tool

   Developer

   Controladoria-Geral da União (CGU), in-house (launched 2015 per the release)Brazil

   Deployer

   Controladoria-Geral da União (CGU)Brazil

   Claim geography

   Brazil

   Case evidence: party or catalog claim3 case publications

   Updated 21 October 2026

   Compare benefit

   CGU, supported by ALICE, cuts R$2.38 million from a 2021 Hospital Federal de Ipanema glove tender and R$5.95 million from an undated INCA glove tender

   Source and detail for CGU, supported by ALICE, cuts R$2.38 million from a 2021 Hospital Federal de Ipanema glove tender and R$5.95 million from an undated INCA glove tender

   overpriced hospital purchases avoided across five preventive audits

   Reported amount

   R$ 8,33 milhões

   Currency

   The procurement amount concerns federal hospitals in Rio de Janeiro, Brazil, and is reported in Brazilian reais.

   USD equivalent

   ≥ ~USD 1.6 million. Converted from BRL at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   the federal government (hospitals and institutes in Rio de Janeiro)

   Paid by

   Not established in the current source review.

   Unit

   five preventive audits of Rio federal hospitals and institutes

   Period

   Reported October 2023; at least one component is dated 2021; full interval is unclear.

   Amount basis

   Unresolved

   Accounting

   Appears to be gross cost avoidance; paid-price outcomes and audit costs are not established.

   Comparison

   Original and revised estimates appear for individual components, not a fully reconciled aggregate.

   AI's part

   Human audit and procurement revisions followed ALICE alerts.

   Reconcile all five audit components

   Full event interval

   Final contract/payment amounts

   Audit and system costs

   budget_revision_not_cash_release

   multi_case_aggregate

   component_baselines_only

   possible_overlap_with_program_aggregates

   CGU, supported by ALICE, cuts R$2.38 million from a 2021 Hospital Federal de Ipanema glove tender and R$5.95 million from an undated INCA glove tender

   A Controladoria-Geral da União (CGU) realizou, preventivamente, cinco auditorias em editais de contratações de hospitais federais e institutos sediados no Rio de Janeiro/RJ. A ação evitou que, pelo menos, R$ 8,33 milhões fossem desperdiçados em aquisições de insumos e serviços hospitalares com sobrepreço.

   Preserve the lower-bound language of at least R$8.33 million. Do not substitute the full original-versus-auction price gap for the cited benefit.

   DIV-0008, AIBD-0082

5. 4 April 2022 Source published

   Vyntra's (formerly NetGuardians) AI risk models may have blocked fraudulent payments at unnamed banks, including a USD 500,000 ATM scheme

   Vyntra's AI risk models may have stopped a USD 500,000 ATM fraud

   KSh 58 million

   Avoided losses

   Awaiting source reviewKES

   1 claim source

   Monetary claim assessed 23 October 2026

   System

   Vyntra (NetGuardians) AI risk models for real-time payment and e-banking fraud detection

   Developer

   Vyntra (NetGuardians SA, merged with Intix in June 2025)Switzerland

   Deployer

   Unnamed client banks

   Claim geography

   Kenya

   Case evidence: independent reporting2 case publications

   Updated 21 October 2026

   Compare benefit

   Fraud Cases - Vyntra

   Source and detail for Vyntra's (formerly NetGuardians) AI risk models may have blocked fraudulent payments at unnamed banks, including a USD 500,000 ATM scheme

   fraudulent ATM withdrawal detected and stopped, with arrests

   Reported amount

   USD 500,000

   Currency

   The bank-fraud example is identified as Kenyan. The source reports the same amount as KSh 58 million and US$500,000.

   USD equivalent

   ~USD 446,000. Converted from KES at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   the second bank, an unnamed client bank (the page labels the case 'Kenya')

   Paid by

   Not established in the current source review.

   Unit

   one attempted withdrawal at an unnamed bank

   Period

   No event date established

   Amount basis

   Unresolved

   Accounting

   Nominal attempted amount, before intervention costs; expected loss without intervention is unknown.

   Comparison

   No independent counterfactual or expected-loss estimate.

   AI's part

   Vendor attributes the intervention to its models; AI and human contributions are not separated.

   Bank identity and confirmation

   Event date

   Expected loss or later recovery without the system

   AI mechanism and intervention cost

   vendor_claim

   dual_currency_same_amount

   undated_case

   attempted_amount_not_expected_loss

   Fraud Cases - Vyntra

   The fraud was detected in real time and the fraudsters were arrested while attempting to withdraw USD 500,000 (KSh 58 million) from the second bank’s ATMs. Vyntra's AI risk models prevented the fraud due to the unusual timing and frequency of ATM withdrawals.

   Kenya - Internal and channels fraud

   Reported by the developerSource checked 2026-10-23

   Source-stated equivalent: KSh 58 million (KES)

   The source gives both currencies for the same amount. It does not specify an exchange-rate date or method.

   Reported amount

   KSh 58 million

   USD basis

   Converted from KES at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Unit

   one attempted withdrawal at an unnamed bank

   Period

   No event date established

   Status

   Awaiting source review

   Accounting

   Nominal attempted amount, before intervention costs; expected loss without intervention is unknown.

   Price basis

   Price year and inflation basis await review.

   Fraud Cases - Vyntra

   USD 500,000 (KSh 58 million)

   Kenya - Internal and channels fraud

   Source checked 2026-10-23

   USD 500,000 and KSh 58 million describe the same attempted withdrawal. The bank and event date remain unnamed, and the amount does not establish the loss that would otherwise have occurred.

   DIV-0009, AIBD-0317

6. 31 March 2023 Source published

   York student has a £60 parking fine withdrawn after sending an appeal letter drafted by ChatGPT

   ChatGPT appeal letter gets a York student's £60 fine withdrawn

   £60

   Cost savings

   Awaiting source reviewGBP

   1 claim source

   System

   ChatGPT

   Developer and deployer

   OpenAIUnited States

   Claim geography

   United Kingdom

   Case evidence: independent reporting4 case publications

   Updated 21 October 2026

   Compare benefit

   York student has a £60 parking fine withdrawn after sending an appeal letter drafted by ChatGPT

   Source and detail for York student has a £60 parking fine withdrawn after sending an appeal letter drafted by ChatGPT

   parking fine withdrawn by York City Council after the ChatGPT-drafted appeal

   Reported amount

   £60

   Currency

   The withdrawn parking fine concerns York, United Kingdom, and is reported in pounds sterling.

   USD equivalent

   ~USD 79.5. Converted from GBP at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   Millie Houlton

   Paid by

   Not established in the current source review.

   Unit

   one parking fine

   Period

   Case reported March 2023; exact withdrawal date absent.

   Amount basis

   Unresolved

   Accounting

   Gross personal payment avoided; transfer effect for the council and appeal costs are separate.

   Comparison

   Narrative baseline is the student paying the fine; no formal comparative design.

   AI's part

   AI drafted the appeal; council decision depended on a valid permit and other facts.

   Council decision and date

   AI contribution relative to entitlement

   Time and other appeal costs

   beneficiary_specific_transfer

   single_case

   counterfactual_self_report

   York student has a £60 parking fine withdrawn after sending an appeal letter drafted by ChatGPT

   A student has successfully appealed against a £60 parking fine by using a letter written by an artificial intelligence chatbot.

   This is a concrete personal financial benefit; it is not automatically a net social gain of the same amount.

   DIV-0012, AIBD-0116

7. 8 September 2023 Source published

   IRS applies AI and machine learning to select large partnerships for examination (IR-2023–166, 8 September 2023); two AI models prioritize partnership audits

   IRS applies AI to select large partnerships for examination

   ~$26.1 million

   Revenue

   Awaiting source reviewUSD

   1 claim source

   System

   IRS AI and machine-learning case-selection models (two partnership prioritization models; a refundable-credit audit selection model in pilot; machine learning for partnership, income and international tax risk)

   Developer

   Internal Revenue Service (in-house data scientists with tax-enforcement experts)United States

   Deployer

   Internal Revenue ServiceUnited States

   Claim geography

   United States

   Case evidence: independent study3 case publications

   Updated 21 October 2026

   Compare benefit

   IRS applies AI and machine learning to select large partnerships for examination (IR-2023–166, 8 September 2023); two AI models prioritize partnership audits

   Source and detail for IRS applies AI and machine learning to select large partnerships for examination (IR-2023–166, 8 September 2023); two AI models prioritize partnership audits

   additional tax assessments yearly that IRS management estimates from reassigning 14 classifiers to examinations after the Form 1040 return classification AI model automated classification (an SB/SE individual-return model, not the partnership models)

   Reported amount

   approximately $26.1 million

   Currency

   The projected tax assessments concern the U.S. Internal Revenue Service and are reported in U.S. dollars.

   Beneficiary

   the Internal Revenue Service (U.S. Treasury)

   Paid by

   Not established in the current source review.

   Unit

   per year after reassigning 14 classifiers

   Period

   Annual projection; start and forecast horizon unstated.

   Amount basis

   Unresolved

   Accounting

   Projected gross assessments, not collected cash or net return.

   Comparison

   No direct monetary comparator or realized collection evidence.

   AI's part

   The value concerns Form 1040 individual-return classification, while the record concerns partnership models.

   Link the claim to the correct model and case

   Projection method

   Actual examinations, assessments and receipts

   Program costs and forecast horizon

   claim_case_scope_mismatch

   assessed_not_collected

   projection

   labor_redeployment_mechanism

   IRS applies AI and machine learning to select large partnerships for examination (IR-2023–166, 8 September 2023); two AI models prioritize partnership audits

   According to IRS management, automating the classification function has allowed the IRS to reassign 14 classifiers to the examination function and they estimate the increased examinations could generate approximately $26.1 million in additional tax assessments yearly.

   Hold: the projected $26.1 million belongs to Form 1040 staff reassignment, not the partnership-model intervention in this AIBD case. A correct claim-to-case link is required.

   DIV-0013, AIBD-0338

8. Date unknown

   France's Foncier innovant aerial-image AI detects 140,000 undeclared pools and about €40 million in extra property tax by 20 March 2024; Dordogne alone 1,500 in 2023

   Foncier innovant finds 140,000 undeclared pools

   40 millions d’euros

   Revenue

   Awaiting source reviewEUR

   1 claim source

   System

   Foncier innovant (DGFiP deep-learning detection of pools and buildings on IGN aerial photographs, built with Capgemini and Google)

   Developer

   Direction générale des Finances publiques (DGFiP) with Capgemini and GoogleFrance

   Deployer

   Direction générale des Finances publiques (DGFiP), ministère de l'Économie et des FinancesFrance

   Claim geography

   France

   Case evidence: independent reporting5 case publications

   Updated 21 October 2026

   Compare benefit

   France's Foncier innovant aerial-image AI detects 140,000 undeclared pools and about €40 million in extra property tax by 20 March 2024; Dordogne alone 1,500 in 2023

   Source and detail for France's Foncier innovant aerial-image AI detects 140,000 undeclared pools and about €40 million in extra property tax by 20 March 2024; Dordogne alone 1,500 in 2023

   additional property tax (taxe foncière) to be collected on the 140,000 detected pools

   Reported amount

   40 millions d’euros

   Currency

   The property-tax amount concerns French local authorities and is reported in euros.

   USD equivalent

   ~USD 45.4 million. Converted from EUR at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   local authorities (collectivités), which receive the taxe foncière

   Paid by

   Not established in the current source review.

   Unit

   tax from 140,000 detected pools

   Period

   Reported March 2024; accounts disagree on the annual detection split; some describe 2023 tax.

   Amount basis

   Unresolved

   Accounting

   Expected gross tax revenue; deployment and verification costs are not deducted.

   Comparison

   Narrative comparison with manual surveying; no comparable revenue baseline.

   AI's part

   Minister attributes the figure to a tool plus human verification.

   Official tax authority figures

   Tax period and collection status

   Cost of verification and false positives

   No-tool revenue counterfactual

   assessed_not_collected

   period_disagreement

   operator_claim

   human_verification_cost_missing

   France's Foncier innovant aerial-image AI detects 140,000 undeclared pools and about €40 million in extra property tax by 20 March 2024; Dordogne alone 1,500 in 2023

   ce sont 140.000 piscines qui ont été détectées et qui permettront à nos collectivités de récolter 40 millions d’euros au titre de la taxe foncière

   Keep the expected revenue separate from actual collected revenue.

   DIV-0015, AIBD-0335

9. 12 November 2024 Source published

   Udbetaling Danmark's fraud-control models may have produced 135, 36 and 72 cases with benefits stopped or reduced in 2023

   UDK fraud models may have yielded 135, 36 and 72 cases

   half a billion Danish crowns

   Avoided losses

   Awaiting source reviewDKK

   1 claim source

   Monetary claim assessed 23 October 2026

   System

   Udbetaling Danmark / ATP Den Faelles Dataenhed fraud-control models (about 60 profiling and machine-learning models as of 2019, including 'Really Single', 'Model Abroad' and 'Fictitious Employment')

   Developer

   ATP (Arbejdsmarkedets Tillaegspension) with NNIT A/S as supplier, for Udbetaling DanmarkDenmark

   Deployer

   Udbetaling Danmark (public authority), administered by ATPDenmark

   Claim geography

   Denmark

   Case evidence: independent study7 case publications

   Updated 25 October 2026

   Compare benefit

   In a quest to optimize welfare management, Denmark built a surveillance behemoth

   Source and detail for Udbetaling Danmark's fraud-control models may have produced 135, 36 and 72 cases with benefits stopped or reduced in 2023

   erroneous payments unveiled by UDK and municipal controls in 2017 and 2018; the source says it is unclear how much was recovered and that the figure includes averted future payments, and it does not isolate the models

   Reported amount

   half a billion Danish crowns

   Currency

   The control aggregate concerns Danish public funds. The source reports DKK 500 million and an approximate EUR 70 million equivalent.

   USD equivalent

   ~USD 75.9 million. Converted from DKK at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   Udbetaling Danmark and the municipalities (the Danish public purse)

   Paid by

   Not established in the current source review.

   Unit

   UDK and municipal control aggregate

   Period

   2017 and 2018; the main case describes model outcomes in 2023.

   Amount basis

   Unresolved

   Accounting

   Mixture of recovery, unrecoverable amounts and future avoidance; net meaning unknown.

   Comparison

   No comparator.

   AI's part

   Historic all-control amount is not isolated to the named models or current outcome period.

   Recovered cash versus future avoided payments versus identified debt

   AI-model contribution

   Correct claim period and cohort

   Costs and claimant consequences

   mixed_accounting

   claim_case_period_mismatch

   whole_control_program

   dual_currency_same_amount

   estimate_flag_too_coarse

   In a quest to optimize welfare management, Denmark built a surveillance behemoth, 2020-08-06

   In 2017 and 2018, controls by UDK and municipalities unveiled half a billion Danish crowns in erroneous payment (approximately 70 million euros). It is unclear what amount was recovered, what was not recovered due to the beneficiary’s incapacity to repay, and what was an accounting trick (the figure includes future payments that were averted because of the control).

   91 referrals to the police

   Independent sourceSource checked 2026-10-23

   Source-stated equivalent: ~70 million euros (EUR)

   The source gives both currencies for the same amount. It does not specify an exchange-rate date or method.

   Reported amount

   approximately 70 million euros

   USD basis

   Converted from EUR at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Unit

   UDK and municipal control aggregate

   Period

   2017 and 2018; the main case describes model outcomes in 2023.

   Status

   Awaiting source review

   Accounting

   Mixture of recovery, unrecoverable amounts and future avoidance; net meaning unknown.

   Price basis

   Price year and inflation basis await review.

   In a quest to optimize welfare management, Denmark built a surveillance behemoth, 2020-08-06

   half a billion Danish crowns in erroneous payment (approximately 70 million euros)

   91 referrals to the police

   Source checked 2026-10-23

   The half-billion-crown amount concerns 2017–2018, while the linked model outcomes concern 2023. Recovered funds, future avoided payments, and the models' contribution remain unresolved.

   DIV-0020, AIBD-0336

10. 30 December 2024 Source published

    CGU ALICE alert leads to a cut of about R$100 million in the reference value of a federal road-works tender (30 December 2024)

    CGU's ALICE alert leads to a R$100 million cut in a road-works tender

    cerca de R$ 100 milhões

    Cost savings

    Awaiting source reviewBRL

    1 claim source

    System

    ALICE (Analisador de Licitações, Contratos e Editais), CGU text-mining and AI alert tool

    Developer

    Controladoria-Geral da União (CGU), in-houseBrazil

    Deployer

    Controladoria-Geral da União (CGU)Brazil

    Claim geography

    Brazil

    Case evidence: independent reporting3 case publications

    Updated 21 October 2026

    Compare benefit

    CGU ALICE alert leads to a cut of about R$100 million in the reference value of a federal road-works tender (30 December 2024)

    Source and detail for CGU ALICE alert leads to a cut of about R$100 million in the reference value of a federal road-works tender (30 December 2024)

    reduction in the reference value of a federal road-works engineering tender after the ALICE alert

    Reported amount

    cerca de R$ 100 milhões

    Currency

    The engineering-procurement amount concerns Brazilian federal public funds and is reported in Brazilian reais.

    USD equivalent

    ~USD 19.2 million. Converted from BRL at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

    Beneficiary

    the federal government (contracting agency not named)

    Paid by

    Not established in the current source review.

    Unit

    one road-works engineering tender

    Period

    Reported December 2024; actual case date is not established.

    Amount basis

    Unresolved

    Accounting

    Gross reduction of a reference budget; actual final expenditure and total costs unknown.

    Comparison

    Original reference price exists but is not quantified; reported delta is about R$100m.

    AI's part

    ALICE alert led to human audit and management adjustment.

    Original and revised prices

    Final award/payment and equivalent scope

    Case date and contracting body

    Overlap with annual R$257.25m aggregate

    budget_revision_not_cash_release

    approximate_amount

    possible_overlap_with_2024_aggregate

    event_date_missing

    CGU ALICE alert leads to a cut of about R$100 million in the reference value of a federal road-works tender (30 December 2024)

    Outro exemplo relevante dos benefícios decorrentes do uso da ferramenta foi a redução de cerca de R$ 100 milhões no valor de referência em um edital para contratação de serviços de engenharia.

    This is a monetary change, unlike a tender face value, but should be labeled reference-price cost avoidance.

    DIV-0021, AIBD-0087

11. 14 January 2025 Source published

    CGU's ALICE robot flags overpricing in three 2024 Rio de Janeiro federal hospital tenders; one suspended, one revoked, one re-estimated

    CGU's ALICE flags R$19.4m overpricing risk in three Rio tenders

    R$ 3.063.774,84

    Cost savings

    Awaiting source reviewBRL

    1 claim source

    System

    ALICE (Analisador de Licitações, Contratos e Editais), CGU text-mining and AI alert tool

    Developer

    Controladoria-Geral da União (CGU), in-houseBrazil

    Deployer

    Controladoria-Geral da União (CGU)Brazil

    Claim geography

    Brazil

    Case evidence: independent reporting5 case publications

    Updated 21 October 2026

    Compare benefit

    CGU's ALICE robot flags overpricing in three 2024 Rio de Janeiro federal hospital tenders; one suspended, one revoked, one re-estimated

    Source and detail for CGU's ALICE robot flags overpricing in three 2024 Rio de Janeiro federal hospital tenders; one suspended, one revoked, one re-estimated

    reduction of the estimated value of tender 90017/2024 after the re-estimate (the release gives no total for the three tenders)

    Reported amount

    R$ 3.063.774,84

    Currency

    The tender correction concerns Hospital Federal de Ipanema in Rio de Janeiro, Brazil, and is reported in Brazilian reais.

    USD equivalent

    ~USD 588,000. Converted from BRL at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

    Beneficiary

    Hospital Federal de Ipanema, federal budget

    Paid by

    Not established in the current source review.

    Unit

    tender 90017/2024 at Hospital Federal de Ipanema

    Period

    2024 tender; result reported January 14, 2025; precise revision date absent.

    Amount basis

    Unresolved

    Accounting

    Gross reduction of an estimate; actual paid saving and audit/system costs unknown.

    Comparison

    Explicit original R$15,450,714.80 versus revised R$12,386,939.96.

    AI's part

    ALICE alert followed by auditors and hospital managers.

    Final contract and payment

    Equivalent quantity and quality

    Revision date

    Avoid overlap with broader audits or annual totals

    budget_revision_not_cash_release

    single_component_of_three_cases

    possible_overlap_with_program_aggregates

    CGU's ALICE robot flags overpricing in three 2024 Rio de Janeiro federal hospital tenders; one suspended, one revoked, one re-estimated

    Como resultado da atuação, os gestores procederam nova estimativa de preços, reduzindo o valor total estimado da compra para R$ 12.386.939,96, o que representou uma economia de mais de R$ 3 milhões (R$ 3.063.774,84).

    The cited R$3,063,774.84 is a documented estimate change, not the sum of all three tenders or all flagged overpricing risks.

    DIV-0022, AIBD-0081

12. 19 March 2025 Source published

    Delhi Police matched a CCTV frame to a prior arrest record with Israeli facial recognition software, arrested two men within about a day of a Rs 80 lakh armed robbery in Chandni Chowk and recovered Rs 79.5 lakh

    Delhi Police facial recognition leads to Rs 79.5 lakh recovery

    Rs 79.5 lakh

    Avoided losses

    Awaiting source reviewINR

    1 claim source

    System

    Israeli facial recognition software used by Delhi Police (not named in the reports)

    Deployer

    Delhi Police, North districtIndia

    Claim geography

    India

    Case evidence: independent reporting5 case publications

    Updated 21 October 2026

    Compare benefit

    Delhi Police matched a CCTV frame to a prior arrest record with Israeli facial recognition software, arrested two men within about a day of a Rs 80 lakh armed robbery in Chandni Chowk and recovered Rs 79.5 lakh

    Source and detail for Delhi Police matched a CCTV frame to a prior arrest record with Israeli facial recognition software, arrested two men within about a day of a Rs 80 lakh armed robbery in Chandni Chowk and recovered Rs 79.5 lakh

    stolen cash recovered

    Reported amount

    Rs 79.5 lakh

    Currency

    The recovered cash concerns a Delhi robbery and is reported in Indian rupees.

    USD equivalent

    ~USD 82,800. Converted from INR at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

    Beneficiary

    the robbed cash courier's firm, R K Enterprises (recovered by Delhi Police from the accused's residence)

    Paid by

    Not established in the current source review.

    Unit

    one robbery and cash seizure

    Period

    Robbery March 17, 2025; recovery by March 19, 2025

    Amount basis

    Unresolved

    Accounting

    Gross cash recovered; return to the victim and investigation costs are not documented.

    Comparison

    Rs 80 lakh stolen and Rs 79.5 lakh recovered provide a recovery denominator; there is no no-AI counterfactual.

    AI's part

    Face matching was one investigative lead; conventional police work produced recovery.

    Cash returned to owner versus held as evidence

    Investigative costs

    AI incremental contribution

    recovery_not_expected_loss

    human_investigation_chain

    restoration_not_new_production

    Delhi Police matched a CCTV frame to a prior arrest record with Israeli facial recognition software, arrested two men within about a day of a Rs 80 lakh armed robbery in Chandni Chowk and recovered Rs 79.5 lakh

    A raid at Samir's residence led to the recovery of Rs 79.5 lakh in cash, a semi-automatic pistol with three bullets, and other incriminating materials.

    Rs 79.5 lakh is the recovery amount; Rs 80 lakh is the robbery total, not another benefit.

    DIV-0023, AIBD-0126

13. Date unknown

    HMRC's Connect data-matching system: about £4.6 billion of additional yield in 2024–25

    HMRC's Connect yields about £4.6 billion of extra tax in 2024–25

    ~£4.6bn

    Revenue

    Awaiting source reviewGBP

    1 claim source

    System

    HMRC Connect (data-matching and risk analytics) and TRUCE (Transaction Risking Upstream in a Connect Environment)

    Deployer

    HM Revenue & Customs (HMRC)United Kingdom

    Claim geography

    United Kingdom

    Case evidence: independent study4 case publications

    Updated 21 October 2026

    Compare benefit

    HMRC's Connect data-matching system: about £4.6 billion of additional yield in 2024–25

    Source and detail for HMRC's Connect data-matching system: about £4.6 billion of additional yield in 2024–25

    additional tax yield from Connect cases in the 2024-25 tax year (average £3.4bn a year)

    Reported amount

    £4.6bn

    Currency

    The revenue claim concerns HMRC and the U.K. Exchequer and is reported in pounds sterling.

    USD equivalent

    ~USD 6.09 billion per year. Converted from GBP at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

    Beneficiary

    HMRC, for the Exchequer, gains

    Paid by

    Not established in the current source review.

    Unit

    annual yield from Connect cases

    Period

    2024-25 tax year

    Amount basis

    Unresolved

    Accounting

    Gross additional yield is suggested; assessment, actual cash and future-benefit components are not separated.

    Comparison

    No monetary comparator beyond an unstated multi-year average.

    AI's part

    Yield attributed to Connect-supported cases; existing sources describe analytics/risk rules without establishing machine learning.

    FOI response and yield definition

    Actual receipts versus assessments or future revenue

    Identifiable AI component

    System and caseworker costs

    yield_definition_missing

    ai_component_needs_review

    whole_program_attribution

    average_not_additive

    HMRC's Connect data-matching system: about £4.6 billion of additional yield in 2024–25

    HMRC, in a freedom of information request to Pinsent Masons, said it generated on average £3.4bn in additional annual yield from Connect cases, with the 2024-25 tax year generating approximately £4.6bn.

    Do not add the £3.4bn average annual figure to the specific-year £4.6bn figure.

    DIV-0024, AIBD-0145

14. 30 June 2025 Source published

    Medicare payments withheld in Operation Gold Rush

    Medicare withholds $4.41 billion in suspected fraudulent payments

    US$4.41 billion in escrow

    Avoided losses

    Withheld in escrowUSD

    1 claim source

    Monetary claim assessed 23 October 2026

    System

    DOJ Health Care Fraud Unit proactive billing analytics

    Developer

    U.S. Department of Justice, Criminal Division Fraud Section, Health Care Fraud Unit Data Analytics TeamUnited States

    Deployer

    U.S. Department of Justice with HHS-OIG and CMSUnited States

    Claim geography

    United States

    Case evidence: party or catalog claim4 case publications

    Updated 25 October 2026

    Compare benefit

    2025 National Health Care Fraud Takedown

    Source and detail for Medicare payments withheld in Operation Gold Rush

    DOJ reports that $4.41 billion was held in escrow; the source describes data analytics without establishing that AI was used.

    Reported amount

    US$4.41 billion in escrow

    Currency

    The amount concerns the U.S. Medicare trust fund and is reported in U.S. dollars.

    Beneficiary

    Medicare trust fund (HHS and CMS)

    Paid by

    Not established in the current source review.

    Unit

    Operation Gold Rush payment block

    Period

    Reported June 30, 2025; transaction interval not fully specified.

    Amount basis

    Avoided expenditure

    Accounting

    Scheduled payments withheld in escrow, not cash already returned to the Medicare trust fund.

    Comparison

    Scheduled payment about $4.45bn versus about $41m paid, with about $4.41bn escrowed.

    AI's part

    Detection is attributed to proactive data analytics; the record states that AI/ML use is unestablished.

    Confirm final escrow disposition

    Identify any actual AI component

    Payment dates and overlap with other fraud aggregates

    Costs

    The source does not identify an AI component in the detection.

    The amount was in escrow, with return to the trust fund still intended.

    held_not_recovered

    ai_component_needs_review

    intended_loss_not_paid_loss

    rounded_amounts

    Included in broader Medicare prevented-payment accounts; not additive to those totals.

    2025 National Health Care Fraud Takedown, 2025-06-30

    HHS and CMS intend to seek to return the $4.41 billion in escrow to the Medicare trust fund

    Reported by the deployerSource checked 2026-10-23

    The monetary amount is supported, but the cited source does not identify AI in the detection. The future AI-enabled Fusion Center is a separate initiative.

    DIV-0026, AIBD-0187

15. Date unknown

    Sheng Siong's facial-recognition CCTV, given a shoplifter's face after a 10 September 2025 stock check at its Punggol Central outlet, matches her to six earlier wine thefts and alerts the manager's phone when she returns at noon the next day; staff stop her with three unscanned bottles, police arrest her, she repays the chain and is jailed eight days on 12 May 2026

    Sheng Siong's facial recognition catches a repeat wine thief

    ~S$470

    Avoided losses

    Awaiting source reviewSGD

    1 claim source

    System

    Sheng Siong's AI-driven facial recognition closed-circuit television system: a suspect's photograph, taken from CCTV footage after a theft is found, is uploaded, the system matches the face across footage and sends mobile alerts through a corporate app to staff when that person enters any Sheng Siong outlet

    Deployer

    Sheng Siong (supermarket chain, 83 outlets), SingaporeSingapore

    Claim geography

    Singapore

    Case evidence: independent reporting4 case publications

    Updated 25 October 2026

    Compare benefit

    Sheng Siong's facial-recognition CCTV, given a shoplifter's face after a 10 September 2025 stock check at its Punggol Central outlet, matches her to six earlier wine thefts and alerts the manager's phone when she returns at noon the next day; staff stop her with three unscanned bottles, police arrest her, she repays the chain and is jailed eight days on 12 May 2026

    Source and detail for Sheng Siong's facial-recognition CCTV, given a shoplifter's face after a 10 September 2025 stock check at its Punggol Central outlet, matches her to six earlier wine thefts and alerts the manager's phone when she returns at noon the next day; staff stop her with three unscanned bottles, police arrest her, she repays the chain and is jailed eight days on 12 May 2026

    restitution paid by the shoplifter for the stolen bottles that were not recovered (three bottles were recovered at the interception)

    Reported amount

    about S$470

    Currency

    The restitution concerns a Sheng Siong store in Punggol, Singapore, and is reported in Singapore dollars.

    USD equivalent

    ~USD 368. Converted from SGD at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

    Beneficiary

    Sheng Siong

    Paid by

    Not established in the current source review.

    Unit

    one defendant and series of thefts

    Period

    Thefts September 2-11, 2025; restitution reported by May 12, 2026.

    Amount basis

    Unresolved

    Accounting

    Gross restitution; investigation and system costs not deducted.

    Comparison

    Manual-review versus alert workflow; no monetary counterfactual.

    AI's part

    AI matched prior thefts and return; human interception, prosecution and restitution followed.

    Date and proof of restitution

    AI contribution relative to human detection

    Operating and enforcement costs

    restitution_transfer

    single_series_of_events

    different_recovery_dates

    Sheng Siong's facial-recognition CCTV, given a shoplifter's face after a 10 September 2025 stock check at its Punggol Central outlet, matches her to six earlier wine thefts and alerts the manager's phone when she returns at noon the next day; staff stop her with three unscanned bottles, police arrest her, she repays the chain and is jailed eight days on 12 May 2026

    The retail manager called the police, and the three bottles of wine were recovered. Tan also made full restitution of about S$470 for the remaining stolen bottles.

    The S$470 excludes three recovered bottles; retain distinct cash and in-kind recoveries.

    DIV-0029, AIBD-0207

16. 8 October 2025 Source published

    Lynn White of Long Beach, California, who lost an eviction jury trial with a lawyer, may have overturned the eviction on appeal with ChatGPT and Perplexity, which she used to find procedural errors and draft it, avoiding roughly $55,000 in penalties and over $18,000 in back rent

    ChatGPT and Perplexity may have won Lynn White her eviction appeal

    roughly $55,000

    Cost savings

    Awaiting source reviewUSD

    1 claim source

    System

    ChatGPT (paid plan, $20 a month) and Perplexity Pro ($20 a month)

    Developer

    OpenAI (ChatGPT); Perplexity AI (Perplexity)United States

    Claim geography

    United States

    Case evidence: independent reporting4 case publications

    Updated 21 October 2026

    Compare benefit

    Lynn White of Long Beach, California, who lost an eviction jury trial with a lawyer, may have overturned the eviction on appeal with ChatGPT and Perplexity, which she used to find procedural errors and draft it, avoiding roughly $55,000 in penalties and over $18,000 in back rent

    Source and detail for Lynn White of Long Beach, California, who lost an eviction jury trial with a lawyer, may have overturned the eviction on appeal with ChatGPT and Perplexity, which she used to find procedural errors and draft it, avoiding roughly $55,000 in penalties and over $18,000 in back rent

    penalties avoided, plus more than $18,000 in overdue rent, by overturning the eviction on appeal

    Reported amount

    roughly $55,000

    Currency

    The avoided penalties concern a tenant in Long Beach, California, and are reported in U.S. dollars.

    Beneficiary

    Lynn White

    Paid by

    Not established in the current source review.

    Unit

    one eviction appeal

    Period

    Reported October 8, 2025 after several months; ruling date unknown.

    Amount basis

    Unresolved

    Accounting

    Personal gross liability reduction, not net legal benefit; costs and counterparty effects unstated.

    Comparison

    Prior jury loss versus later appeal success; nonexperimental and not an equivalent legal comparison.

    AI's part

    AI assisted research and drafting; legal rules and personal advocacy also contributed.

    Court ruling and actual legal effect

    Whether rent was waived or deferred

    Separate penalties and rent

    Case date and full litigation costs

    multiple_amount_components

    court_record_missing

    deferral_vs_extinguishment_unclear

    beneficiary_specific_transfer

    Lynn White of Long Beach, California, who lost an eviction jury trial with a lawyer, may have overturned the eviction on appeal with ChatGPT and Perplexity, which she used to find procedural errors and draft it, avoiding roughly $55,000 in penalties and over $18,000 in back rent

    Several months of litigation later, White managed to overturn her eviction notice and avoid roughly $55,000 in penalties and more than $18,000 in overdue rent.

    The amount field covers roughly $55,000; what/quote separately describe over $18,000 rent. Keep both claims separate pending legal documentation.

    DIV-0030, AIBD-0150

17. 16 January 2026 Source published

    Ghana Revenue Authority launches Publican AI trade analytics, which may have recovered GH¢15 million

    GRA's Publican AI may have recovered GH¢15 million

    GH¢15 million

    Revenue

    Awaiting source reviewGHS

    1 claim source

    System

    Publican AI Trade Solution

    Developer

    A Cyprus-based firm, unnamed in the read article

    Deployer

    Ghana Revenue Authority, Customs DivisionGhana

    Claim geography

    Ghana

    Case evidence: independent reporting3 case publications

    Updated 21 October 2026

    Compare benefit

    Ghana Revenue Authority launches Publican AI trade analytics, which may have recovered GH¢15 million

    Source and detail for Ghana Revenue Authority launches Publican AI trade analytics, which may have recovered GH¢15 million

    recovered from fraudulent importers

    Reported amount

    GH¢15 million

    Currency

    The recovered import revenue concerns Ghana Revenue Authority in Tema and is reported in Ghanaian cedis.

    USD equivalent

    ~USD 1.29 million. Converted from GHS at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

    Beneficiary

    Government of Ghana (Ghana Revenue Authority)

    Paid by

    Not established in the current source review.

    Unit

    program recovery to date

    Period

    Reported January 2026; recovery interval unspecified.

    Amount basis

    Unresolved

    Accounting

    Gross recovery claim; costs and refund/reversal adjustments are absent.

    Comparison

    Assessment-time comparison only; claimed revenue-gain percentages refer to pilots.

    AI's part

    AI analytics claim and recovery appear in a registry summary; source account does not establish mechanism or incremental share.

    Original GhanaWeb source and authority confirmation

    Recovery dates and accounting definition

    AI role

    Costs

    source_chain_needs_review

    recovery_period_missing

    pilot_vs_rollout_timing

    Ghana Revenue Authority launches Publican AI trade analytics, which may have recovered GH¢15 million

    The system has already enabled the government to recover GH¢15 million from fraudulent importers and aims to address significant revenue leakages in Ghana's customs operations.

    The 40-45% pilot revenue claim and the GH¢15m recovered need separate claim and cohort identifiers.

    DIV-0033, AIBD-0181

18. 6 March 2026 Source published

    Gyeonggi ARES AI pig-weighing pilot cuts shipping age 5.6 days and lifts income per head 18% on six farms

    Gyeonggi AI pig-weighing pilot lifts income per head 18%

    KRW118,000 per pig (reported level)

    Other monetary claim

    Awaiting source reviewKRW

    1 claim source

    System

    Fattening-pig shipping selection system (AI non-contact weighing) and sow farrowing management system (ICT monitoring)

    Developer

    Not named in the sources

    Deployer

    Gyeonggi-do Agricultural Research and Extension Services with six pig farms in Anseong, Pocheon and Yeoju (weighing) and Pocheon farms (sow system)South Korea

    Claim geography

    South Korea

    Case evidence: independent reporting3 case publications

    Updated 21 October 2026

    Compare benefit

    Gyeonggi ARES AI pig-weighing pilot cuts shipping age 5.6 days and lifts income per head 18% on six farms

    Source and detail for Gyeonggi ARES AI pig-weighing pilot cuts shipping age 5.6 days and lifts income per head 18% on six farms

    income per pig shipped with the AI shipping selection system, up from 10만 원 (100,000 won) per head, an 18% increase

    Reported amount

    KRW118,000 per pig (reported level)

    Currency

    The income figure concerns pilot pig farms in Anseong, Pocheon and Yeoju, South Korea, and is reported in South Korean won.

    USD equivalent

    ~USD 87.2 per pig. Converted from KRW at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

    Beneficiary

    the six pilot pig farms in Anseong, Pocheon and Yeoju

    Paid by

    Not established in the current source review.

    Unit

    per pig shipped across six pilot farms

    Period

    2025 pilot, reported March 2026; within-year observation length unspecified.

    Amount basis

    Outcome level

    Accounting

    The source term income does not establish revenue, gross margin or net profit.

    Comparison

    Explicit 100,000 won before versus 118,000 won after, an 18% increase.

    AI's part

    Uncontrolled before/after pilot combines weighing, shipping advice and farm responses.

    Definition of income and included costs

    Farm/sample weighting and observation period

    Other changes and market prices

    System cost

    outcome_level_not_increment

    income_definition_missing

    per_animal

    uncontrolled_before_after

    Gyeonggi ARES AI pig-weighing pilot cuts shipping age 5.6 days and lifts income per head 18% on six farms

    출하일령이 182.3일에서 176.7일로 5.6일 줄었다. 또 1+등급 출하율은 35.3%에서 37.7%으로 2.4%p 증가하고, 두당 소득이 10만 원에서 11만 8000원으로 18% 늘었다.

    The KRW118,000 figure is a post-intervention per-pig income level, not the increment. KRW100,000 is the reported earlier level; any derived difference needs a verified common basis and attribution review.

    DIV-0034, AIBD-0202

19. 17 June 2026 Source published

    Retirement holdings in TIAA’s reported scam intervention

    TIAA’s AI may have helped stop a scam involving retirement savings

    over $3 million

    Avoided losses

    Awaiting source reviewUSD

    1 claim source

    Monetary claim assessed 24 October 2026

    System

    Unnamed TIAA AI tool used to flag suspected fraud for human review

    Deployer

    TIAA (Teachers Insurance and Annuity Association of America)United States

    Claim geography

    United States

    Case evidence: party or catalog claim1 case publication

    Date assessed 24 October 2026 Updated 24 October 2026

    Compare benefit

    Interview with Thasunda Brown Duckett

    Source and detail for Retirement holdings in TIAA’s reported scam intervention

    Reported holdings; the attempted withdrawal amount is unspecified.

    Reported amount

    over $3 million

    Currency

    The reported retirement-account amount is in U.S. dollars. The TIAA participant's place of residence is not reported.

    Beneficiary

    the 76-year-old TIAA participant

    Paid by

    Not established in the current source review.

    Unit

    one participant’s retirement holdings

    Period

    Recently before June 17, 2026; exact date unstated.

    Amount basis

    Unresolved

    Accounting

    Reported holdings; the attempted withdrawal amount is unspecified.

    Comparison

    No counterfactual; the withdrawal could also have triggered ordinary review.

    AI's part

    AI flag started a human intervention chain; the CEO states that AI alone would not necessarily have protected the participant.

    Independent case confirmation

    Event date, attempted withdrawal amount, and transfer records

    Counterfactual risk and recovery potential

    Intervention cost

    holdings_not_verified_avoided_loss

    single_unnamed_case

    operator_anecdote

    human_intervention_chain

    Interview with Thasunda Brown Duckett, 2026-06-22

    had over $3 million

    “How TIAA is approaching AI—and whether it threatens the retirement market,” participant example.

    Reported by the deployerSource checked 2026-10-24

    Withdrawal amount unverified.

    DIV-0035, AIBD-0210

No matching claims.

Context

7 amounts that describe costs, scale, or setting.

1. 27 September 2017 Source published

   TCU's Alice robot flags DNIT notices for BR-153 maintenance on stretches already under concession; TCU Acórdão 1116/2017 cancels the tender

   TCU's Alice flags DNIT's BR-153 notices; TCU cancels the tender

   R$ 39 milhões

   Contextual amount

   Context onlyBRL

   1 claim source

   System

   Alice (Análise de Licitações e Editais), TCU/CGU text-analysis and data cross-checking robot

   Developer

   Tribunal de Contas da União (TCU) with the Ministério da Transparência e Controladoria-Geral da União (CGU), in-houseBrazil

   Deployer

   Tribunal de Contas da União (TCU), Goiás unitBrazil

   Claim geography

   Brazil

   Case evidence: independent reporting3 case publications

   Updated 21 October 2026

   Compare benefit

   TCU's Alice robot flags DNIT notices for BR-153 maintenance on stretches already under concession; TCU Acórdão 1116/2017 cancels the tender

   Source and detail for TCU's Alice robot flags DNIT notices for BR-153 maintenance on stretches already under concession; TCU Acórdão 1116/2017 cancels the tender

   value of the two DNIT tender notices for BR-153 maintenance that a TCU acórdão blocked

   Reported amount

   R$ 39 milhões

   Currency

   The tenders concern federal highway maintenance in Goiás, Brazil, with amounts reported in Brazilian reais.

   USD equivalent

   ~USD 7.48 million. Converted from BRL at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   DNIT's Goiás unit, which had published the tenders for stretches whose maintenance falls to the concessionaire

   Paid by

   Not established in the current source review.

   Unit

   two tender notices

   Period

   Tenders reported in September 2017; precise occurrence dates absent.

   Amount basis

   Reference

   Accounting

   A contract or budget value is not a gross or net benefit.

   Comparison

   No expenditure reduction comparator.

   AI's part

   AI flagged tenders, but the amount describes their value rather than a demonstrated monetary gain.

   Reconcile R$39 million across two notices with R$20 million for one tender

   Original court/audit decision

   What payments actually would have occurred and were avoided

   contract_value_not_saving

   conflicting_scope_and_amount

   ai_increment_unresolved

   TCU's Alice robot flags DNIT notices for BR-153 maintenance on stretches already under concession; TCU Acórdão 1116/2017 cancels the tender

   Em Goiás, o Tribunal de Contas, via Alice, descobriu que o DNIT estdual publicou dois editais, de R$ 39 milhões, para manutenção da BR 153, mas em trechos que estão sob concessão – portanto com manutenção a cargo da concessionária.

   DIV-0004, AIBD-0331

2. Date unknown

   TCE-MG's Alice audit of 2,297 August 2022 tenders detects 60 irregularities worth R$113.7 million; 53 alerts answered and notices worth R$91.7 million corrected

   Alice audit at TCE-MG corrects notices worth R$91.7 million

   R$ 91.723.011,31

   Contextual amount

   Context onlyBRL

   1 claim source

   System

   Alice (Análise de Licitações e Editais), the CGU/TCU tender-analysis system as adopted by TCE-MG

   Developer

   Controladoria-Geral da União (CGU) and Tribunal de Contas da União (TCU)Brazil

   Deployer

   Tribunal de Contas do Estado de Minas Gerais (TCE-MG)Brazil

   Claim geography

   Brazil

   Case evidence: party or catalog claim3 case publications

   Updated 21 October 2026

   Compare benefit

   TCE-MG's Alice audit of 2,297 August 2022 tenders detects 60 irregularities worth R$113.7 million; 53 alerts answered and notices worth R$91.7 million corrected

   Source and detail for TCE-MG's Alice audit of 2,297 August 2022 tenders detects 60 irregularities worth R$113.7 million; 53 alerts answered and notices worth R$91.7 million corrected

   value of the tender notices corrected after Alice's alerts in the audit of August 2022 tenders

   Reported amount

   R$ 91.723.011,31

   Currency

   The tender corrections concern state and municipal bodies in Minas Gerais, Brazil, and are reported in Brazilian reais.

   USD equivalent

   ~USD 17.6 million. Converted from BRL at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   the audited state and municipal bodies in Minas Gerais, which republished corrected notices; the page's title calls the result a saving without saying who saved

   Paid by

   Not established in the current source review.

   Unit

   corrected notices in one audit cohort

   Period

   August 2022 tender cohort; audit described in December 2022 and reported June 2023.

   Amount basis

   Reference

   Accounting

   Value of corrected notices, not a gross or net saving.

   Comparison

   No actual expenditure reduction or monetary counterfactual.

   AI's part

   ALICE alerts supported corrections; no benefit amount is isolated.

   Amounts changed, canceled, saved or recovered

   Final procurement outcomes

   Separate corrected-contract value from benefit value

   contract_value_not_saving

   headline_accounting_mismatch

   audit_cohort

   TCE-MG's Alice audit of 2,297 August 2022 tenders detects 60 irregularities worth R$113.7 million; 53 alerts answered and notices worth R$91.7 million corrected

   Dos 60 alertas encaminhados aos jurisdicionados, 53 apresentaram respostas que permitiram a verificação do controle, alcançando a correção das irregularidades na republicação dos editais ou adoção de providências em futuros procedimentos licitatórios. Com isso, a auditoria conseguiu a correção dos editais equivalente a R$ 91.723.011,31.

   DIV-0011, AIBD-0332

3. Date unknown

   Cheshire Police facial recognition matches CCTV to Keenan Hughes; county lines dealers jailed for over five years

   Cheshire facial recognition match jails county lines dealers

   ~£990

   Contextual amount

   Context onlyGBP

   1 claim source

   System

   Cheshire Constabulary facial recognition system (retrospective)

   Developer

   Not named in the sources

   Deployer

   Cheshire Constabulary, Northwich Proactive Policing UnitUnited Kingdom

   Claim geography

   United Kingdom

   Case evidence: independent reporting4 case publications

   Updated 21 October 2026

   Compare benefit

   Cheshire Police facial recognition matches CCTV to Keenan Hughes; county lines dealers jailed for over five years

   Source and detail for Cheshire Police facial recognition matches CCTV to Keenan Hughes; county lines dealers jailed for over five years

   crack cocaine seized from Hughes (99 wraps, approximate street value)

   Reported amount

   £990

   Currency

   The seized-drug valuation concerns Northwich, United Kingdom, and is reported in pounds sterling.

   USD equivalent

   ~USD 1,310. Converted from GBP at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   Cheshire Constabulary's Northwich Proactive Policing Team (seized from Keenan Hughes)

   Paid by

   Not established in the current source review.

   Unit

   99 wraps seized in one arrest

   Period

   May 17, 2024

   Amount basis

   Reference

   Accounting

   Contraband valuation, not a cash saving or net social benefit.

   Comparison

   Identification-speed comparison, not a monetary benefit comparator.

   AI's part

   Face match helped identify a suspect; seizure followed patrol and investigation.

   A defensible monetary benefit endpoint if one is sought

   Valuation basis

   Costs and consequences of intervention

   seized_goods_value_not_benefit

   estimated_street_price

   estimate_flag_too_coarse

   Cheshire Police facial recognition matches CCTV to Keenan Hughes; county lines dealers jailed for over five years

   99 individual wraps of crack cocaine, worth approximately £990

   DIV-0016, AIBD-0232

4. Date unknown

   Taoyuan's AI sound cameras locate a white BMW that had been roaring down Jieshou Road at night, within a day of residents' complaint

   Taoyuan's AI sound cameras pinpoint a noisy BMW within a day

   ≤ 3,600元 (New Taiwan dollars)

   Contextual amount

   Context onlyTWD

   1 claim source

   System

   AI sound camera equipment (AI聲音照相) of the Quiet Taoyuan (靜桃) project, part of the bureau's network of 'improved mobile sound cameras' (改良式移動聲音照相)

   Deployer

   Taoyuan City Department of Environmental Protection (桃園市政府環境保護局), with the Taoyuan City Police Department for the stopTaiwan

   Claim geography

   Taiwan

   Case evidence: independent reporting5 case publications

   Updated 21 October 2026

   Compare benefit

   Taoyuan's AI sound cameras locate a white BMW that had been roaring down Jieshou Road at night, within a day of residents' complaint

   Source and detail for Taoyuan's AI sound cameras locate a white BMW that had been roaring down Jieshou Road at night, within a day of residents' complaint

   maximum fine under the Noise Control Act to be imposed for the noise violation

   Reported amount

   3,600元 (New Taiwan dollars)

   Currency

   The fine ceiling concerns Taoyuan, Taiwan, and is reported in New Taiwan dollars.

   USD equivalent

   ≤ ~USD 113. Converted from TWD at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   the BMW's owner pays

   Paid by

   Not established in the current source review.

   Unit

   one noise violation; repeat penalties possible

   Period

   Violation detected April 23, 2025; future fine mentioned.

   Amount basis

   Reference

   Accounting

   Legal ceiling, not realized gross receipt or monetized community benefit.

   Comparison

   Identification-speed comparison only.

   AI's part

   AI-assisted detection may lead to a fine; no assessed or paid amount is established.

   Actual assessed penalty

   Collection status

   Measured noise or welfare improvement

   maximum_not_expected_value

   possible_penalty_not_outcome_amount

   transfer_perspective

   Taoyuan's AI sound cameras locate a white BMW that had been roaring down Jieshou Road at night, within a day of residents' complaint

   後續將依噪音管制法進行裁罰，最高可處3,600元罰鍰，並通知限期改善，屆期未改善者將按次處罰直至符合機動車輛噪音管制標準為止。

   DIV-0025, AIBD-0129

5. Date unknown

   Taichung EPB's AI smoke-recognition system catches a serial roadside waste burner in Dadu District

   Taichung EPB's AI smoke detection catches a serial waste burner

   新臺幣10萬元以下 (up to NT$100,000)

   Contextual amount

   Context onlyTWD

   1 claim source

   System

   AI智慧判煙系統 (AI smart smoke-recognition system) with air micro-sensors and surveillance cameras, operated by the Taichung City Environmental Protection Bureau; 36 hotspot monitoring stations as of July 2025 and over 100 AI and sensor sites citywide by September 2025

   Deployer

   Taichung City Government Environmental Protection Bureau (臺中市政府環境保護局)Taiwan

   Claim geography

   Taiwan

   Case evidence: independent reporting11 case publications

   Updated 21 October 2026

   Compare benefit

   Taichung EPB's AI smoke-recognition system catches a serial roadside waste burner in Dadu District

   Source and detail for Taichung EPB's AI smoke-recognition system catches a serial roadside waste burner in Dadu District

   fine under Article 32 of the Air Pollution Control Act for the open burning; CTS reports the fine as NT$100,000

   Reported amount

   新臺幣10萬元以下 (up to NT$100,000)

   Currency

   The fine ceiling concerns Taichung, Taiwan, and is reported in New Taiwan dollars.

   USD equivalent

   ≤ ~USD 3,140. Converted from TWD at October 2026 rates (InforEuro monthly accounting rates). Original unit, period, and qualifier retained.

   Beneficiary

   the 74-year-old burner pays

   Paid by

   Not established in the current source review.

   Unit

   one enforcement case

   Period

   Detections July-August 2025; announcement September 13, 2025.

   Amount basis

   Reference

   Accounting

   Legal maximum or reported fine, not a valuation of avoided pollution.

   Comparison

   No monetary comparator.

   AI's part

   AI detected smoke; identification and enforcement depended on human investigation.

   Resolve up-to versus actual NT$100,000 penalty

   Payment status

   Environmental outcome and any defensible valuation

   maximum_vs_imposed_conflict

   possible_penalty_not_outcome_amount

   transfer_perspective

   Taichung EPB's AI smoke-recognition system catches a serial roadside waste burner in Dadu District

   依違反空氣污染防制法第32條規定，處新臺幣10萬元以下罰鍰

   DIV-0027, AIBD-0147

6. 8 August 2025 Source published

   DARPA's AIxCC final: autonomous systems find 54 of 63 synthetic vulnerabilities, patch 43, and surface 18 real bugs in open-source code

   AIxCC systems find 54 of 63 synthetic vulnerabilities, patch 43

   $152

   Contextual amount

   Context onlyUSD

   1 claim source

   System

   AIxCC finalist cyber reasoning systems (Team Atlanta, Trail of Bits, Theori and four others)

   Developer

   Seven finalist teams (Team Atlanta et al.), using LLM credits from Anthropic, Google and OpenAI

   Deployer

   DARPA (with ARPA-H), AIxCC Final CompetitionUnited States

   Claim geography

   United States

   Case evidence: independent reporting3 case publications

   Updated 21 October 2026

   Compare benefit

   DARPA's AIxCC final: autonomous systems find 54 of 63 synthetic vulnerabilities, patch 43, and surface 18 real bugs in open-source code

   Source and detail for DARPA's AIxCC final: autonomous systems find 54 of 63 synthetic vulnerabilities, patch 43, and surface 18 real bugs in open-source code

   average cost per competition task for the systems' bug reports and patches, which DARPA calls a fraction of the cost of traditional methods

   Reported amount

   $152

   Currency

   The task-cost figure concerns the U.S. DARPA AIxCC competition and is reported in U.S. dollars.

   Beneficiary

   DARPA (cost of running the competitors' cyber reasoning systems per task)

   Paid by

   Not established in the current source review.

   Unit

   per competition task

   Period

   AIxCC final competition in 2025

   Amount basis

   Reference

   Accounting

   A cost input; inclusions such as training, development and staff time are unknown.

   Comparison

   Structured comparator concerns prior-round vulnerability detection; traditional monetary cost is only described qualitatively.

   AI's part

   Cost of running the cyber-reasoning systems, not a monetary benefit amount.

   Traditional cost for comparable tasks

   Which cost components are included

   Task success and quality denominators

   Generalizability beyond competition

   cost_not_benefit

   per_task

   comparison_cost_missing

   synthetic_task_scope

   DARPA's AIxCC final: autonomous systems find 54 of 63 synthetic vulnerabilities, patch 43, and surface 18 real bugs in open-source code

   Competitor CRSs proved they can create valuable bug reports and patches for a fraction of the cost of traditional methods, with an average cost per competition task of about $152.

   DIV-0028, AIBD-0238

7. 20 October 2026 Source published

   VTTI/VDOT thermal-camera AI wildlife detection pilot on Route 460 near Grundy, Virginia: 97% detection accuracy and a 3 mph average speed drop at activated signs

   Drivers slowed down when signs warned of wildlife ahead

   $80,771.49 per elk crash

   Contextual amount

   Context onlyUSD

   2 claim sources

   Monetary claim assessed 24 October 2026

   System

   Animal Detection and Driver Warning System (ADDWS), using Crosstek thermal-camera AI detection

   Developer

   Crosstek (cameras and detection software); system integration and evaluation by Virginia Tech Transportation InstituteUnited States

   Deployer

   Virginia Department of Transportation with Virginia Tech Transportation InstituteUnited States

   Claim geography

   United States

   Case evidence: independent reporting4 case publications

   Date assessed 24 October 2026 Updated 24 October 2026

   Compare benefit

   Donaldson et al. An Evaluation of Wildlife Crossing Design, Placement, Costs, and Funding Opportunities for Corridor Q. VTRC 24-R8, September 2023.

   Source and detail for VTTI/VDOT thermal-camera AI wildlife detection pilot on Route 460 near Grundy, Virginia: 97% detection accuracy and a 3 mph average speed drop at activated signs

   Estimated comprehensive societal cost per elk crash, including economic losses and monetized pain and suffering. No AI-attributable avoided crash or saving is established.

   Reported amount

   $80,771.49 per elk crash

   Currency

   The original analysis estimates $80,771.49 in comprehensive societal cost per elk crash for the Virginia project. This remains context only.

   Beneficiary

   Road users and society bear the economic and human costs represented by this valuation.

   Paid by

   Not established in the current source review.

   Unit

   per elk collision

   Period

   Crash records from five states cover different periods between 2009 and 2021. The estimate was published in September 2023 and uses 2020 U.S. dollar valuations.

   Amount basis

   Reference

   Accounting

   Comprehensive societal valuation per elk crash. Includes tangible economic impacts and monetized pain and suffering. This is neither a cash payment nor a measured benefit of AI.

   Comparison

   Measured speed reduction, but the record explicitly states that prevented crashes were not established.

   AI's part

   AI triggers alerts; a crash-reduction or loss-avoidance quantity is not supplied.

   No quantity of collisions prevented by AI

   Severity mix drawn from reported crashes in five other states

   Monetized pain and suffering is not a cash expense

   System costs and any net AI benefit not established

   For 104 of 171 California crash records, injury severity was missing and was treated as property damage only in the calculation.

   harm_unit_cost_not_avoided_loss

   missing_avoided_quantity

   estimate_not_benefit_projection

   Donaldson et al. An Evaluation of Wildlife Crossing Design, Placement, Costs, and Funding Opportunities for Corridor Q. VTRC 24-R8, September 2023.

   Average elk crash cost

   pp. 41–42, table 12 and note c; the calculated value is $80,771.49.

   Reported by the deployerSource checked 2026-10-24

   Cole, Mark A. Memorandum: VDOT Traffic Crash Costs—Version 1. Virginia Department of Transportation, March 2, 2022., 2022-03-02

   Virginia KABCO Crash Unit Costs (2020)

   1, tables 1–2; 3, table 6 and price adjustment method. Preserved copy of the original VDOT memorandum.

   Reported by the deployerSource checked 2026-10-24

   DIV-0037, AIBD-0359

No matching amounts.