Skip to content

Reviewed edition · qasb report 10/2026 · Topic report · Public sector

AI in the public sector: benefits mostly estimated rather than measured

Audit reports and official transparency records show AI in justice, welfare and tax administration in the UK and the US. The benefits are mostly estimated by the agencies themselves and rarely measured.

Published
October 27, 2026 · as of October 26
Reading time
12 min
Publisher
qasb Research
Method
qasb methodology v1

Key findings

  1. 266,667 hsaved so far by the AI note-taking tool Justice Transcribe, as estimated by the agency HMPPS, with no measured data.1Evidence: model estimate
  2. 20%of examined final case records in the pilot had inaccuracies, mainly in names, with no comparison rate.1Evidence: observed
  3. 715 casesa month were additionally kept from maintenance arrears by model-prompted action, a DWP estimate with no control group.2Evidence: model estimate
  4. 92%of closed examinations in the IRS program with AI selection ended in no change (preliminary), share of AI cases unknown.6Evidence: observed

1.0Summary

The NAO, GAO and TIGTA audit reports are the strongest evidence. Only one study is randomised, a retinal screening trial in Tanzania. For Germany and the EU, outcome measurements and dated evidence on market surveillance under the AI Act are missing from the material checked.

  • The largest benefit estimate, with no published method: about a million hours a year if all police forces in England and Wales used AI redaction and new processes (older source).
  • The UK transparency records on AI in justice, tax administration and at the Disclosure and Barring Service are self-descriptions by the agencies without evaluation results.
  • The US Department of Defense’s AI office had filled only 3 of 10 responsible-AI posts in February 2026. The Swiss federal government lacks an overview of AI projects.
  • In the eGovernment MONITOR, 23% of participants in Germany said they had used AI chatbots for administrative services. The question does not separate types of chatbot.

2.0Starting point: many projects, gaps in governance

Audits and inventories from Switzerland, the US and France describe how administrations record, oversee and assess their AI projects. They show gaps in overview, staffing and ethical review. The French findings and the GAO audit of AI acquisitions date from before the reporting period.

Figure 1

AI use cases at France Travail

As of April 2025, according to France's court of audit (older context study)

Older context study from before the reporting period. In the court's assessment, most of the use cases fall into the high-risk category of the EU AI Act.Evidence: observed · Sources 19

In its report of 2 October 2026 the Swiss Federal Council states that the investment volume for AI and process automation in the federal administration cannot be quantified because there is no overall view of all AI and RPA projects. More than 130 reported processes yielded more than 70 use cases, 39 of them with mostly medium to high automation potential.5A1

According to CDAO officials cited in the GAO report, the US Department of Defense’s central AI office (CDAO) had filled only three of ten authorised responsible-AI positions in February 2026; according to CDAO documentation, 29 of the 64 lines of effort in its own implementation pathway were incomplete in April 2026.4A2

The investment volume for AI and process automation in the Swiss federal administration cannot be quantified because there is no overall view of projects.

Older context study: France’s court of audit counted 87 AI use cases at the employment service France Travail up to April 2025, 27 of them deployed at scale and 17 abandoned; a formalised ethical analysis had been started for only 18, and in the court’s assessment most of the use cases fall into the high-risk category of the EU AI Act.19A3

Further findings

  • Older context study: France’s digital directorate DINUM counted over 140 AI products in central government, 0 to 74 projects per ministry.12A4
  • Older context study: across 13 AI acquisitions reviewed, four US federal agencies were not yet systematically collecting lessons learned.20A5

3.0Justice and home affairs: no measured time savings reported

UK justice agencies and the Disclosure and Barring Service (DBS) use or are trialling AI for transcription and summarisation. The Home Office estimates the benefit of AI redaction tools together with more efficient processes if all police forces in England and Wales adopted them (factsheet of June 2026, updated in October, older context source). None of these sources reports measured time savings.

For Justice Transcribe, the AI transcription tool used by the probation service in England and Wales, the agency HMPPS estimates about 266,667 staff hours saved across more than 1.6 million summarised meetings; according to the NAO, no measured time data exist.1A6

In the Justice Transcribe pilot, internal quality assurance found inaccuracies in 20% of the final case records examined, mainly misspelt names and agencies; according to the NAO there is still no mechanism to check transcript accuracy.1A7

For Justice Transcribe the agency estimates staff hours saved, but according to the NAO no measured time data exist.

In September 2026 the UK transparency register published records for two AI tools in the justice system: BenchNotes transcribes orally dictated decisions of up to 430 judges of the First-tier Tribunal Immigration and Asylum Chamber, and Beam Notes produces transcripts and summaries of video evidence for the Crown Prosecution Service in about 11,000 cases a year. Neither record reports measured time savings; for Beam Notes only a vendor-reported transcription accuracy of 98% is given.16,17A8

Older context study: in its factsheet on police AI, first published on 9 June 2026 and updated on 22 October 2026, the UK Home Office estimates that AI-enabled audio-visual redaction tools together with more efficient processes would save around a million hours per year if adopted by all forces in England and Wales; a public registry of the AI in use is to start as a pilot by the end of 2026.8A9

Further findings

  • The DBS record for AI summaries of barring referrals mentions comparison tests without results on time or accuracy.7A10

4.0Welfare and tax administration: models in casework

In the UK and the US, AI models support casework in welfare and tax agencies: on child maintenance, tax debts and the selection of cases for examination, along with pilots of generative AI. On effects there are mostly only agency estimates or no performance values. Full evaluations are still missing. At the IRS it is unknown what share of examinations was selected with AI.

Figure 2

Rollout and review in welfare and tax administration

Selected dates at the UK DWP and the US tax agency IRS, 2024 to 2027

  1. 2024Child maintenance model in routine useSince March 2024 a model predicting payment breakdowns has supported caseworkers in the DWP Child Maintenance Service.
  2. 2025GAIL rollout beginsAfter a proof of concept in 2024, DWP rolls out its tool for producing training material in phases from March 2025; there is no long-term evaluation so far.
  3. Oct 2026TIGTA report on the IRSIt cites an audit report of March 2026: in the large partnership program, 92% of closed examinations ended with no change (preliminary). How many were AI-selected is not stated.
  4. 2027Data extension planned for late 2027DWP plans the full evaluation of its child maintenance model only after further data are added in late 2027.
Information from NAO and TIGTA audit reports. The timeline shows rollouts and review stages, not measured effects. The 2027 date is a DWP plan.Evidence: model estimate, observed · Sources 2, 6

The UK Department for Work and Pensions (DWP) estimates that interventions prompted by its machine-learning model kept an additional 715 child maintenance cases per month from falling into arrears; a full evaluation is planned only after further data are added in late 2027.2A11

Each month the UK tax authority HMRC uses the machine-learning model Business Propensity to Pay to calculate scores for about 4 million customers that are meant to predict how businesses will engage with the recovery of tax debts; the transparency record names evaluation metrics but no performance values, and those affected cannot appeal the score because it is formally not a decision.15A12

In the Large Partnership Compliance Program of the US tax agency IRS, which according to the oversight body TIGTA has begun using AI to select large partnerships for examination, 92% of closed examinations resulted in no change; TIGTA reads these preliminary results as a sign that the selection models need further refinement. TIGTA does not say what share of these examinations was selected with AI.6A13

Further findings

  • Only half of the AI projects in DWP’s Lighthouse programme got past the pilot, and the training tool GAIL lacks a long-term evaluation.2A14

5.0Health: process effects, rarely randomised

The month’s health studies come from Tanzania, from a hospital in China and from UK general practitioners. Only the Tanzanian study is randomised; the other two rest on a before-and-after comparison and on self-reports.

Health service, not administration: in a randomised trial in Tanzania, 35.9% of participants needing referral attended the eye clinic within eight weeks after AI-supported retinal screening with an immediate result and face-to-face counselling, compared with 18.2% under the standard pathway.3A15

In a large Chinese hospital, after an AI and robotic process automation financial system was introduced, daily data-collection time fell from 132.5 to 18.6 minutes and the data-entry error rate from 3.5% to 0.6%.10A16

After AI screening with counselling, 35.9% of participants needing referral attended the eye clinic within eight weeks, against 18.2% on the standard pathway.

Further findings

  • Per the GPAI-UK 2026 survey, 52% of 1,004 UK GPs use generative AI, and 70% of these users report a reduced work burden.9A17

6.0Citizens and staff: use, barriers, expectations

This chapter brings together surveys: on AI chatbots for digital administrative services in Germany, Austria and Switzerland, on acceptance of government AI in Germany and the Netherlands, on barriers in two Finnish public services and on AI use at Brazil’s federal court of audit. They show use, expectations and perceived barriers, not measured effects.

Figure 3

Use of AI chatbots for digital administrative services

Share of participants who have already used AI chatbots as support, eGovernment MONITOR 2026

One survey in three countries: internet users aged 16+ in private households, self-reports (Germany n = 8,091, Austria n = 1,002, Switzerland n = 1,005). The question does not distinguish between government and general-purpose chatbots. No uncertainty reported in the source.Evidence: observed · Sources 11

In the eGovernment MONITOR 2026, 23% of participants in Germany said they had already used AI chatbots as support for digital administrative services (Austria 32%, Switzerland 34%). Regarding AI in public administration, 35% in Germany expect more and 18% less equal treatment, while 38% fear that exceptional cases will be considered less well.11A18

In two Finnish surveys of public-service staff with no prior AI use, 84% in a regional wellbeing services organisation named limited knowledge of applications and 83% inadequate training as at least moderate barriers, while 76% in municipal technical services named doubts about the accuracy and reliability of generative AI.14A19

In an internal survey at Brazil’s federal court of audit TCU, 88.1% of participants said they use AI at work and 45.1% daily; 64.5% perceived time savings of more than 30% on their tasks, and 77.8% said they always fully review AI output before using it in official documents.18A20

Further findings

  • Survey in Germany and the Netherlands: acceptance of government AI was associated with whether participants presumed efficiency motives.13A21

7.0Evidence

Not every statement in this report is equally well supported. The table shows the key statements with their evidence type, AI relationship and sources.

StatementEvidenceAI linkSources
Swiss federal administration: AI investment cannot be quantified, no overall viewobservednot stated5
266,667 h saved so far by the AI note-taking tool Justice Transcribe, as estimated by the agency HMPPS, with no measured datamodel estimatenot stated1
20% of examined final case records in the pilot had inaccuracies, mainly in names, with no comparison rateobservednot stated1
Estimate: AI redaction plus more efficient processes, about a million hours a yearmodel estimatenot stated8
715 cases a month were additionally kept from maintenance arrears by model-prompted action, a DWP estimate with no control groupmodel estimatenot stated2
92% of closed examinations in the IRS program with AI selection ended in no change (preliminary), share of AI cases unknownobservednot stated6
AI retinal screening in Tanzania: 35.9% instead of 18.2% attended the clinicobservednot stated3
23% in Germany had used AI chatbots for digital administrative servicesobservednot stated11

The 21 reviewed statements from 20 original sources fall into chapters 2.0 (5), 3.0 (5), 4.0 (4), 5.0 (3) and 6.0 (4), of which 18 observed, 3 model estimate. “Not stated”: the edition does not assess the AI relationship of these statements separately. The scope and limitations of every statement are listed in the appendix (A1 to A21).

Original sources checked on October 26, 2026.

8.0What public administrations can do now

  1. Measure effects rather than assume them: record time savings and quality with measured data and a comparison group. The largest benefit figures in this report are estimates, such as the assumed saving of 10 minutes per meeting for Justice Transcribe.
  2. Check accuracy regularly: set up sample checks with a comparison value for AI-generated records and summaries. According to the NAO, Justice Transcribe still has no mechanism to check transcript accuracy.
  3. Create an overview: record AI projects centrally with costs and results and publish performance values. In the Swiss federal administration the investment volume cannot be quantified for lack of an overall view, and the HMRC record gives no performance values.
  4. Train staff and set review rules: in Finland, surveyed non-users named knowledge gaps, inadequate training and doubts about accuracy as barriers. At the TCU, 77.8% of survey participants say they always fully review AI output before official use.

9.0Method and limitations

Reporting period: October 1, 2026 up to the data cutoff of October 26, 2026 at 1:00 PM GMT+1 (Berlin time). This topic report is published before the end of the month; sources published after the cutoff follow in the next edition of the series. Compiled: October 27, 2026 at 12:23 AM GMT+1. This is the first baseline. The sources come from different years; their publication dates are not events in the reporting month. No change from an earlier edition was measured.

The report uses only reviewed public original sources. Customer conversations, private dossiers and personal customer assessments are not inputs. Observed, model estimate, scenario and open describe the evidence for each finding. Source values retain their original units and baselines; they are not combined into an overall score.

Editorial approval: qasb editorial team · Approved: October 27, 2026 at 12:25 AM GMT+1. The editorial presentation arranges the reviewed statements and contains no figure that is not in them. Presentation approved by the qasb editorial team · October 27, 2026 at 2:27 PM GMT+1.

Limits

  • This topic report covers one topic area: “Public systems”. It is not an exhaustive survey of all studies or incidents.
  • A narrow study finding is not a forecast for a particular customer, occupation or security system.
  • No qualified finding does not prove no effect; source-access gaps are disclosed separately.
  • This report contains no illustrative measurement series or invented confidence rating.

10.0Sources

  1. 1Productivity improvement: The Justice Transcribe AI-enabled transcription tool (Case study, HC 615)National Audit Office (UK), Report by the Comptroller and Auditor GeneralOctober 14, 2026www.nao.org.uk/insights/productivity-improvement-the-justice-transcribe-ai-enabled-transcription-tool/
  2. 2DWP’s approach to innovation (HC 616)National Audit Office (UK), Report by the Comptroller and Auditor GeneralOctober 21, 2026www.nao.org.uk/reports/dwps-approach-to-innovation/
  3. 3Artificial intelligence-supported diabetic retinopathy screening in Tanzania: a randomised controlled trialCleland et al., The Lancet Global HealthOctober 12, 2026doi.org/10.1016/j.langlo.2026.104062
  4. 4Artificial Intelligence: Actions and Opportunities to Implement Responsible Use in DOD (GAO-27-107837)U.S. Government Accountability OfficeOctober 14, 2026www.gao.gov/products/gao-27-107837
  5. 5Effizienzsteigerung von Verwaltungsprozessen durch Prozessautomatisierung und künstliche Intelligenz. Bericht des Bundesrates in Erfüllung des Postulates 24.3582 Silberschmidt vom 12. Juni 2024Schweizerischer BundesratOctober 2, 2026cms.news.admin.ch/fileservice/sdweb-docs-prod-nsbcch-files/files/2026/10/02/3336b959-b060-41fd-bf86-df4ca8ffd924.pdf
  6. 6Management and Performance Challenges Facing the Internal Revenue Service for Fiscal Year 2027Treasury Inspector General for Tax Administration (TIGTA), USAOctober 26, 2026www.tigta.gov/reports/top-management-challenges/management-performance-challenges-facing-internal-revenue-service-fiscal-year-2027
  7. 7Disclosure and Barring Service Barring Referral AI Summary Tool (Algorithmic Transparency Record)Disclosure and Barring Service / Home Office, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK)October 14, 2026www.gov.uk/algorithmic-transparency-records/disclosure-and-barring-service-barring-referral-ai-summary-tool
  8. 8Police use of artificial intelligence (AI): factsheetHome Office (Vereinigtes Königreich)June 9, 2026www.gov.uk/government/publications/police-use-of-artificial-intelligence-ai-factsheet/police-use-of-artificial-intelligence-ai-factsheet-accessible
  9. 9Generative artificial intelligence adoption and the future of the primary care profession: an online survey of UK general practitioners’ experiences and opinionsBMJ Digital Health & AI (BMJ Publishing Group)October 1, 2026doi.org/10.1136/bmjdh-2026-000253
  10. 10Effectiveness of a hospital intelligent financial system based on AI and RPAZhao, Zhang, Pan, Frontiers in Public HealthOctober 15, 2026doi.org/10.3389/fpubh.2026.1945091
  11. 11eGovernment MONITOR 2026Initiative D21 e. V. und Technische Universität München (Durchführung: Kantar)September 29, 2026initiatived21.de/uploads/03_Studien-Publikationen/eGovernment-MONITOR/2026/D21-eGovMon2026.pdf
  12. 12The roll-out of artificial intelligence in public administrations: comparative analysis, challenges and success factorsInspection générale des affaires sociales (IGAS), Inspection générale des finances (IGF), Inspection générale de l’administration (IGA), FrankreichJuly 2, 2026igas.gouv.fr/roll-out-artificial-intelligence-public-administrations-comparative-analysis-challenges-and-success-factors
  13. 13Why Government Turns to AI – What Citizens Believe and How it Matters. Survey Evidence from Germany and the NetherlandsPascal D. König, The American Review of Public Administration (SAGE)September 2, 2026doi.org/10.1177/02750740261481775
  14. 14Why generative AI adoption stalls across public service domains: an AMO comparison of Finnish wellbeing and municipal technical servicesPulkkinen, Rantamäki, Laukka, Transforming Government: People, Process and Policy (Emerald)September 11, 2026doi.org/10.1108/TG-05-2026-0261
  15. 15HMRC: Business Propensity to Pay (Algorithmic Transparency Record)HM Revenue & Customs, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK)September 9, 2026www.gov.uk/algorithmic-transparency-records/hmrc-business-propensity-to-pay
  16. 16HM Courts & Tribunals Service: BenchNotes (Algorithmic Transparency Record)Ministry of Justice / HM Courts & Tribunals Service, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK)September 9, 2026www.gov.uk/algorithmic-transparency-records/hm-courts-and-tribunals-service-benchnotes
  17. 17The Crown Prosecution Service: Beam Notes (Algorithmic Transparency Record)Attorney General’s Office / Crown Prosecution Service, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK)September 9, 2026www.gov.uk/algorithmic-transparency-records/the-crown-prosecution-service-beam-notes
  18. 18Pesquisa sobre Uso de IA Generativa no TCU (Infográfico)Tribunal de Contas da União (TCU), BrasilienSeptember 18, 2026portal.tcu.gov.br/uploads/noticias/pdf/2026/09/18/Infografico_Pesquisa-sobre-Uso-de-IA-Generativa-no-TCU_v1_5_(correcao)(1)_1_1.pdf
  19. 19France Travail et l’intelligence artificielle (Observations définitives)Cour des comptes (Frankreich)January 8, 2026www.ccomptes.fr/fr/publications/france-travail-et-lintelligence-artificielle
  20. 20Artificial Intelligence Acquisitions: Agencies Should Collect and Apply Lessons Learned to Improve Future Procurements (GAO-26-107859)U.S. Government Accountability OfficeApril 13, 2026www.gao.gov/products/gao-26-107859
Authors, reuse basis and rights records
  1. 1National Audit Office (UK), Report by the Comptroller and Auditor General · Reuse basis: © National Audit Office 2026; reproduction permitted for non-commercial purposes only, with acknowledgement; our own factual paraphrase and link only, no figures · Rights and license record
  2. 2National Audit Office (UK), Report by the Comptroller and Auditor General · Reuse basis: © National Audit Office 2026; reproduction permitted for non-commercial purposes only, with acknowledgement; our own factual paraphrase and link only, no figures · Rights and license record
  3. 3Cleland et al., The Lancet Global Health · Reuse basis: Inconsistent licence information: the publisher’s Crossref record gives CC BY 4.0 for the version of record, the PubMed record says “All rights reserved”; handled cautiously: our own factual paraphrase and link only, no figures or tables · Rights and license record
  4. 4U.S. Government Accountability Office · Reuse basis: Work of the U.S. government (public domain in the United States), may include third-party copyrighted material; our own paraphrase and link, no figures · Rights and license record
  5. 5Schweizerischer Bundesrat · Reuse basis: Copyright held by the Swiss federal authorities; per the admin.ch legal notice, reproduction only with prior written consent; our own factual paraphrase and link only, no text reproduction or figures · Rights and license record
  6. 6Treasury Inspector General for Tax Administration (TIGTA), USA · Reuse basis: Work of the U.S. government (not protected by copyright in the United States under 17 U.S.C. § 105); our own paraphrase and link, no figures · Rights and license record
  7. 7Disclosure and Barring Service / Home Office, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK) · Reuse basis: GOV.UK content under the Open Government Licence v3.0 unless otherwise stated; our own paraphrase with attribution and link · Rights and license record
  8. 8Home Office (Vereinigtes Königreich) · Reuse basis: © Crown copyright; GOV.UK content under the Open Government Licence v3.0; our own paraphrase with attribution and link (first published 9 Jun 2026; updated version of 22 Oct 2026 cited) · Rights and license record
  9. 9BMJ Digital Health & AI (BMJ Publishing Group) · Reuse basis: CC BY 4.0 (per the publisher’s licence entry in the Crossref record); our own paraphrase and link only, no figures · Rights and license record
  10. 10Zhao, Zhang, Pan, Frontiers in Public Health · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures · Rights and license record
  11. 11Initiative D21 e. V. und Technische Universität München (Durchführung: Kantar) · Reuse basis: Copyright held by the publishers; per the initiatived21.de imprint, texts and graphics may not be reproduced or used in other publications without express consent, links are not restricted; our own factual paraphrase and link only, no text reproduction or figures · Rights and license record
  12. 12Inspection générale des affaires sociales (IGAS), Inspection générale des finances (IGF), Inspection générale de l’administration (IGA), Frankreich · Reuse basis: Licence Ouverte / Etalab 2.0 (per the IGAS website, third-party rights excepted); our own paraphrase with attribution and link (the English version of 9 Sep 2026 is linked; French original of 2 Jul 2026) · Rights and license record
  13. 13Pascal D. König, The American Review of Public Administration (SAGE) · Reuse basis: CC BY 4.0 (per the publisher’s licence entry in the Crossref record); our own paraphrase and link only · Rights and license record
  14. 14Pulkkinen, Rantamäki, Laukka, Transforming Government: People, Process and Policy (Emerald) · Reuse basis: CC BY 4.0; our own paraphrase and link only, no figures or tables · Rights and license record
  15. 15HM Revenue & Customs, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK) · Reuse basis: GOV.UK content under the Open Government Licence v3.0 unless otherwise stated; our own paraphrase with attribution and link · Rights and license record
  16. 16Ministry of Justice / HM Courts & Tribunals Service, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK) · Reuse basis: GOV.UK content under the Open Government Licence v3.0 unless otherwise stated; our own paraphrase with attribution and link · Rights and license record
  17. 17Attorney General’s Office / Crown Prosecution Service, veröffentlicht im Algorithmic Transparency Recording Standard (GOV.UK) · Reuse basis: GOV.UK content under the Open Government Licence v3.0 unless otherwise stated; our own paraphrase with attribution and link · Rights and license record
  18. 18Tribunal de Contas da União (TCU), Brasilien · Reuse basis: All rights reserved; per the TCU terms of use, republication of public content is allowed with source, title and author, commercial use only with permission; our own factual paraphrase and link only · Rights and license record
  19. 19Cour des comptes (Frankreich) · Reuse basis: Per the Cour des comptes legal notice, reproduction is allowed with explicit attribution to www.ccomptes.fr and without distorting changes; our own paraphrase and link, no figures · Rights and license record
  20. 20U.S. Government Accountability Office · Reuse basis: Work of the U.S. government (public domain in the United States), may include third-party copyrighted material; our own paraphrase and link, no figures · Rights and license record

How to cite

qasb Research (2026). AI in the public sector: benefits mostly estimated rather than measured. Topic report 10/2026, data cutoff October 26, 2026. https://www.qasb.ai/en/reports/2026-10-oeffentlicher-sektor

Appendix

For every statement: evidence type, change status, scope and limitations, and the observation period, source value, comparison and uncertainty where the source gives them. The texts are the approved entries of the edition, as is its summary.

Scope and limitations of all 21 statements (A1–A21)

Edition summary

In October 2026 the strongest evidence on AI in the public sector comes from audit and oversight reports (NAO, GAO, TIGTA) and official transparency records: they show wide deployment in justice, welfare and tax administration whose benefits are mostly estimated rather than measured; at the IRS, in a program that uses AI for case selection, preliminarily 92% of closed examinations ended with no change, without the share of AI-selected cases being known. The month’s health studies report process effects mostly from self-reports and before-and-after comparisons; only the retinal screening trial in Tanzania is randomised. For Germany and the EU the evidence consists mainly of surveys and inventories (eGovernment MONITOR, Finland, plus Switzerland and, as an older context source, France); outcome measurements, court rulings and dated evidence on market surveillance under the AI Act are missing from the material checked.

  1. A1Back to chapter 2.0

    Evidence: observed · baseline · Source 5

    Scope
    Swiss federal administration; cross-departmental survey in the fourth quarter of 2025 and ten workshops on reference processes in the first quarter of 2026; report in response to postulate 24.3582.
    Limitations
    The automation potential rests on initial assessments by workshop participants, supplemented by the project team; there are no measured effects. Reporting AI projects to the competence network is voluntary. Because funds are tight, not all use cases will be pursued; each department is to take part in at least three. Switzerland, not the EU.
    Observation period
    Survey in Q4 2025; workshops in Q1 2026; report of 2 Oct 2026
    Source value
    39 use cases with mostly medium to high automation potential
    Baseline
    more than 70 identified use cases from more than 130 reported processes
    Comparison
    potential assessed by value added, feasibility and risk control
    Uncertainty
    No uncertainty reported in the source; initial assessment by workshop participants

  2. A2Back to chapter 2.0

    Evidence: observed · baseline · Source 4

    Scope
    Chief Digital and AI Office (CDAO) of the US Department of Defense and ten surveyed DOD components, each with one selected AI use case; GAO audit September 2024 to October 2026.
    Limitations
    Governance audit: it measures oversight capacity, not effects or error rates of AI applications. Staffing and implementation figures rest on CDAO documents and officials’ statements. Defence context, not readily transferable to civilian administration.
    Observation period
    As of February 2026 (positions) and April 2026 (lines of effort); audit September 2024–October 2026
    Source value
    3 of 10 authorised positions filled
    Baseline
    10 positions under a 2023 resourcing memorandum
    Comparison
    staffing in February 2026
    Uncertainty
    No uncertainty reported in the source; figure from CDAO officials

  3. A3Back to chapter 2.0

    Evidence: observed · baseline · Source 19

    Scope
    France Travail (public employment service, the largest state operator by staff); Cour des comptes audit of the deployment, effects and governance of AI; investment of about €93 million between 2017 and 2024.
    Limitations
    Status April 2025, published January 2026, outside the reporting period. The count includes variants of the same tool. According to the court, effects on work and working conditions are not adequately tracked; the more recent figure (68 use cases in early 2026) comes from the inspectorates’ report.
    Observation period
    2017 to April 2025; report of 8 Jan 2026
    Source value
    18 use cases with a formalised ethical analysis begun
    Baseline
    87 identified use cases (as of April 2025)
    Comparison
    use cases with an initial ethical analysis
    Uncertainty
    No uncertainty reported in the source

  4. A4Back to chapter 2.0

    Evidence: observed · baseline · Source 12

    Scope
    Central government, social security and the public employment service in France; joint report of the three general inspectorates (cover dated April 2026, French original published 2 Jul 2026, the English version linked here 9 Sep 2026); figures from the mission’s data collection and agencies’ questionnaire answers.
    Limitations
    First published in July 2026, before the reporting period. An inventory without outcome measurement; counting units differ (products, projects, use cases). Benefit examples in the report are feedback without a comparison group; for one municipal chatbot the summary and the main report give different baseline values. Costs are said to be hard to estimate and often exceed plans.
    Observation period
    Data collected up to early 2026; French original of 2 Jul 2026, English version of 9 Sep 2026

  5. A5Back to chapter 2.0

    Evidence: observed · baseline · Source 20

    Scope
    Four US federal agencies (DOD, DHS, GSA, VA) with 13 selected AI acquisitions; on challenges, GAO interviewed officials from five agencies (adding the Department of Commerce); review of OMB guidance on AI acquisition.
    Limitations
    Published April 2026, outside the reporting period; small, selected number of cases; nothing on costs or outcomes of the acquired systems. US context, no framework contracts in the EU.
    Observation period
    Audit up to April 2026

  6. A6Back to chapter 3.0

    Evidence: model estimate · baseline · Source 1

    Scope
    About 11,000 sentence-management staff in the Probation Service (England and Wales), national rollout February–March 2026; partial automation of meeting records (transcript and summary, reviewed by the officer). Case study by the National Audit Office (NAO).
    Limitations
    Agency modelling based on an assumed saving of 10 minutes per meeting; no measured time data and no comparison group. The NAO did not independently verify the reported results. Nothing on how freed-up time is used or how record quality develops.
    Observation period
    7 Oct 2025–14 Sep 2026 (usage figures); forecast of about 450,000 hours per year
    Source value
    266,667 staff hours (agency estimate, cumulative)
    Baseline
    assumption: 10 minutes saved per meeting compared with manual note-taking
    Comparison
    more than 1,600,000 meetings summarised with the tool
    Uncertainty
    No uncertainty reported in the source; no measured time data available

  7. A7Back to chapter 3.0

    Evidence: observed · baseline · Source 1

    Scope
    Pilot with 40 probation officers in Kent, Surrey, Sussex and Wales (May–July 2025; 30 completed the final survey); checks by HMPPS’s Performance Assurance and Risk group; state of quality assurance after the national rollout.
    Limitations
    The number of records examined and a comparison rate for manually written records are not reported. The 73% time saving on write-ups reported in the pilot is self-reported. Not independently verified by the NAO.
    Observation period
    Pilot May–July 2025; quality-assurance status 2026
    Source value
    20 percent of examined final case records with inaccuracies
    Baseline
    No comparison rate for records without AI support reported
    Comparison
    pilot records with an AI-generated summary after officer review
    Uncertainty
    No uncertainty reported in the source; sample size not stated

  8. A8Back to chapter 3.0

    Evidence: observed · baseline · Sources 16, 17

    Scope
    HM Courts & Tribunals Service (phase “Production” per the record) and Crown Prosecution Service (private beta across all 14 CPS areas in England and Wales); speech recognition via Azure and ElevenLabs respectively, summarisation via OpenAI models for Beam Notes; ATRS records.
    Limitations
    Self-descriptions by the agencies without evaluation results. The BenchNotes record is internally inconsistent: phase “Production” but “pilot” in the risk section, and a human rights impact assessment is listed both as completed and as an open action. Judges and prosecutors must check all outputs.
    Observation period
    Records of 9 Sep 2026

  9. A9Back to chapter 3.0

    Evidence: model estimate · baseline · Source 8

    Scope
    Police forces in England and Wales; Home Office; new National Centre for AI in Policing (PoliceAI) hosted by the College of Policing; funding from the Police Reform White Paper (£115 million over three years). First published 9 Jun 2026; the version of 22 Oct 2026 does not mark which statements are new, so it is treated as an older context source.
    Limitations
    Departmental modelling without a published method; the estimate covers tools and process changes together. Nothing on the number of AI tools in use, error rates or outcomes. The change note of 22 Oct 2026 only says the factsheet was updated; whether the estimate and the registry timetable were already in the first version could not be checked. A policy document, not an evaluation.
    Observation period
    First published 9 Jun 2026, updated version of 22 Oct 2026; funding since 2025-26
    Source value
    1,000,000 hours per year (departmental estimate if adopted by all forces)
    Baseline
    audio-visual redaction without AI tools and without the funded process changes
    Comparison
    AI-enabled redaction tools and more efficient processes used by all forces in England and Wales
    Uncertainty
    No uncertainty reported in the source; worded as “around a million”

  10. A10Back to chapter 3.0

    Evidence: observed · baseline · Source 7

    Scope
    DBS Barring Referral Service (UK according to the record); triage team of about 16 caseworkers handling about 1,200 cases per month; phase Public Beta; models via AWS Bedrock; record in the UK Algorithmic Transparency Recording Standard (ATRS).
    Limitations
    Self-description by the agency without independent review; the evaluation covered about 140 production cases in private beta, and the metrics are named but not quantified. The statement that summaries support faster case orientation is not quantified.
    Observation period
    Tests in alpha and private beta (period not stated); record of 14 Oct 2026

  11. A11Back to chapter 4.0

    Evidence: model estimate · baseline · Source 2

    Scope
    DWP Child Maintenance Service; model predicting payment breakdowns (70% accuracy in testing) as AI support for caseworkers, in routine operation since March 2024; case example in the NAO report “DWP’s approach to innovation”.
    Limitations
    Departmental estimate from a test-and-learn approach and staff feedback, without a control group; the time gain equivalent to 49 full-time staff is also an estimate. The NAO relied on the department’s documents, data and interviews; it does not report its own check of model performance.
    Observation period
    Routine operation since March 2024; NAO evidence collected March–September 2026
    Source value
    715 additional cases per month not moving into arrears (departmental estimate)
    Baseline
    about 8,000 cases per month with an unplanned payment breakdown (DWP estimate)
    Comparison
    cases with model-prompted early intervention
    Uncertainty
    No uncertainty reported in the source

  12. A12Back to chapter 4.0

    Evidence: observed · baseline · Source 15

    Scope
    HM Revenue & Customs, Debt Management; gradient-boosted model for VAT, PAYE and corporation tax debts; the score shapes the frequency and content of letters; record in the UK transparency register ATRS of 9 Sep 2026.
    Limitations
    Self-description by the agency without independent review; nothing on effects on payments or on particular groups. Staff see the score in casework, while monitoring is monthly and only at aggregate level.
    Observation period
    Monthly operation; record of 9 Sep 2026

  13. A13Back to chapter 4.0

    Evidence: observed · baseline · Source 6

    Scope
    US tax agency IRS, Large Partnership Compliance Program; TIGTA report on management challenges for fiscal year 2027. According to TIGTA, the IRS reported 175 projects involving AI in November 2025 and also uses AI to offset workforce reductions.
    Limitations
    Preliminary results; neither the share of AI-selected cases among the closed examinations nor a comparison rate without AI selection is given, so no link between AI selection and outcomes is shown. The figure comes from an earlier TIGTA audit report (March 2026) that was not opened here. Nothing on additional tax revenue.
    Observation period
    Closed examinations up to the reporting date (period not stated); AI project count as of November 2025; report of 26 Oct 2026
    Source value
    92 percent of closed examinations with no change
    Baseline
    No comparison rate for examinations without AI-supported selection reported
    Comparison
    closed examinations in the Large Partnership Compliance Program
    Uncertainty
    No uncertainty reported in the source; described by TIGTA as preliminary

  14. A14Back to chapter 4.0

    Evidence: observed · baseline · Source 2

    Scope
    DWP’s Lighthouse Programme for generative AI (now part of the AI and Innovation Directorate); case example GAIL (Generative AI in Learning): proof of concept funded with £428,000, followed by a further £10 million for digital and AI activities including GAIL; NAO report “DWP’s approach to innovation”.
    Limitations
    The NAO gives neither the number of Lighthouse projects nor the reasons for not taking projects forward. The statement that tasks taking days or weeks could be done in minutes comes from DWP’s early testing with no published method. Nothing on quality or costs in routine operation.
    Observation period
    Proof of concept January–May 2024; rollout from March 2025, extended March 2026; NAO evidence March–September 2026

  15. A15Back to chapter 5.0

    Evidence: observed · baseline · Source 3

    Scope
    2,378 people with diabetes at diabetes/NCD clinics of 24 hospitals in the Kilimanjaro and Arusha regions; 904 with confirmed referable retinopathy; individually randomised; comparator: deferred human image grading with telephone counselling.
    Limitations
    Health-care study, not an administrative process. The intervention bundles the AI result with immediate face-to-face counselling; the AI’s share of the effect is not isolated. Clinic attendance was measured, not treatment outcomes. Software vendor not named in the abstract; only the abstract was checked.
    Observation period
    8 Mar 2023–31 Aug 2024; eight-week follow-up
    Source value
    17.7 percentage points higher attendance within eight weeks
    Baseline
    18.2% (85 of 467) under the standard pathway
    Comparison
    35.9% (157 of 437) under the AI-supported pathway
    Uncertainty
    95% CI 12.0–23.4 percentage points; odds ratio 2.52 (1.86–3.42)

  16. A16Back to chapter 5.0

    Evidence: observed · baseline · Source 10

    Scope
    One tertiary A general hospital in China; comparison of the six months before and after the June 2025 launch; finance and billing processes including risk prediction for health-insurance violations.
    Limitations
    Single centre, non-randomised, no control group (before-and-after comparison); the authors report no other system, staffing or policy changes between the periods but acknowledge that trends, seasonal differences and learning effects cannot be ruled out. The hospital’s ownership is not stated explicitly; staffing effects, costs and loss of skills were not studied.
    Observation period
    January to December 2025 (six months each before and after the June 2025 launch)
    Source value
    18.6 minutes of daily data-collection time after introduction
    Baseline
    132.5 minutes before introduction
    Comparison
    six months after introduction of the AI-RPA financial system
    Uncertainty
    P < 0.001 per the abstract; no confidence intervals reported

  17. A17Back to chapter 5.0

    Evidence: observed · baseline · Source 9

    Scope
    General practitioners in the UK (primary care); annual online convenience survey via Doctors.net.uk within a monthly omnibus; most frequently named tools ChatGPT, Heidi Health and Microsoft Copilot.
    Limitations
    Convenience sample, not representative; use and relief are self-reports, not measured time savings. Nothing on patient safety or documentation quality; the abstract does not distinguish practice types or ownership. Only the abstract deposited by the publisher with Crossref was checked (the publisher page blocks automated retrieval).
    Observation period
    2026 survey wave (exact period not stated in the abstract)
    Source value
    52 percent of GPs using generative AI in clinical practice (522 of 1,004)
    Baseline
    total sample of 1,004 GPs
    Comparison
    70% of users (363 of 522) report a reduced work burden
    Uncertainty
    No uncertainty reported in the abstract; convenience sample

  18. A18Back to chapter 6.0

    Evidence: observed · baseline · Source 11

    Scope
    Online survey (CAWI) by Initiative D21 and the Technical University of Munich, fielded by Kantar; internet users aged 16+ in private households; Germany n = 8,091 (about 500 per federal state), Austria n = 1,002, Switzerland n = 1,005.
    Limitations
    Internet users only; self-reports and expectations, no measurement of actual AI effects in administrative procedures. The question does not distinguish between government and general-purpose chatbots. The study is supported by numerous IT companies as partners.
    Observation period
    April and May 2026
    Source value
    23 percent of participants in Germany who have used AI chatbot support for digital administrative services
    Baseline
    Austria 32%, Switzerland 34%
    Comparison
    Germany (n = 8,091)
    Uncertainty
    No uncertainty reported in the source

  19. A19Back to chapter 6.0

    Evidence: observed · baseline · Source 14

    Scope
    Finland; one regional wellbeing services organisation (4,226 invited, 437 responses) and municipal land-use planning and building control (2,646 invited, 218 responses from 104 municipalities); analysis of 321 non-users; web surveys in September 2025.
    Limitations
    Low response rates (about 10% and 8%) and self-selection; non-users only; shares exclude “don’t know” answers; perceptions, no measurement of use or effects. The survey predates publication by a year.
    Observation period
    September 2025

  20. A20Back to chapter 6.0

    Evidence: observed · baseline · Source 18

    Scope
    Participants in an internal survey at Brazil’s federal court of audit (Tribunal de Contas da União) on the use of generative AI at work; the most used tool is the in-house ChatTCU (76.6%); TCU infographic on the survey.
    Limitations
    Self-reports; time savings are perceived, not measured. Number and composition of participants, recruitment and fieldwork period are not stated in the infographic; a bias towards users is possible.
    Observation period
    2026 (exact fieldwork period not given); published 18 Sep 2026
    Source value
    88.1 percent of participants using AI at work
    Baseline
    No earlier survey reported
    Comparison
    45.1% use AI daily
    Uncertainty
    No uncertainty reported in the source

  21. A21Back to chapter 6.0

    Evidence: observed · baseline · Source 13

    Scope
    Citizens in Germany and the Netherlands; four types of government AI use involving different citizen-state relationships; survey data, article in a peer-reviewed journal.
    Limitations
    Associations from a survey, no causal claim; the abstract does not say whether the four types of use are real or hypothetical cases. Only the abstract deposited by the publisher with Crossref was checked; the publisher page and full text block automated retrieval, so fieldwork period, sampling and effect sizes were not verified.
    Observation period
    Fieldwork period not stated in the abstract; published 2 Sep 2026

And in your case? Check it with sources.

Start an analysisFor organizations
All reports