How RAG is changing access to knowledge in governments around the world

An agent has to answer a simple question: which procedure applies today to this case? The answer exists. It is probably in a PDF, perhaps in two, most likely in a version that is no longer in force. The problem is not that the administration lacks information. It is that it has too much of it, spread across too many places, and nobody can say which one is authoritative.

First article of a series of five. This one draws the panorama and states the thesis; the four that follow open the hood.

1. A worldwide problem, not a model problem

Document dispersion is not a French peculiarity. In France, according to a 2017 survey by the Cour des comptes (the national audit court), only 16% of the State’s archives are kept at the Archives nationales; more than 60% are kept in the departmental archives ¹. The Archives nationales hold 395 linear kilometres, 135 terabytes and 14.9 million digitised documents available online ². These are not saturation figures: they are distribution figures. The document that answers the question exists, but you have to know in which silo, in which version and under which authority to look for it.

Elsewhere, the same structure appears in other forms. At KIRA, in Germany, analysing an IT project contract means cross-checking pieces spread over several documents ³. At Digdir, in Norway, the tool meant to extract measures from strategic documents first runs into the heterogeneity of formats ⁴. In Brazil, the public guide to RAG places revoked norms among the first risks: a document still in the base may have lost all legal value ⁵. Three countries, three trades, one finding: the corpus is the problem, long before the model.

How much does this searching cost? Nobody measures it. JuriScope, an investigative project of the French Ministry of Foreign Affairs, mentions 40 to 60% of the time of some projects spent on document research, with no protocol that would turn it into a general figure ⁶. The Cour des comptes, for its part, establishes that existing mechanisms allow neither a firm commitment on the productivity gains of digital projects nor a reliable measurement of them ⁷. This is an important precision: the cost of the problem is documented qualitatively, not quantitatively. Anyone who puts forward a percentage of lost time must say where it comes from.

What this article does not claim

It does not promise a demonstrated average gain: none is published. It does not promise automatic compliance: a RAG does not check that a norm is applicable. It does not promise sovereignty through hosting alone: an “on-premise” system can be less traceable than a cloud under contract. It does not promise infallibility: the systems that work are bounded. It gives no unit cost: no public case publishes one.

A note on method. The corpus of this article holds 23 RAG systems or projects spread over twelve countries, plus three European institutions (four projects). A source enters only if it explicitly names RAG, or unambiguously describes the “document retrieval → generation” cycle. Borderline cases (the Pair chatbot in Singapore, Jüri in Estonia, Hankintavälkky in Finland) are cited as counter-examples, never counted as RAG. All pages were consulted on 19 October 2026; the statuses are those of the sources, not a field update.

2. Four things not to confuse

The word “chatbot” now covers four different things, and the confusion is the first source of error in public projects.

Object	What it does	Example	What it does not do
Information chatbot	answers from public content or a decision tree	Ask Laura, El Paso, retired 8	does not cite a business corpus, does not abstain
General-purpose assistant	writes, summarises, free conversation	Pair, Singapore 9; L’Assistant, France 10	does not guarantee grounding in documents
Semantic search	finds passages close to a question	Jüri, Estonia, according to a secondary source 11	does not generate a written answer
Document RAG	searches passages, gives them to the model, cites the sources	Assist 12, China Maritime 13, KIRA 3	does not make the model infallible

The most economical demonstration comes from Singapore. GovTech runs two products that everyone calls “AI chatbots”: AIBots and Pair. AIBots is described as a retrieval-augmented generation system, on internal documents that can go up to the Restricted / Sensitive level ¹⁴. Pair is a generative assistant, with no document retrieval described ⁹. Both are useful; only one is a RAG. The difference is not cosmetic: it determines what you can demand of the system on citation, permissions and abstention.

In Finland, Hankintavälkky illustrates another boundary. The assistant of the public procurement portal is indeed in service, with about 700 conversations a month at the end of 2024 and favourable feedback ¹⁵. But no source describes its architecture as a RAG. Calling it “RAG” would be a convenience of language, not a fact. The same caution applies to Jüri: the figure of 83% of routine questions handled comes from a secondary source ¹¹.

This clarification is not an exercise in vocabulary. A chatbot that does not cite cannot be audited. A general-purpose assistant that does not abstain cannot be deployed on sensitive procedures. A RAG that does not filter rights is not a document tool: it is a leak waiting to happen. The requirements are not the same, and neither is the budget.

3. Three models of organisation

The panorama is not read as a maturity ladder, from the most advanced country to the least advanced. It is read according to three ways of organising the service. One country can practise all three.

1

Shared platform

An operator provides the tooling; agencies bring their corpora.

• AIBots · Singapore115 agenciesService
• North Carolinaagency widgetService
• Albert API · France70+ projectsService

2

Business product

One team, one corpus, one job.

• Assist · United Kingdom5 documentsService
• China Maritime5,488 question-answer pairsService
• Seoul Data Hubmunicipal dataService
• GSA Leasing Botinternal guideService
• KIRA · GermanyIT contractsPilot

3

Guide and capabilities

A document, not a system: it spreads practice, it does not run anything.

• GOV.UK AI Insightstwo guidesService
• Brazil · MGIpublic guideService
• Rwanda · Policy Navigatoraward-winning project, not deployedProject

The shared platform. A central operator provides the tooling, agencies bring their documents. AIBots, in Singapore, claims 40,000 users, 115 agencies, 12,000 bots and more than a million messages over the period August 2024 – February 2025 ¹⁴. North Carolina offers an information widget to State agencies, hosted and maintained by NCDIT ¹⁶. In France, Albert API exposes generation, RAG, vectorisation, reranking and OCR, with more than 70 public projects and more than 100,000 requests per week. These figures are the operator’s, not audited ¹⁷.

The business product. One team, one corpus, one trade. Assist, in the United Kingdom, helps public communicators write and review, over five explicitly maintained reference documents ¹². China Maritime answers regulatory questions for seafarers and companies, on a base of 5,488 question-answer pairs ¹³. Seoul Data Hub queries municipal public data in natural language ¹⁸. KIRA analyses contracts of large IT projects, locally, on behalf of the ASDN ³.

The guide and the capabilities. A document, not a system. GOV.UK AI Insights publishes two guides on RAG, one of them on agentic RAG ^{19, 20}. Brazil publishes an introductory guide to RAG in the public sector, which deals with revoked norms ⁵. Rwanda designed a Policy Navigator that won an award at the AI for Good 2026 summit, with no established entry into service ²¹.

Each model solves one problem and leaves another untouched. The platform solves availability, not the quality of the corpora poured into it. The business product solves precision on a narrow perimeter, not reuse by other services. The guide solves the spread of practice, not execution. This is why a country can have, at the same time, a national platform, a business product that works and a guide: these are not three stages, they are three answers to three different constraints.

The trap, when writing a panorama, is to rank countries. France has a platform (Albert API) and business experiments (Loiret, RAGaRenn) ^{17, 22}. The United Kingdom has a guide (GOV.UK) and a product (Assist). Singapore has a platform (AIBots) and a general-purpose assistant (Pair), which is not a RAG. A table by country would give a false impression of progress; a table by model gives a decision grid.

4. What services in production show

Six cases let us see what a public service in production actually publishes, and what it does not.

AIBots (Singapore). The system is in production at government scale, on the government network, with classified documents. What it publishes: adoption volumes. What it does not publish: any measure of quality. We do not know what rate of correct answers it reaches, nor how permissions go down to the passages ¹⁴. This is the paradox of the most cited case in the corpus: its success is organisational, not technically documented.

Assist (United Kingdom). This is the most complete case. The algorithmic transparency record describes a RAG combined with the GOV.UK Search API, with citations, over five reference documents ¹². The service claims about three hours saved per week and 98% declared usefulness among 8,500 civil servants in July 2026. The method is not published. Hosting relies on AWS Bedrock, with an “escrow” account that prevents the provider from accessing the data ¹². It is better documented than average, and it is still not an evaluation: no question set, no recall, no exact citation rate.

China Maritime. The Ministry of Transport publishes a rare figure: more than 90% announced accuracy, over 3,570 answers, since 25 June 2025 ¹³. The protocol is not published, the sample is not described, the abstention rate is absent. The system provides for double checking and a handover to a human. It is a real business service, with a figure, but a figure that allows no comparison.

Seoul Data Hub. 67,300 consultations, 13,800 users, about 80% satisfaction, according to a city press release of February 2025 ¹⁸. The survey method is not published. Satisfaction measures the experience, not accuracy. A user can be satisfied with a wrong answer if it is fluent.

Ask the EP Archives. The European Parliament makes more than 100,000 historical documents, from 1952 to 1994, searchable, with machine translation into 55 languages ²³. No metric is published. It is a remarkable case for its language coverage and the age of the corpus; it says nothing about the quality of the answers.

ECBot (ECB). The public chatbot combines predefined answers and a RAG layer for open questions; it displays three main sources, and the IBM Watson Discovery infrastructure is located in Germany and Romania ²⁴. It is an example of a “second layer” RAG: the RAG is not the product, it catches the questions the decision tree does not cover.

To which a useful counter-example is added: the American GSA declares a Leasing Desk Guide Bot deployed, a Pricing Policy Chatbot in pilot and LaborMatch IQ in pre-deployment (three statuses for three cases, none of which publishes a measured benefit) ²⁵. The analytical mirror of the federal inventory fills in none of the eight governance questions it tracks ²⁶.

The American inventories illustrate the gap at federal scale. The GAO measures spectacular adoption (32 use cases recorded in 2023, 282 in 2024) and, in the same report, the obstacles: technical and budget resources, updating usage rules, privacy ²⁷. NARA documents a natural-language search in production and an automatic labelling of about two million documents, with no RAG architecture described and no impact metric ²⁸. The Department of Labor declares two tools deployed without publishing a single figure ²⁹. An inventory says that a system exists; it does not say that it works.

The common thread of all these cases fits in one sentence: usage is proven, quality is not. We know how many people use them, how many questions are asked, how many documents are indexed. We almost never know whether the answer is right, complete and applicable.

5. What the pilots admit

Pilots are more instructive than successes, because they say where the project stops.

Minnesota. The nebulaONE platform was meant to let State agencies prototype chatbots and RAGs on data of Low and Moderate level. The project cost the TMF fund $64,000 and was closed on 11 March 2025, with no result published ³⁰. This is not a documented failure: it is a stop without a review. The difference matters to anyone looking for lessons learned.

KIRA (Germany). The prototype was presented in March 2025, with an MVP planned for the end of August and daily use envisaged in the fourth quarter of 2025. No confirmation followed ³. The case is nonetheless interesting: it is on-premise, on sensitive contracts, with citations. It shows that the confidentiality constraint can be handled, not that it has been at scale.

Rwanda. RISA’s Policy Navigator received an international distinction in 2026, for an announced perimeter of 50 agents and 24 institutions. Entry into service is not established ²¹. An award is not a deployment; it is a recognition of design.

Kenya. GovBot is a 2024–2026 pilot, with a detailed architecture: ChromaDB, LlamaIndex, PydanticAI, citations and confidence scores ³¹. It is one of the few cases where the architecture is public. No usage or quality metric is.

JRC (European Commission). Two demonstrators: crisis analysis, in December 2024, and nuclear safeguards, in September 2026 ^{32, 33}. Both are aimed at analysts; the second provides for human review. These are demonstrators, not services.

Japan and India. SEABIS, the assistant of Japan’s Digital Agency, was shown in a demonstration on 25 June 2025, with citations and page numbers, without deployment ³⁴. In India, the NFRA challenge is a call for proposals, not a system ³⁵.

Network Rail (United Kingdom). The case is technically accomplished and not yet in service. The lessons-learned chatbot rests on Azure AI Search and Azure OpenAI, with citations; the proof of concept took four to five weeks, the chatbot could be queried after two weeks, and nine weeks were spent consolidating security, infrastructure and formats ³⁶. No gain, no adoption, no accuracy is published. A figure does circulate: the 29,000 hours a year and the roughly three million pounds attributed to Network Rail concern another initiative, on the data of the Wales & Western region, and must not be attached to the chatbot ³⁷. This is the kind of confusion that turns a proof of concept into an industrial success in a presentation.

The Norwegian case deserves a place of its own. Digdir does not publish a success: it publishes five failures observed on 58 strategic documents: wrong page references, a perimeter not respected, answers produced even when the base does not contain the information ⁴. It is the only public self-evaluation of this kind in the corpus. For a project manager it is worth more than ten deployment announcements.

Two lessons emerge. First, the move to service is a phase of its own: a prototype that works does not become a service because someone decides it should. Second, stops are documented, causes are not. MILEVA was stopped on 31 January 2026 without publishing code, budget or usage statistics ⁶; Intelligence Entreprise, at the DGE, stopped in 2022 after trying a semantic engine on more than a million documents ⁶. These two stops say the same thing: the need was real, the industrialisation was not.

6. The four missing pieces of evidence

This is where the panorama stops being a catalogue. Of 23 systems or projects, four publish a quality or benefit figure: China Maritime, China NDA (a figure from a supplier file), Seoul and Assist ^{13, 38, 18, 12}. None publishes a reproducible protocol. None is evaluated by a third party.

What 23 public RAG systems and projects document

4 / 23publish a quality or benefit figure

1 / 23document how documents are classified

0 / 23publish a unit cost

Service deployed or in usePilotDemoProject documented not documented

Evidence matrix of 23 public RAG systems and projects. Four columns: status, citation documented (7 of 23), access rights documented (1 of 23) and evaluation published (4 of 23). A filled dot means the public source documents it; an empty dot means the source says nothing about it.

System	Status	Cites	Rights	Eval.
AIBotsSingapore	Service	citation not documented	rights documented	evaluation not published
AssistUnited Kingdom	Service	citation documented	rights not documented	evaluation published
Leasing Desk Guide BotGSA, United States	Service	citation not documented	rights not documented	evaluation not published
Pricing Policy ChatbotGSA, United States	Pilot	citation not documented	rights not documented	evaluation not published
LaborMatch IQGSA, United States	Project	citation not documented	rights not documented	evaluation not published
AI ChatbotNorth Carolina	Service	citation not documented	rights not documented	evaluation not published
nebulaONEMinnesota, United States	Pilot	citation not documented	rights not documented	evaluation not published
Ask the EP ArchivesEuropean Parliament	Service	citation not documented	rights not documented	evaluation not published
ECBotECB	Service	citation documented	rights not documented	evaluation not published
RAG crisis analysisJRC	Demo	citation not documented	rights not documented	evaluation not published
RAG nuclear safeguardsJRC	Demo	citation not documented	rights not documented	evaluation not published
KIRAGermany	Pilot	citation documented	rights not documented	evaluation not published
Digdir toolNorway	Pilot	citation not documented	rights not documented	evaluation not published
Seoul Data HubSouth Korea	Service	citation documented	rights not documented	evaluation published
China MaritimeChina	Service	citation documented	rights not documented	evaluation published
RAG annotationChina, NDA	Service	citation not documented	rights not documented	evaluation published
Sanshui AssistantChina, Foshan	Service	citation not documented	rights not documented	evaluation not published
Policy NavigatorRwanda	Project	citation not documented	rights not documented	evaluation not published
GovBotKenya	Pilot	citation documented	rights not documented	evaluation not published
CGU-InsightBrazil	Service	citation not documented	rights not documented	evaluation not published
ChatTCUBrazil	Service	citation not documented	rights not documented	evaluation not published
NFRA Insight BotIndia	Project	citation not documented	rights not documented	evaluation not published
SEABISJapan	Demo	citation documented	rights not documented	evaluation not published

A fifth figure exists, French, outside the matrix: L’Assistant, a general-purpose assistant, with 75% of civil servants finding it useful and gains of 12 to 16% observed in discussion groups ¹⁰. It does not change the finding, it confirms it: perception is measured, not document fidelity.

First missing piece: recall. No case publishes the rate of relevant passages found on a set of real questions. Yet it is the measure that governs everything else: if the passage is not in the list of candidates, no model, no reranker can bring it back ³⁹. A system can look excellent and silently miss half of the difficult questions.

Second missing piece: permissions. Only one case explicitly mentions classified documents: AIBots, up to the Restricted / Sensitive (Normal) level ¹⁴. No source describes how rights are inherited down to the passages, nor the filtering at search time. Yet OWASP puts precisely this point at the centre of its sheet on RAG security ⁴⁰. The gap between the importance of the subject and its public documentation is a result in itself, not a research gap.

Third missing piece: cost. Only two elements: $64,000 for the Minnesota pilot ³⁰ and billing per million tokens in North Carolina, with no rate published ¹⁶. No cost per indexed document, per query or per active user. Any comparison between architectures would therefore be an estimate disguised as a result.

Fourth missing piece: evaluation. Four cases out of 23 systems, no published question set, no exact citation rate, no abstention rate. Seven cases document citations, which is better than nothing, but a displayed citation is not a verified citation ^{12, 24, 3, 31, 18, 34, 13}.

A reading precaution is needed here: “not documented” does not mean “absent”. The matrix records what the sources say, not what the systems do. An empty dot means that no public source describes the function, not that it does not exist. It is a difference that hurried readers often confuse, and that leads some to say that public systems have no permissions because nobody described them.

The sentence to remember is not a condemnation: the public sector knows how to count users; it does not yet know how to prove that an answer is right. It is a work in progress, not a verdict.

7. Why RAG remains the right answer

If the evidence is missing, should we conclude that the technology is not worth the investment? No. RAG answers a real problem, and it does so better than the available alternatives for procedures that change.

It makes the answer verifiable. A model alone produces a plausible text; a RAG produces a text accompanied by the passages that grounded it. The civil servant can open the document, check the page, contest. This is the difference between an answer and an assertion ¹⁹.

It lets you add a document without retraining the model. A procedure published today can be indexed tomorrow. This is what makes RAG suited to a living corpus, unlike fine-tuning, which changes the behaviour of the model without guaranteeing freshness ¹⁹.

It must not be confused with two other things. RAG is not fine-tuning: one brings documents, the other changes behaviour. RAG is not agentic either: GOV.UK recommends classic RAG by default, and presents the agentic approach as longer, more expensive and harder to debug ²⁰. These are architecture choices, not degrees of modernity.

The cases that work are bounded. Assist works on five maintained documents ¹²; China Maritime on a base of 5,488 question-answer pairs ¹³; North Carolina on the public content of agency websites ¹⁶. None claims to cover “all the knowledge of the administration”. The narrow perimeter is not a limit of RAG: it is its condition of success.

One question remains that the panorama does not settle: is RAG the right answer for every need? No. For a question whose answer fits in a structured sheet, a classic search is enough. For a synthesis over an unstable corpus, a poorly governed RAG will add confidence to errors. The technology is suited when the answer must be grounded, cited and kept up to date, not when it merely has to be fluent.

8. What this changes for a public project

The IGF–IGAS–IGA report makes thirteen proposals for deploying AI in administrations: mapping use cases, integration into information systems, reversibility, training, governance ⁴¹. A well-designed document architecture directly serves several of them. It does not replace them.

Five decisions weigh more than the choice of the model.

1. A narrow, business-oriented perimeter. One corpus, one trade, a team that maintains it. The opposite, “all the knowledge of the local authority”, produces an index with no authority and an answer with no one responsible.

2. A real question set, before the tool. Fifty to two hundred questions asked by real civil servants, about 20% of which the corpus cannot answer ⁴². Without unanswerable questions, you do not measure abstention; you measure the politeness of the model.

3. Recall before generation. Measure whether the right passages are found, separately from the quality of the writing. A system that answers well from the wrong passages is a confidence generator, not a document tool ³⁹.

4. Human validation on sensitive uses. The cases that mention it (China Maritime, JRC, Digdir) present it as necessary, not as a leftover of the transition ^{13 33, 4}. A RAG that produces a decision without review is no longer an assistant.

5. Stop criteria written in advance. MILEVA and Intelligence Entreprise stopped ⁶; Minnesota closed its pilot without a review ³⁰. A project must say, before it starts, what would make it stop: insufficient recall, a cost per query out of budget, unfunded document maintenance.

The five questions to ask before signing

1. Which exact corpus, maintained by whom, how often?
2. On which real questions will you be evaluated, and how many are unanswerable?
3. How are access rights carried by the passages and filtered at search time?
4. What exact citation rate and what abstention rate do you accept?
5. What is the cost per indexed document and per query, and who pays for the update?

The fifth question is the one forgotten fastest. An index is not a one-off investment: it is a document subscription. Assist holds only because five documents are explicitly maintained ¹²; AIBots holds only because agencies contribute and update their corpora ¹⁴. Cutting the maintenance budget means scheduling an index that degrades without anyone noticing.

None of these questions is about the model. That is the point: the model is the most visible and least decisive part of the project.

9. What the series will show

This article set the scene and the thesis. The next four open the hood, each with a single question.

Article	Question	What it will bring
2. How does a document RAG really work?	What happens between a PDF and a cited answer?	OCR, structure, chunking, hybrid search, reranking, citations
3. Securing a RAG: permissions, injections and confidentiality	What must a system demonstrate before a security officer authorises it?	per-passage permissions, indirect injection, lifecycle, sovereignty
4. Why RAG projects fail in production	What breaks, and how do you measure it before promising?	stops, document quality, evaluation, costs, observability
5. Building a reliable document assistant, from A to Z	Which concrete choices, with which trade-offs?	corpus, repository, tests, demonstration, architecture decisions

The four missing pieces of evidence are not a condemnation of RAG. They are the four workstreams of the following articles: the mechanics (article 2), security (article 3), measurement (article 4) and construction (article 5).

10. Conclusion

RAG really does change access to knowledge in administrations: it makes searchable what was scattered, and it makes verifiable what was asserted. But the transformation is not where one expects it. It is not in the model, whose versions keep succeeding one another. It is in the documentation: an identified, versioned corpus, whose rights are known and whose quality is measured.

That is the lesson of the 23 systems surveyed. The best documented are not the most spectacular; they are those of which one can say, precisely, what they do, on what, for whom, and with what limits. The others produce usage figures and leave the most important question unanswered: was the answer right?

At Seshat, we apply this rule to our own work: distinguishing what is available, what is experimental and what is not measured. A document system that cannot say where its answer comes from is not progress; it is one more black box.

Sources

The sources are cited with their original titles, in the language of the source.

1. Cour des comptes, enquête sur les Archives nationales. link
2. Archives nationales, rapport d’activité 2025. link
3. BDBOS / ASDN (Lower Saxony, Germany), KIRA. link
4. Digdir (Norway), KI-verktøy for identifisering av digitaliseringstiltak. link
5. Ministério da Gestão e da Inovação (Brazil), Guia de introdução ao RAG no setor público. link
6. Fiches beta.gouv.fr (project status sheets). link
7. Cour des comptes, Mieux suivre et valoriser les gains de productivité de l’État issus du numérique. link
8. El Paso, Ask Laura chatbot. link
9. GovTech Singapore, Pair. link
10. DITP / DINUM, L’Assistant IA : bilan de l’expérimentation menée auprès des agents. link
11. Open Contracting Partnership, AI in public procurement. link
12. Cabinet Office (United Kingdom), Assist, algorithmic transparency record. link
13. Ministry of Transport (China), China Maritime Intelligent Customer Service. link
14. GovTech Singapore, AIBots. link
15. Hansel (Finland), Hankintavälkky in the Hilma portal. link
16. NCDIT (North Carolina), AI Chatbot. link
17. Albert API (DINUM / Etalab). link
18. City of Seoul, Seoul Data Hub. link
19. GOV.UK / GDS, AI Insights: RAG Systems. link
20. GOV.UK / GDS, AI Insights: Agentic RAG. link
21. RISA (Rwanda), GoR AI-Powered Policy Navigator. link
22. Localtis / Banque des Territoires, collectivités et IA générative. link
23. European Parliament, Ask the EP Archives. link
24. ECB, ECBot, privacy statement and technical description. link
25. GSA (United States), 2025 GSA AI use cases. link
26. U.S. Federal AI Observatory, analytical mirror of the OMB 2025 inventory. link
27. GAO (United States), Generative AI Use and Management at Federal Agencies. link
28. NARA (United States), AI use case inventory. link
29. US Department of Labor, use case inventory. link
30. Minnesota IT Services (MNIT), TMF nebulaONE project. link
31. GovStack, GovBot, AI Chatbot for Discoverability of Government Services (Kenya). link
32. JRC (European Commission), AI applied to disasters and crises management. link
33. JRC (European Commission), Applying Generative AI to Enhance Nuclear Safeguards Information Analysis. link
34. Digital Agency of Japan, Co-creation PF Camp workshop (SEABIS). link
35. IndiaAI / NFRA, Financial Reporting Compliance Challenge. link
36. Network Rail (United Kingdom), RAG chatbot on lessons learned. link
37. Network Rail, Wales & Western data strategy (APM). link
38. National Data Administration (China), RAG annotation case. link
39. Late interaction and reranking. link
40. OWASP, Top 10 for LLM Applications 2026 and RAG Security Cheat Sheet. link
41. IGF, IGAS and IGA (joint report), Le déploiement de l’intelligence artificielle dans les administrations publiques : analyse comparée, enjeux et conditions de réussite. link
42. Evaluation. link