Artificial Intelligence (AI) platforms are meaningfully evolving every month. Frontier systems now draft and check documents, detect patterns across millions of transactions, triage casework, answer citizens in natural language around the clock, and watch networks for intrusion. Governments are built largely out of documents, casework, transactions and queues. That makes this the largest productivity opportunity the public sector has seen in a generation — and among the hardest to capture.

The development community has done the diagnostic work with unusual speed and candor. We know that more than seventy countries have national AI strategies, that many stall on the way to operational reality, that public administrations face acute shortages of technical expertise, and that the shortage produces over-reliance on external vendors.¹ The problem has been named precisely.

The 2026 World Development Report, published in August, sets out the destination: adopt existing tools, adapt them to local conditions, then advance toward building your own.² What is less settled is how an administration with three technical staff climbs the first two rungs. Our review found programs that embed advisers, support pilots and strengthen technical capacity. What we did not find clearly documented is a repeatable facility combining four things: platform-neutral delivery, co-production with ministry staff, transfer to a permanent government team, and measurement of whether that capability endures.

We write about Latin America and the Caribbean, where the Partnership for the Americas works. The evidence below is drawn from wherever it exists — much of the best-documented material comes from the United Kingdom and the United States — and we say plainly when an example comes from a setting the region does not share.

That is the gap this paper is about, and it can be tested through a limited feasibility and design phase. The objective matters as much as the method: not to deliver AI systems to governments, and not to hand one over and leave, but to help build the permanent internal team an administration needs to select, adapt, govern and sustain them.

We suggest a modest first move: that multilaterals commission a short feasibility and design phase — roughly three months — to test demand with real ministries, work out the procurement path, and return with a costed pilot design.

The stakes are higher for government than for most sectors, because so much of government is made of the work these systems are now good at. Tax authorities process declarations; ministries process files; courts process cases; cities process permits, complaints and traffic; security agencies process alerts. The region has evidence of what is possible. Buenos Aires' WhatsApp assistant Boti fields more than 1.3 million users a month across 450 topics and 70 system integrations. Among Latin American and Caribbean tax administrations surveyed in 2022, 71 percent used data science or analytics, 38 percent virtual assistants, and 24 percent artificial intelligence — building on the region's world-leading digital tax infrastructure.^{3, 4} But that capacity is concentrated. The median ministry or medium-sized city still has no path from idea to production.

Strategies outnumber systems

The World Bank's assessment of public institutions in the age of AI puts it plainly: converting strategies into operational reality is the hard part. The countries that manage it tend to pair ambition with enforceable directives, risk assessments, budgets and performance indicators — practices associated with success rather than proven causes of it.¹ Its GovTech Maturity Index, covering 197 economies, finds something more troubling still: asked whether they track actual usage of the systems they have built, many governments could not say. Policies and platforms exist; the means to tell whether they work often does not.⁵

One central constraint is a shortage of people able to combine technical, operational and commercial judgment. Without that capability, the Bank warns, governments risk data leaks and over-reliance on vendors; public servants already using AI tools without guidance it calls a shadow AI workforce.¹ Harvard's Belfer Center documents the same from the hiring side in Latin America: a thin technical pipeline, private-sector competition, and public hiring that struggles to keep pace.⁶

A gap that is not closing

The IMF's AI Preparedness Index averaged 0.68 for advanced economies, 0.46 for emerging markets and 0.32 for low-income countries in 2023 — a snapshot rather than a trend.⁷ But the World Bank's GovTech data does track movement, and points the same way: higher-income economies have advanced while low-income ones have regressed.⁵

The upside is real but should be described carefully. The UK's trial of Microsoft Copilot across twenty thousand civil servants reported average savings of twenty-six minutes per working day — self-reported, no control group, and the evaluators could not establish how the time was used.⁸ A signal of promise, not evidence of better services. The gap between those two is the problem: the question is not whether the technology can help, but whether governments can get it into production and then tell whether it worked.

Somebody has to be in the room

When large enterprises hit this wall, they paired the software with people. Forward-deployed engineers embed inside the client organization, work on its actual systems, and integrate in place — alongside the product, not instead of it. The model is most closely associated with Palantir and has spread fast. OpenAI has stood up a separate deployment business and agreed to acquire a consultancy to staff it; Accenture describes a practice numbering in the thousands; Deloitte has launched one of its own.^{9, 10, 11}

The analogy has limits: these are partly distribution strategies for platforms, and ministries differ from corporations in ways that matter — statutory authority, procurement law, due process, continuity across administrations. Government digital delivery reached the same place by its own route. The UK's discovery–alpha–beta–live model embodies the principle directly: public services are developed iteratively with users, tested in prototype code expected to be thrown away, and assessed at each boundary before moving toward live operation.¹²

Why AI is harder than ordinary government software

Three things separate AI from a payroll system, and all three argue for permanent capability rather than a one-time build. Outputs are probabilistic, so a system cannot be tested once and declared correct. Performance drifts as data and models change underneath it, making monitoring a standing function rather than a closing task. And the vendor landscape moves fast enough that portability — changing supplier without rebuilding — is a design requirement. A government that cannot evaluate, monitor and switch is exposed. The corollary: AI should not be used where a rule change or conventional software would be safer or cheaper, and part of a competent team's job is saying so.

The intelligent client

There is a second reason to send people, and it may be the more important one. Even an administration that intends to buy rather than build needs enough internal capability to specify what it wants, judge what it is offered, and supervise what arrives. British government guidance calls this the intelligent client function — written for property and estates rather than technology, and borrowed here deliberately. The UK's National Audit Office now uses the term in exactly the sense meant here, recommending that departments strengthen the intelligent-client role in technology procurement and emphasise outcomes over inputs.^{13, 14} Where it is missing in digital procurement, the UK's audit office has documented the results: delays, cost overruns, commercial teams without the depth to manage digital contracts, and growing dependence on suppliers that, in its words, are bigger than governments themselves.¹⁵ A ministry that relies on vendors for advice about the vendors' own products is not buying. It is being sold to.

Capability varies widely and some administrations are formidable buyers. But where a government cannot independently assess a proposal's performance, cost and risk, it is not really negotiating. The Bank's recommendation that governments decide carefully between internal expertise and external procurement presupposes the very capability that is often absent.¹ That capability does not appear to come from training courses alone; it seems to come from having built something once, alongside someone who had done it before — a hypothesis worth testing, not a settled finding.

A review of published documentation, current to August 2026, shows a somewhat populated landscape.

• The World Bank, through GovTech, benchmarks digital government across 197 economies, convenes an AI working group spanning 35 governments and institutions, runs AI bootcamps, and has devoted its flagship World Development Report to AI.^{2, 1, 5, 16}
• The Inter-American Development Bank, through fAIr LAC, has built the reference tools for applying ethical principles across an AI project and supports pilots with its partners and hubs.¹⁷
• UNDP deploys senior technical talent embedded in government — in its own description, for vendor selection, IT architecture and recruitment — alongside local capability building.¹⁸
• GovStack supplies specifications, building blocks and training that let countries design, prototype and scale services, through partner institutions.¹⁹
• CAF, with UNESCO, has taken AI readiness assessment and ethics into member countries, part of a program reaching more than fifty countries.²⁰

Several are excellent, and the overlap with what we propose is real — embedded advisers, prototyping support, pilot financing and capability building all exist today. What the record does not describe is a repeatable facility combining four things at once: platform-neutral embedded delivery, co-production with the ministry's own staff, explicit transfer to a permanent internal team, and measurement of whether that team still runs the system a year later. That is a preliminary finding from a documentation review, and we would rather be corrected early than late.

Three non-commercial efforts come closer than the rest, each stopping short in a different place. eGov Foundation's international arm takes DIGIT, its open-source government platform, into ten countries — Jamaica among them — with solution engineering and local capability building alongside it; but the engagement is organised around one stack and delivered largely through partners, rather than around whatever the ministry already runs.²¹ Bloomberg Philanthropies' innovation teams are the closest precedent for what we would want to leave behind: multidisciplinary units established inside city halls in more than a hundred cities, staffed to implement rather than to advise — municipal, and closer to service design than to production engineering.²² The Center for Civic Futures has committed US$8.5 million to seven two-year projects applying AI to benefits delivery, selected from more than four hundred applications — a signal of how far demand outruns supply, but grants and a learning cohort rather than engineers in the building.²³ Only the first of the three touches this region at all, and it arrives with a platform to install.

The model is no longer the expensive part — in some countries

Something shifted in 2025 that the development conversation has not caught up with, though it shifted in the United States first. Through the US General Services Administration, both Anthropic and OpenAI now offer their enterprise products to federal agencies for one dollar per agency per year — Anthropic's across all three branches, at FedRAMP High, with technical support to help agencies implement.^{24, 25} Those are US federal procurement vehicles. No equivalent exists for a ministry in Bogotá or Kingston.

What is on offer beyond the United States differs in kind. OpenAI for Countries proposes national partnerships built around in-country data-centre capacity, localized deployments and co-invested startup funds — infrastructure and capital, pitched at sovereign scale rather than at the working level of a line ministry.²⁶ The companies are also arriving commercially: OpenAI opened its Latin American headquarters in São Paulo in August 2026, where Brazil is now among its three largest markets worldwide and its named public-sector work is with the city's technology agency.²⁷ Arrival is not delivery capacity, and a commercial office is not a public mandate.

These companies have built serious deployment capacity, and some of it is now turning toward government. OpenAI's four-billion-dollar deployment business sells mainly to the Fortune 500, but the company is now hiring AI deployment engineers for the public sector — to embed with what it calls its most sophisticated public-sector customers as their technical lead, and to build prototypes alongside them.²⁵ That is the closest thing to this model now operating anywhere. It is also a vendor's engineers, working on that vendor's products, hired into Washington for United States federal agencies. Anthropic's mission-driven team, Beneficial Deployments, works with nonprofits, health research and science rather than public administrations.^{9, 28} The capability exists and it is being built out fast. It is simply not aimed at a finance ministry in a middle-income country.

The consultancies, honestly

Which leaves the consulting firms, where the evidence deserves more care than it usually gets. The UK's National Audit Office examined government's use of external consultants in November 2025 and found that 86 percent of officials considered them a valuable contribution, 40 percent extremely so. More precisely: consultants add most value when brought in for specialist expertise the civil service lacks, particularly on digital projects — which is exactly the work described here — and least when used as temporary staff during busy periods.¹⁴ The problem is not that they are bad at this. They are good at it.

The documented failure is what happens afterwards. The NAO found that learning is captured only at the end of a project, when attention has moved on, and is obstructed by staff turnover, intellectual-property terms and vague contracting — with no mechanism for sharing lessons between departments. Australia and France now write knowledge-transfer clauses into consultant contracts as standard. The auditor's recommendations read, almost line by line, as the design brief for what follows: strengthen the intelligent-client role, avoid repeatedly buying in the same task, and start from the assumption that a government's own staff are the better use of resources.¹⁴

The scale of the alternative is worth registering. In 2026 the United Kingdom awarded KPMG and EY up to £456 million to train civil servants in AI and related skills through 2028 — the largest single Big Four government contract on record — to be phased out, the government says, once an in-house model exists.²⁹ That is close to half a billion pounds buying capability as courses, in a wealthy state with a mature digital service. Few governments in Latin America and the Caribbean will ever have that option. The question is whether some of that spending, at any scale, is better used building something alongside the people who will run it.

A proposed shape, not a settled method — anyone claiming to know the right duration of each phase before running one is guessing. It borrows from the discovery-and-assessment model already established in government digital delivery.¹²

A short visit, and an honest list

A small team spends roughly two weeks inside an agency that has asked for help. It examines real systems, real data and real workflows rather than presentations.

It leaves behind four things:

• Diagnosis of where AI would and would not help;
• Shortlist of candidate use cases scored for feasibility and value;
• Costed roadmap written so the ministry's finance officials can read it;
• List of what the government must have in place before anyone can responsibly build anything — data access authority, a named product owner, security clearances, infrastructure, and staff assigned rather than borrowed.

Return when it is ready

The team comes back to build once the government has completed that list. This is the decision that carries the most weight. Nothing is promised that cannot be delivered, and the government's own follow-through — not an external rating of its readiness — determines what happens next. A ministry that assembles the prerequisites has shown commitment in a way no assessment can; one that does not was unlikely to succeed against a fixed calendar anyway. The list should carry a validity window. That commitment is necessary but not sufficient: legal authority, data quality, security and rights risk still require independent verification. And where an otherwise suitable government cannot yet meet the prerequisites, the answer should be a preparatory track rather than exclusion.

The build phase is a route to production rather than production itself: a first working version tested with real users, then the security, legal and rights assurance a public service requires before it carries real cases. Nothing goes live without a named permanent owner, an operating budget, a monitoring plan and an exit route. Whether the right length is six weeks or twelve, and whether a team works best at four people or six, a first set of engagements should answer. What is clear already is that engineering alone is insufficient: this work needs product judgment, service design, security and privacy, and someone who understands both the sector and the language. Teams should draw heavily on local and regional practitioners, both for context and to strengthen the surrounding talent ecosystem.

One tension sits at the centre of this and deserves stating plainly. The administrations that most need help are the ones with the fewest technical staff — and yet the model depends on leaving a permanent team behind. It can only work where a government commits named people, institutional authority and recurrent financing. An embedded team can strengthen a structure the government agrees to maintain; it cannot create civil-service posts or fix underlying pay and retention problems. The realistic destination is not a self-sufficient engineering department but a small government-owned product function able to direct its own staff and its suppliers.

A successful engagement produces three things, listed in the order they appear rather than by importance.

• Something that works. One bounded, useful system operating under government control — with clear rights over data, access, audit, portability and exit — on a problem the agency's own staff selected.
• People who can run it. Named counterparts who built alongside the team and can manage the system's operation, improvement and suppliers without continued dependence on it.
• A team that outlasts the engagement. A small permanent group with the authority, budget and standing to select, adapt, govern and sustain these systems — and to specify and supervise the next project, whether they build it, buy it, or simply ask a supplier a harder question.

The third is what makes this a development proposition rather than a procurement of services. The external team is a transitional mechanism whose job is to help stand up a permanent internal one — the direction the World Development Report points, and the question it leaves open is how an administration with few technical staff gets onto the ladder at all.² Evaluation should follow: alongside service quality, cost and safety, the test at six, twelve and where feasible twenty-four months is whether the service still functions, the accountable people are still named, the financing recurs, and the government can still monitor performance and direct its suppliers.

The evidence does not yet support a large commitment, and we are not proposing one. Whether ministries will commit staff and data access, what this should cost, and which financing instrument fits are all unestablished — and answerable.

A limited three-month technical-assistance assignment can test demand, institutional fit, readiness, procurement feasibility and delivery economics. It cannot establish whether capability endures — that is what a pilot would be for. Five outputs:

• A demand test. Structured conversations with governments and country offices, aimed at specifics: which agency, which problem, which official would sponsor it, and what the government itself would commit.
• A partnership map. What peers already provide, and where this would complement rather than duplicate them — including where a government should simply be referred elsewhere.
• A tested delivery model. The engagement design worked out in practice, ideally including one or two real discovery visits, so it is described from experience rather than invented.
• A financing and procurement path. Worked through with the institution's own legal and procurement staff: how a build phase is tendered so a discovery visit confers no unfair advantage, how continuation is financed, and what the rules actually permit.
• A costed pilot design and a recommendation — unit costs, staffing, travel, procurement lead times, country contributions, and the recurring cost of running a system after handover — including, explicitly, a recommendation not to proceed if the evidence does not support it.

This commits an institution to learning something rather than building anything. Whether the delivery facility itself sits inside the institution or with an external operator is a question the design phase should answer; the country-level capability it strengthens stays government-owned either way.

Better to set out the open questions than paper over them. Will governments complete a prerequisites list, and fund a permanent team afterward? Those are the real tests of demand; enthusiasm in a meeting is no substitute. What team composition works? How is a build phase procured so a discovery visit does not prejudice competition — settled before the first engagement, not after. And how much transfers between countries?

Three risks deserve naming. The first is dependency: embedded assistance can substitute for capability rather than create it. The answer is the objective itself, made measurable — engagements judged on whether officials can run the system without the team at six and twelve months. If they cannot, the engagement failed regardless of what shipped.

The second is rights. The World Bank's synthesis documents what failure looks like at scale in AI and automated decision systems — the Dutch childcare benefits scandal, where discriminatory algorithmic risk-profiling wrongly accused tens of thousands of families;³⁰ the UK's pandemic-era exam grading algorithm, a statistical allocation model rather than machine learning — and concludes such failures violate rights at scale and fall hardest on marginalized populations.¹ The implication is a formal risk assessment, and a presumption that early engagements begin away from automated decisions on benefits, health, enforcement, policing or justice.

Risk should be assessed properly rather than by category. A bounded internal workflow is not automatically safe: case triage shapes who gets seen first, internal search can surface what a user should not see, summarization can distort the record a decision rests on. What matters is the consequence for the person affected, the sensitivity of the data, how much of the decision is automated, the scale, whether it can be reversed, and whether appeal and fallback exist. Human review is necessary but not sufficient — reviewers tend to defer to the machine.

The third is duplication, which is why the partnership map belongs in the first three months rather than a footnote. If another institution is better placed, that is a useful finding reached early. And teams rotating across finance, tax, health and security institutions concentrate sensitive knowledge in a small group — requiring compartmentalization by country, personnel vetting, and a written record that separates patterns safe to publish from specifics that plainly are not.

Governments are being handed a technology capable of augmenting a great deal of public work, at a moment when few people can help them use it well and many are positioned to sell it to them. The open question is how administrations with thin technical benches take the first step. Teams that build alongside public servants — and leave behind a permanent group more capable than the one they found — are one answer worth testing, and development institutions are among the few actors with the reach and relationships to test it.

1. 1World Bank, "Public Institutions in the Age of AI: Emerging Practices" — source for the seventy-plus national AI strategies, the literacy and expertise gaps, over-reliance on vendors, the "shadow AI workforce," the internal-expertise-versus-procurement recommendation, and the failure cases in Section VII. The high project-failure figure it reports is a cross-sector estimate drawn from RAND, not a public-sector-specific rate.
2. 2World Bank, World Development Report 2026: The Promise of Artificial Intelligence (August 2026), and the accompanying release, "AI Offers Lifeline to Developing Economies in an Era of Weak Growth" — source for the adopt–adapt–advance sequence and the emphasis on procurement frameworks, institutional quality and skilled public workforces.
3. 3OECD Observatory of Public Sector Innovation, "Boti, the City's WhatsApp!", on the Buenos Aires city assistant.
4. 4Inter-American Center of Tax Administrations (CIAT), "Digitalization and Digital Transformation of the Tax Administration in Latin America and the Caribbean".
5. 5World Bank, GovTech Maturity Index 2025 Update: 197 economies assessed; monitoring of actual usage identified as a significant gap; higher-income economies advancing while low-income economies regress. The index measures maturity rather than performance and relies substantially on self-reported survey responses.
6. 6Belfer Center for Science and International Affairs, Harvard Kennedy School, "Closing the Government Tech Talent Gap: Lessons for Latin America".
7. 7IMF, AI Preparedness Index, covering 174 economies on a 0–1 scale. Group averages for 2023: 0.68 advanced economies, 0.46 emerging markets, 0.32 low-income countries. A single-year cross-section, not a time series.
8. 8UK Government, "Microsoft 365 Copilot Experiment: Cross-Government Findings Report": 20,000 participants across 12 organisations, September–December 2024. Time savings were self-reported, there was no control group, and the report states it was not possible to identify how the saved time was spent.
9. 9OpenAI, "OpenAI Launches the Deployment Company" (May 2026): an initial investment of more than US$4 billion and an agreement to acquire Tomoro, bringing approximately 150 forward-deployed engineers. Cited as evidence of announced investment, not of results.
10. 10Accenture, "Accenture Launches Microsoft Forward Deployed Engineering Practice" (March 2026).
11. 11Deloitte, "Announcing Forward Deployed Engineering" (December 2025). The announcement does not state the practice's scale.
12. 12UK Government Digital Service, GOV.UK Service Manual. Note that alpha explicitly produces prototype code teams should "expect to throw away" — it is not a production build, and the phases after it are where a service becomes real.
13. 13UK Government, "Intelligent Client Roles: Functional Guidance". This guidance sits under Government Functional Standard GovS 004: Property; the term is borrowed here by analogy to digital and AI procurement, where no equivalent named function exists.
14. 14UK National Audit Office, "Government's Use of External Consultants" (November 2025) — source for the 86 percent and 40 percent figures, the finding that consultants add most value on specialist digital work and least as temporary staff, the knowledge-transfer failures, the Australian and French contract clauses, and the recommendations on the intelligent-client role and consultant dependency. On the risks of buying analysis rather than capability, see also the Australian Department of Employment and Workplace Relations report for which Deloitte agreed a partial refund after fabricated citations and a false judicial quotation were identified, reported in Fortune (October 2025).
15. 15UK National Audit Office, "Government's Approach to Technology Suppliers: Addressing the Challenges".
16. 16World Bank, Global Program on GovTech and Public Sector Innovation, including the GovTech Maturity Index, the AI working group and the AI bootcamps.
17. 17IDB, fAIr LAC.
18. 18UNDP, "UNDP Highlights Successful Digital Transformation Strategies at Inaugural GovTech Congress".
19. 19GovStack, Country Collaborator.
20. 20UNESCO, "UNESCO to Support More Than 50 Countries in Designing Ethical AI Policy", the program under which CAF and UNESCO have worked with member countries in Latin America.
21. 21eGov Global, "Your Partners for DPI in Public Service Delivery": DIGIT, an open-source platform of reusable building blocks and shared data registries, deployed across ten countries — Mozambique, Liberia, Senegal, Burundi, Ethiopia, Nigeria, Jamaica, Sri Lanka, Indonesia and India — with solution engineering and local capability building offered alongside. The foundation describes DIGIT-led implementations as led by partners, of which it lists more than fifty across technology, government and civil society. See also its partnership with the Clinton Health Access Initiative for low- and middle-income countries.
22. 22Bloomberg Philanthropies, "i-Teams and Innovation Programs": an innovation team is described as "a multidisciplinary, mission-minded unit" within city government, charged with helping officials design solutions and implement "resident-centered interventions that improve people's lives." The program reports reaching more than 100 cities representing over 100 million residents. Reach figures are the program's own.
23. 23Center for Civic Futures, "Center for Civic Futures and Partners Commit $8.5M for AI Solutions That Improve Safety Net Program Delivery" (December 2025): seven two-year awards selected from more than 400 applications across 45 US states, drawn from the US$20 million Public Benefit Innovation Fund. Awardees receive grants and technical assistance through a cohort model rather than embedded engineering teams. United States only.
24. 24US General Services Administration, "GSA Strikes OneGov Deal with Anthropic" (August 2025): up to a year of Claude for Enterprise and Claude for Government for US$1 per participating agency across the executive, legislative and judicial branches, at FedRAMP High, with technical support for implementation.
25. 25OpenAI, "Providing ChatGPT to the Entire U.S. Federal Workforce" (August 2025). On deployment engineering for government, see OpenAI's posting for AI Deployment Engineer, Public Sector — Federal Civilian (Washington, DC), which describes the role as to "deeply embed with our most sophisticated public sector customers as the technical lead" and to "intervene directly to accelerate customer time to value through building hands-on prototypes." A posting is evidence of a function being built, not of delivery at scale.
26. 26OpenAI, "Introducing OpenAI for Countries" (May 2025): in-country data-centre capacity, localized deployments and co-invested national start-up funds, with an initial target of ten country or regional projects.
27. 27Infobae, "OpenAI inicia operaciones en América Latina con una oficina en Brasil" (August 2026): São Paulo regional headquarters; Brazil among the company's three largest markets; named work with Prodam, the São Paulo city technology agency, alongside academic and health institutions.
28. 28Fast Company, "Anthropic Hires a Top Biden Official to Lead Its New AI for Social Good Team". The team's stated mandate is to support deployment "in ways that are good for the world but might not be incentivized by the market," working with nonprofits, health research, education and science. Disclosure: this paper was prepared with the assistance of Claude, a model made by Anthropic.
29. 29Consultancy.uk, "Government Hands KPMG and EY Huge Contract for AI Training of Civil Servants": up to £456 million across 2026–2028, the largest single Big Four UK government contract on record, which the Cabinet Office states will be phased out once an in-house training model is in place.
30. 30Amnesty International, "Xenophobic Machines: Discrimination Through Unregulated Use of Algorithms in the Dutch Childcare Benefits Scandal" (2021). Cited alongside the World Bank synthesis as the primary account.