One hundred meters south of the gap

AILA identifies which AI capabilities the country lacks to take advantage of artificial intelligence; the Human Development Atlas shows where. Combined, they stop being diagnostics and become a tool for designing public policy.

UNDP Costa Rica published two instruments in 2026 that cannot be read separately. The Evaluation of the Artificial Intelligence Landscape (AILA), developed in agreement with MICITT, measures national preparedness to adopt AI. The Cantonal Human Development Atlas, describes development conditions in the 84 cantons. Read together they answer the question that matters for designing public policy: what capabilities to build, where, and in what order.

The starting point is recognizing what each instrument sees and what it doesn't. AILA has thematic resolution, but not territorial: it says which capabilities are weak, not where. The Atlas has territorial resolution, but doesn't yet measure AI: its chapter 5 anticipates risks and opportunities without measuring adoption. That asymmetry is not a flaw in either one. It's the reason they need each other.

What AILA says beyond the average#

Reading AILA only by its stage, "systematic," or by its three pillars (ecosystem 2.7; governance 2.6; regulation and ethics 2.4) wastes its most useful information. That information is in the dispersion of its eleven sub-pillars, which range from 2.0 to 3.0 following a clear pattern.

The highest scores correspond to leadership capabilities: AI governance and regulation. Green computing also scores high, but the report itself clarifies that this result is sustained by connectivity and the renewable electricity matrix, not specialized computing (p. 26). The lowest correspond to operational capabilities: trust and security, applied ethics, public sector capacity to use AI, talent and technology. The country has a clearer sense of where to go than how to operate along the way.

That reading defines what kind of policy is appropriate. Leadership gaps are addressed through strategy and regulation; operational ones, through implementation tools: training civil servants, shared platforms, incident protocols and oversight. Surveys confirm this. Responses predominate that rate AI literacy of civil servants as low or very low, and those that deny the existence of mechanisms to monitor ethical principles (pp. 31 and 43).

AILA also contains a clue it cannot solve alone: territory. According to the ITU, 86% of urban households have internet access, compared to 76% of rural ones. The workshops, for their part, pointed to connectivity outside the Greater Metropolitan Area as a direct constraint (pp. 10 and 25). The instrument detects that the gap has geography, but its national design doesn't allow it to locate it. That's where the Atlas comes in.

What the Atlas adds: four countries in one#

The Atlas shows that the average also deceives in human development. In 2024, 60% of the population lived in cantons with very high human development; when adjusted for inequality, that proportion dropped to 17.7% (chart 2, p. 17). For AI policy, its most useful contribution is the typology it builds by crossing the HDI with the Index of Vulnerability to Drugs and Related Activities (IVDAC) (tables 11 and 12, pp. 54–55).

Averages by type, 2024. Source: Cantonal Human Development Atlas 2026, tables 11 and 12.

Cantons in transition have an HDI nearly equal to critical ones, but better citizen security and much lower vulnerability. Territories under tension exceed the national HDI and yet their IVDAC is 66% higher than that of consolidated ones. The correlation between HDI and IVDAC is only moderate (−0.532): development does not guarantee low vulnerability (table 10, p. 51).

The consequence for policy design is direct. Two territories with the same development level may require different interventions. An investment in connectivity and digital literacy finds a more favorable institutional and security environment in a canton in transition than in a critical one with equivalent HDI. HDI alone does not distinguish these cases; the typology does.

The intersection: what capability, where and in what order#

Crossing both instruments produces something that neither offers separately: a design matrix. The operational gaps that AILA points out do not weigh equally in all territories, and the Atlas makes it possible to locate where they accumulate. Three intersections show this with clarity.

The first is trust and security, AILA's lowest sub-pillar (2.0). Table 13 of the Atlas associates territories under tension with risks of mass surveillance, cybercrime and data misuse, and identifies them as the space where AI can be applied to urban management, mobility and security (p. 57). The nation's weakest capacity coincides with the territories where the most sensitive uses are most likely to be deployed first. There, the correct sequence puts safeguards before scale.

The second is institutional capacity. AILA places public sector capacity to use AI at 2.3. The Atlas adds the local scale: six of the ten cantons with the highest vulnerability do not provide municipal services for security and community surveillance, according to the Comptroller's maturity measurement that feeds into the IVDAC (table 8, p. 47). In critical territories the capacity gap is twofold, national and municipal, and no technology deployment compensates for it on its own.

The third is access. The Multidimensional Poverty Index of the Atlas already includes internet use as an indicator of deprivation (annex 7.2, p. 84), and its value in critical territories (0.074) more than doubles that of consolidated ones (0.031). The rural gap that AILA detects with ITU data is, in the Atlas, given a canton name.

The last two columns result from the crossing; they are not textual recommendations from either report.

The matrix also orders time. In critical territories, what is urgent is not deploying AI, but building the conditions that AILA identifies as prerequisite: interoperable data, connectivity, and capacities. In consolidated ones, the challenge is ensuring innovation doesn't stop there. A single national strategy can sustain all four sequences; what it cannot do is apply only one.

This reading reframes a recurring discussion: whether the challenge is one of strategy or execution. A strategy executed the same way in four different realities would not fail from lack of vision or execution, but from lack of resolution. Territorializing means, at once, revising the design and improving execution.

Measuring to correct#

The Atlas leaves a direct methodological lesson: the national average can rise while the distance between territories grows. If the use of a tool goes from 70% to 90% where conditions were better and from 20% to 25% where conditions were worse, adoption improves and the gap grows from 50 to 65 points. A follow-up that only reports users, pilots, or registered licenses would record that result as a success.

Table 14 of the Atlas proposes nine dimensions for building canton-level indicators on AI: digital infrastructure, digital inclusion, use and appropriation, data, human talent, digital government, productive sector, cybersecurity, and governance (p. 61). Most have a direct equivalent in the AILA's subpillars. In practice, that table is the skeleton of a canton-level AILA: the same capability architecture, with the resolution of the 84 territories.

That convergence also covers the limits of each instrument. Two of the three AILA pillars focus on the State: its readiness to use AI and its capacity to regulate it. This is coherent with an instrument oriented toward public policy, but it gives less weight to knowledge production, investment, businesses, and computing. Table 14 explicitly incorporates the productive sector (companies using AI, SME digitalization, automation), precisely one of the dimensions where a next national measurement would gain depth.

A readiness assessment is not a census of AI in the country. It measures conditions, and some of those conditions are the perceptions of those who must execute. That's why its methodology combines surveys with documentary review, international indicators, expert consultations, and workshops (AILA, pp. 14–16). And that's why it matters that both instruments publish their data and formulas (AILA, p. 59; Atlas, annexes 7.1 and 7.2): a public instrument gains legitimacy when others can recalculate it, question it, and improve it.

Where design happens#

The Atlas itself warns that its results do not constitute public policy recommendations nor do they substitute institutional planning (p. 62). The same applies to AILA. Neither decides alone, but together they allow for better decisions: AILA says what capabilities are missing and the Atlas says where that lack weighs most and with what risk. The design of public policy occurs at that intersection.

The underlying question that both reports leave is not how much the average rose, but what happened in each of the four countries that fit inside one. If within a few years Costa Rica has more systems, more users, and more licenses, it will have shown that it knows how to adopt technology. If the Atlas typology also changes shape, it will have shown that it knew how to use evidence to expand opportunities.

Author's note. I participated in the research and analysis process of AILA. The interpretations in this article are exclusively personal and do not represent the position of MICITT or UNDP. The crossings presented are my own analytical exercise.

Sources: UNDP and MICITT, Evaluation of the Artificial Intelligence Landscape: Costa Rica (2026), pp. 10, 14–16, 18–19, 25–26, 30–31, 39, 43, and 59. UNDP, Atlas of Human Development by Canton 2026, chart 2 (p. 17), table 8 (p. 47), tables 10 to 14 (pp. 51–61), p. 62, and annex 7.2 (p. 84).