In an SME, poorly applied AI costs money and time. In government, it costs rights: a poorly audited appointment algorithm can leave without care those who need it most —and the citizen has no other hospital to go to. An opaque decision on health, education, security or resources affects people who did not choose that system and cannot switch providers. The bar is higher: what is good practice in business is a requirement here — auditability, equity, transparency, accountability.
Government also concentrates three ills AI can ease or worsen: bureaucracy that wears down civic trust, data silos between agencies, and distrust when the algorithm decides without explanation.
What I do not propose
I do not propose installing systems or buying software as if modernization were a catalogue. I do not propose replacing the public worker: I propose freeing their capacity to manage and serve. AI is an extension of their judgment, not a substitute. And I do not propose black boxes: in decisions about rights, a model that cannot be audited cannot be used.
Three principles in public terms
The citizen at the center. The procedure is redesigned from the experience of whoever suffers it, not from the org chart.
AI as extension of the public worker. Public judgment is not replaced; it is amplified with evidence. The employee decides better and faster, not less.
Auditability as a condition, not decoration. Governance enables trust, and trust enables adoption. The more autonomous the agentic layer —a prioritized file, an assigned slot, a case routed alone— the higher the requirement: human control is exercised over the design and audit of the rules, not case by case on every transaction.
Multichannel access when rights are at stake
Modernization that removes the counter without a real alternative is not efficiency: it is institutionalized exclusion. A 100% digital procedure with no assisted kiosk, phone slot or in-person point in neighborhoods with unstable connectivity leaves out those policy should serve. The bridge demands mandatory multichannel when the procedure affects rights or opportunities, and published equity metrics: completion by segment, funnel drop-off, disaggregated resolution time. State efficiency is also measured by who did not get left behind.
Before code: impact assessment
For high-impact treatments —health, minors, biometrics, scoring with legal effects— public hygiene starts before deployment: a light but non-delegable impact assessment (PIA/AIA). Document purpose, legal bases, affected groups, bias risks, mitigations and a human appeals channel. Three questions before each public release: what personal data? is there a less intrusive alternative? who answers if the algorithm is wrong?
Governance of public data
Breaking silos without breaking privacy is the data link’s job: reliable, interoperable information available for evidence-based decisions. Without a data owner, lineage and shared definitions, the ministry dashboard and the neighboring agency’s tell different stories —and public policy decides blind.
IMIA (AI maturity index) is the instrument I use to measure whether an agency is ready for AI to add value before installing anything: seven dimensions, six levels, and a governance gate that caps the level when verifiable rules are missing. It is not marketing: it is a design condition.
Minimum legal frame (without replacing counsel)
Before signing contracts or tendering platforms, close predictable questions: ownership of code developed with public funds, controller and processor roles for personal data, open-source dependency licences, digital signature validity when the act requires it, and exit paths against vendor lock-in. The bridge does not replace legal counsel; it avoids surprises that later cost years of litigation or dependency.
Sustainability without greenwashing
Sensors installed with international funding, maps published, and two years later nobody calibrates or integrates with real works: eternal pilot. Smart city without an integrated municipal ticket is showcase, not service. Technological sustainability includes infrastructure footprint, digital inclusion and equity in the transition —who absorbs the cost of digitization— not only a green label on the report.
The 2025 state of the art: the OECD as map, CEPAL as anchor
Two 2025 publications refresh the evidence. The OECD, in Governing with Artificial Intelligence (2025), surveys around 200 real cases of AI use in governments across 11 core functions of the State, under a framework of three pillars: enablers, guardrails and engagement. That ordering confirms governance is not an annex to the project but one of the three axes —the public-sector links are enablers; impact assessment and auditability are guardrails; multichannel access and the citizen at the center are engagement.
With a geographic caveat the bridge keeps: OECD evidence is global and illustrates good practice; the Latin American scale is anchored in the ILIA 2025 (CEPAL/CENIA). The Latin American Artificial Intelligence Index, in its third edition, covers 19 countries of the region across three dimensions —enabling factors; R&D and adoption; governance— and classifies national ecosystems as pioneers, adopters or explorers by maturity. When Argentina or the region must be placed against the frontier, the reference is ILIA, not a global average.
What to do differently
- Diagnose the real procedure, not the org chart in PowerPoint.
- Require multichannel when the procedure touches rights.
- Measure maturity (IMIA) before scaling AI pilots.
- Publish equity, not only volume processed.
- Design audit as part of the product, not as a later phase.
Modernization of the State in Latin America is not bought: it is built by strengthening institutional capacity and putting the citizen at the center (CEPAL, 2024). A transformation that does not replace the public worker but frees their capacity to manage and serve —using AI as an extension of their judgment.
Full framework in The Sociotechnical Bridge and My approach. Portfolio case: Public-sector process redesign. Related essay: SME maturity is not a straight line.