Peru's new framework for public-sector Artificial Intelligence Officers raises a practical question for every administration adopting AI: who has the authority, knowledge and resources to keep its use accountable? The answer depends on how responsibility is organised throughout a service, from procurement to everyday decisions
From a national announcement to an institutional question
An AI system can enter public administration through many doors: a supplier’s software update, an experimental chatbot, a document-processing service or a tool that helps officials prioritise cases. Each route creates a similar challenge. Someone must understand what the system does, how it affects a public service and where concerns should be raised.
Peru provides a timely starting point for examining that challenge. Published on 21 October 2026, Resolution 004-2026-PCM/SGTD approves a directive establishing the profile and responsibilities of the Oficial de Inteligencia Artificial, or OIA. The resolution situates the role within the country’s existing AI and digital-government framework, including its national AI strategy for 2026–2030. It connects specialised coordination with the public bodies responsible for digital transformation.
The significance extends beyond a new job title. An institution needs a reliable way to connect technical knowledge with administrative authority. The discussion below examines how that connection could work in practice; the examples and suggested measures are editorial analysis, rather than reported outcomes of Peru’s reform.
Responsibility needs a route through the organisation
Consider a hypothetical authority using AI to sort incoming applications. The technology team may maintain the software, a contractor may supply the model, and a service manager may determine how its recommendations enter the workflow. A caseworker then sees a suggested priority on screen. If an application is repeatedly overlooked, responsibility can become difficult to locate across those separate functions.
A designated officer can help make those relationships explicit. A useful governance arrangement would identify the service owner, the person authorised to approve changes, the team that investigates errors and the official empowered to pause use. The officer’s coordinating role would connect these responsibilities rather than absorb all of them.
This distinction matters because accountability cannot depend on one specialist reviewing every transaction. It needs to survive staff turnover, supplier changes and pressure to accelerate processing. Written decision rights, escalation routes and records of significant changes would make oversight more durable.
Supplier oversight begins before a contract is signed
The Peruvian resolution expressly requires safeguards for personal data and institutional information when third parties are engaged to assist the OIA. It also provides for implementation within existing institutional budgets. These provisions bring information protection and administrative capacity into the same conversation.
For an administration, a practical implication is to ask governance questions while defining a purchase. Can the supplier explain relevant limitations? Will the authority receive notice when the underlying model changes? Can it export its records and continue the service if the contract ends? What evidence will be available when a resident challenges an outcome?
These are suggested procurement questions, not a description of specific clauses mandated by the resolution. Their purpose is to expose dependencies before an institution becomes reliant on a tool. A low initial price may offer little value if the public body cannot investigate failures, obtain its information or replace the service at a reasonable cost.
Human oversight must be possible during real work
Keeping an official in a workflow does not automatically produce meaningful supervision. A person asked to approve hundreds of recommendations with little time or supporting information may have few opportunities to exercise judgement. An interface that makes acceptance easy and disagreement cumbersome can reinforce that problem.
Effective oversight would therefore require more than a formal approval button. Staff need understandable reasons to question an output, access to the underlying case material and permission to depart from a recommendation. Managers need to know whether those departures reveal recurring weaknesses or legitimate differences between cases.
A proportionate approach is essential. An assistant drafting an internal meeting agenda and a system influencing access to an essential service present different consequences if they fail. Review arrangements should reflect the task, the people affected and the practical reversibility of an error.
Training is an organisational capability
The resolution also envisages coordination on professional competency standards and encourages training partnerships. Its provisions make clear that developing certification arrangements is not a condition that suspends officers’ designation or operation. Building capability and establishing responsibility must therefore proceed alongside one another.
The broader challenge is to distribute enough understanding across an organisation. Procurement staff need to recognise contractual dependencies; service managers need to evaluate whether the tool improves delivery; frontline officials need to spot unsuitable outputs. Senior leaders need to understand what assurance they are receiving and what remains uncertain.
The OECD’s 2025 report, Governing with Artificial Intelligence, provides a useful comparative perspective. Drawing on 200 government use cases, it identifies governance, data, infrastructure, skills, investment, procurement and partnerships as enabling conditions. It also highlights obstacles such as skills gaps, legacy systems and limited impact measurement. This evidence helps explain why appointing an officer is only one component of institutional readiness.
Measure the service, including the work AI creates
A hypothetical document assistant might produce a draft in seconds while creating additional checking work. A resident-facing chatbot might answer more questions yet struggle with unusual circumstances. Counting generated documents or automated responses would miss those effects.
A more informative evaluation would compare the service before and after adoption: completion times, correction rates, unresolved cases, staff effort and the experience of people who need additional support. The authority should also examine who benefits and who encounters new barriers. Average performance can conceal a recurring problem affecting a smaller group.
Evaluation needs to support decisions. If evidence shows little improvement, an administration should be able to change the workflow, narrow the system’s use or discontinue it. Continuing a pilot indefinitely can consume resources while leaving its value unresolved.
What would demonstrate that the framework is working?
The publication of a framework establishes expectations; it does not demonstrate that public services have improved. Evidence of implementation would include clear allocation of responsibilities, staff who can raise concerns, credible supplier oversight and documented responses when systems underperform.
For Peru, the practical test will be whether the officer role connects these activities with everyday administrative decisions. For other governments, the case offers a wider lesson: accountable AI requires people with a defined mandate, access to evidence and the ability to act.
The question is ultimately about public authority. Residents should be able to understand who is responsible for a service and how problems can be corrected, even when several technologies and organisations contribute to it. An AI officer can strengthen that chain of responsibility when the surrounding institution gives the role practical effect.
Illustrative photograph: Government Palace, Lima, photographed in 2009. Photo: Charles Gadbois, via Wikimedia Commons, CC BY 3.0. No editorial alterations; WordPress resizes and may crop display versions.