Artificial intelligence in administrative work: what to delegate and what not to.
Drafting, summarizing and preliminary analysis, yes. Decisions about people and confidential data, only with written rules.
The conversation about artificial intelligence in mid-sized Colombian companies swings between two equally unproductive extremes: banning it outright, or letting everyone use whatever they want with whatever information they want. Both positions have the same origin, which is not having defined rules.
Where AI adds real value in an administrative area is in language tasks with human supervision: drafting a first version of an internal communication, summarizing a long document, rewording a procedure to make it clearer, preparing interview questions, or translating. In all of them, the output is an input that a person reviews before using.
Where it should not enter is decisions about people. Evaluating a résumé, rating performance, deciding a promotion or determining a salary adjustment are decisions with legal and human consequences that require an identifiable person responsible. Using a model to rank candidates, moreover, introduces bias risks that are hard to audit and to defend against a claim.
The most urgent practical problem, however, is not philosophical but about confidentiality. Pasting a payroll file, a list of employees with salaries or a client contract into a public tool means handing personal information to a third party without the data subject's authorization. That is a matter of Colombia's data-protection law (Law 1581 of 2012), not a debate about technology.
The minimum guidelines we help our clients write fit on one page: which tools are authorized, which categories of information are never pasted into them, how data is anonymized when the analysis requires it, who reviews the output before using it, and what to do if someone made a mistake. That last question is the one almost always missing, and the one that determines whether incidents get reported or hidden.
Team training has to run in parallel. A policy is useless if people do not know what can be asked or how to verify an answer. In the programs we deliver, the most useful exercise is not learning to write prompts but learning to detect when the answer is wrong: an invented regulatory reference, a figure that does not add up, a recommendation that ignores the Colombian context.