The figure is attractive but treacherous. Asobancaria confirms that 73% of Colombian financial institutions already use artificial intelligence. Most read the figure as a sign of progress. It is the opposite. It is the sign that the baseline moved and real competition has just begun.
What the 73% does not say
Asobancaria reported that the two technologies most implemented by the financial industry are AI and Big Data, mainly in asset management, algorithmic trading, credit underwriting, product personalization, and cybersecurity. The aggregate picture looks strong.
The granular picture looks different. Most of that adoption is in assistive generative AI: customer service chatbots, analyst copilots, product recommendation engines. These are applications that improve processes without transforming them. They reduce time, they do not redefine the operating model. And everyone is doing them.
The real frontier: AI that executes without asking permission
The industry has spent several quarters talking about agentic AI, but few institutions operate it. The difference is structural. An agentic AI system does not assist a human who decides. It decides and executes autonomously within a defined governance framework. It closes the loop between the client's digital request and the bank's internal validation, which is exactly where Colombian banking has its biggest operational bottleneck.
Gartner has projected that by 2028, one third of enterprise applications will incorporate autonomous agentic AI. Whoever starts operating it in 2026 captures three years of advantage in learning curve, proprietary data, and auditable governance models. Whoever starts in 2028 enters an already consolidated market.
Three common counterarguments and why they fail
"Regulation does not allow operating autonomous AI in banking." Partially false. The Superintendencia Financiera and Colombian regulation on AI in financial services are under active construction, but the current framework allows autonomous agents provided they operate with explainability, audit, and human-in-the-loop oversight. The regulatory constraint is real but not blocking. It is design.
"My bank is too small to invest in agentic AI." Fully false. Agentic AI is precisely the opportunity for mid-market financial institutions: credit unions, trust companies, fintech, second-tier banks. The large players already have scale and need it less. The small ones can skip entire layers of legacy infrastructure that the large carry as debt. This is asymmetry in favor of the mid-market.
"Let us wait for the technology to mature." The most expensive argument. The underlying technology is already mature. What is still being built are corporate governance patterns, control metrics, and operational frameworks. Whoever participates in that construction shapes it to their advantage. Whoever waits receives it designed by the competition.
Having AI no longer differentiates. Having AI that executes does. Mid-market banks that grasp this in 2026 gain three years of advantage.
What someone who sees the play actually does
Institutions already operating agentic AI in 2026 share four structural decisions that distinguish them from the 73% generic adoption.
They define a closed-loop use case, not an open pilot. The difference is that the chosen use case (typically onboarding, credit validation, or claims) is measured in cycle time, not NPS. If cycle time drops measurably, the use case justifies expansion. If not, it is discarded without sentimentality.
They build explainability before speed. The agentic AI models that scale are those that can defend themselves to an auditor or regulator in five minutes, without recourse to the data science team. That requires architecture, not good intentions.
They separate the agent's decision from the agent's execution. In institutions that do it well, the agent decides but execution has a control layer that can revert, pause, or audit any action. This reduces catastrophic incident risk without sacrificing operational autonomy.
They invest in mixed talent, not pure data scientists. McKinsey has documented that enterprise AI projects that scale have a typical ratio of one data scientist for every four product, risk, and operations profiles. Most of the Colombian mid-market invests in the inverse ratio. The result is predictable: brilliant pilots that do not scale.
The competitive window
The next 18 months define the split of the Colombian mid-market financial market. Whoever starts operating agentic AI in specific closed loops captures proprietary data, organizational learning curve, and cost advantage. Whoever waits for full technology maturity later competes against institutions that have already closed that gap.
The question is not whether your institution uses AI. 73% already do. The question is whether your institution uses AI that decides and executes autonomously or AI that assists and recommends. The difference is not marketing. It is business model.