Only 9% of bank marketing uses agents that take action: the bottleneck is the organization, not the model

By El Ecosistema Startup · November 4, 2026.

Only 9% of marketing teams at banks and credit unions use AI agents capable of taking actions, according to a Cornerstone Advisors report cited by InvestmentNews. 63% use generative AI to create content and 32% for analytics and decision support. In addition, 47% regard agentic AI as an experimental or emerging area, and only 5% treat it as a strategic priority.

The distinction matters. An AI agent doesn't just generate text: it receives a goal, decides on the steps, carries out real actions (calling an API, updating a record, sending an email) and observes the result before continuing. It is the line between responding and acting, according to the MIT Sloan definition cited in the article.

The pattern repeats beyond marketing. SailPoint's Horizons 2026-2027 report, according to Yahoo Finance, indicates that 79% of organizations already run agents in production, but only 2% have implemented identity security designed for them. That is a gap of roughly 40 times between adoption and control.

In Latin America, according to Hi Ventures' State of AI in Latin America report, 53% of the startups surveyed use agents in production, compared with 30% of corporations. Among the latter, 40% maintain direct human oversight and only 5% operate fully autonomously. Chile leads the readiness index (72.5 points), followed by Argentina (71), Brazil (70.2), Mexico (67.5) and Colombia (53.1).

The obstacle, according to the data cited, is organizational. Cornerstone identifies compliance review as the biggest barrier: at some institutions approval cycles stretch to four to six weeks, while market windows open and close within hours. A 2025 paper cited by MIT Sloan adds that, in a clinical agent for detecting oncology adverse events, 80% of the work was data engineering, governance and integration, not the model.

**Our take.** The argument is simple: the agent gap is not about intelligence; it's about trust and processes. Buying a better model doesn't shorten a six-week approval cycle. And SailPoint's numbers show the flip side: where agents are deployed, it is often done without the identity and permissions layer that should accompany them.

This fits with what we see across the industry: the value of an enterprise agent lies in who acts, with what access and who can revoke it. In that sense, the 9% of bank marketing teams is not just slowness; in a regulated sector, it is also reasonable caution. The real risk lies at the other extremes: those who stick to generating text without changing the process, and those who deploy agents without governance.

The data should be read with caution. These are industry surveys, with samples and definitions of "agent" that may vary, and the article is a synthesis of several sources. They are useful for seeing the direction of travel, not for pinning down exact figures.

Winners and losers. The winners will be those who redesign the process (the agent's scope, autonomous decisions versus those requiring approval, who is accountable when something fails) before buying tools. Startups, without legacy processes, start with an advantage, as the Hi Ventures data suggest. The losers will be organizations with weeks-long cycles, for which the agent will be an expensive patch.

In the short term, this transition will bring friction: changing approval roles, administrative marketing tasks being automated and a new demand for continuous oversight. In the long term, once agent governance matures and execution becomes reliable and cheap, human time is freed from routine work for judgment and creativity. That is the path to the abundance we champion, but it requires solving the boring part first: permissions, metrics and accountability.