AI agent deployments are accelerating faster than the enterprise foundations needed to support them, leaving many organizations struggling with the processes, data, costs and controls required to run agents at scale.

The assessment comes from the latest edition of Prompt, AI Business's weekly briefing on the shifting AI landscape, which pairs an analytical look at the week's biggest developments with a curated roundup of the stories that matter.

Agentic AI describes systems that can plan and carry out multi-step tasks with limited human oversight — drawing on enterprise data, calling software tools and triggering actions across business systems — a step beyond chatbots built to answer a single prompt. Enterprises have big plans for AI agents. Whether they are ready for them is another question. A new Deloitte survey found that roughly 75% of U.S. business leaders expect AI agents to reshape about half of their organizations' processes within four years. Yet just 20% believe their companies are currently equipped to redesign workflows for autonomous agents.

The finding lands at a notable moment. Enterprises are deploying more agents, vendors are making them easier to launch by embedding agent-building tools directly into mainstream enterprise software, and AI capabilities continue to advance — but the organizational foundations needed to support them are not moving at the same speed. The result is a widening gap between what agentic AI can do and what enterprises are actually prepared to manage.

Even so, deployment continues to accelerate. Salesforce research found the average number of AI agents across organizations nearly tripled over 15 months, while the time required to create and activate an agent fell 53% to less than two days. Agents are also handling more work: the average number of actions per account grew at a compound monthly rate of 31% over the same 15 months. In other words, enterprises can deploy agents faster than ever — preparing the organization to use them effectively is proving far more complicated.

The biggest barriers sit inside the enterprise. Deloitte found many organizations simply are not ready, with unclear business processes, disconnected data and systems, and resistance to changing established ways of working standing in the way. Such issues predate AI, but they carry more weight here: unlike script-based automation, agents determine their own steps as they go, which makes clearly defined processes and reliable data prerequisites rather than refinements. Those barriers become more consequential as agents take on greater responsibility: without reliable access to data, agents can struggle to produce trustworthy results.

Cost is another concern. Gartner found that agentic AI may not benefit from traditional economies of scale, as more complex reasoning and planning hike inference costs. Because agents can make many model calls as they plan, reason and use tools, spending can climb with the complexity of each task rather than fall as usage grows. Questions about reliability also remain. Just 35% of executives surveyed by HFS Research and TCS said AI consistently delivers business outcomes, earns regulator confidence and provides sufficient control.

Taken together, the developments point to a growing disconnect. Organizations are deploying more agents and asking them to do more, while the processes, data, economics and controls needed to support them are still catching up. For now, the surveys point to organizational readiness, not model capability, as the limiting factor. Agentic AI may be arriving quickly — enterprise readiness is proving much harder to accelerate.

También en las noticias de IA esta semana

• La desaceleración del modelo de OpenAI ofrece a los CIO una lección sobre planificación de IA — La decisión del laboratorio de IA de ralentizar el desarrollo de sus modelos más avanzados destaca por qué los CIO necesitan hojas de ruta de IA flexibles que puedan adaptarse a medida que cambian las capacidades del modelo, la disponibilidad y los plazos de los proveedores.
• Micron destina $10B a la investigación de memoria para IA en EE. UU. — El fabricante de chips está invirtiendo $10 billion en un nuevo centro de investigación en Idaho centrado en avanzar tecnologías de memoria, cómputo y empaquetado para sistemas de IA cada vez más exigentes.
• Nvidia’s SONIC enseña a los humanoides a moverse — El modelo fundacional permite que los robots humanoides aprendan una amplia gama de movimientos de cuerpo completo a partir de demostraciones humanas, avanzando el impulso más amplio de la empresa hacia la IA física.
• Starling Bank agrega habilidades impulsadas por IA a su chatbot orientado al cliente — El banco está ampliando su asistente de IA orientado al cliente con nuevas funciones diseñadas para gestionar consultas bancarias más complejas y brindar un soporte más personalizado.
• Las ganancias de productividad de la IA no se están traduciendo en alivio de la carga de trabajo de TI — La IA está ahorrando a los profesionales de TI horas cada semana, pero la mayoría dice que sus cargas de trabajo se han mantenido igual o han aumentado, ya que nuevas responsabilidades compensan esas ganancias de productividad.
• CISA y FBI advierten sobre una campaña respaldada por IA dirigida a dispositivos Siemens S7 vulnerables — Las agencias advierten que los atacantes están usando técnicas asistidas por IA para apuntar a sistemas de control industrial Siemens vulnerables en sectores de infraestructura crítica, incluidos energía y agua.
• Pony.AI planea 4,000 robotaxis fuera de China — El proveedor chino planea desplegar 4,000 robotaxis a nivel mundial mientras acelera la expansión de su negocio de conducción autónoma.
• Cómo las empresas pueden ponerse al día con el rápido ritmo de los avances en IA — Las empresas que buscan seguir el ritmo de los rápidos avances en IA necesitan una comprensión clara de sus procesos de negocio y la flexibilidad para cambiar entre modelos y agentes a medida que cambian las capacidades y los costos.

Cobertura relacionada: El nuevo modelo abierto de Nvidia es para casos de uso específicos | Meta cambia de rumbo con Muse Glimmer de peso abierto | Marco de Microsoft para reducir los costos de entrenamiento de agentes de IA

Fuente: AI Business | Por Liz Hughes, Contributing Writer | August 21, 2026