AI agents are on the rise, and cash is flowing to them. Over half of companies budget $500,000 or more annually for AI agents. The catch is most firms lack the technology infrastructure stack to deploy those agents effectively.
These are the takeaways from a new survey of 1,045 enterprise technology managers and professionals published by Tray.ai. The survey found that 42% of respondents expect to build or prototype more than 100 AI agents over the coming year, and 36% expect to see more than 100 put into production. The survey, commissioned by Tray.ai, found a similar percentage (41%) expect to address more than 20 distinct business problems.
Also: 25% of enterprises using AI will deploy AI agents by 2025
AI agents are intelligent assistants that can autonomously make decisions and deliver actions. They are intended to serve as digital co-workers, assistants, or customer service representatives, communicating via natural language processing.
<!–>
The projections for AI agents are significant. By the end of 2025, one in four say most of their companies’ core business processes will run with AI agents. Another 41% believe that between 26% and 50% of their core processes will be enabled with AI agents.
Also: 7 ways to write better ChatGPT prompts – and get the results you want faster
The survey also found that more than 86% of professionals report a need to upgrade their existing tech stacks to prepare for AI agents. Widespread integration challenges may hamper AI agent deployments, with 42% of respondents indicating their enterprises need access to eight or more data sources to deploy AI agents successfully.
At least 42% of professionals also said they must make major infrastructural upgrades, such as adding new vendors to their tech stack or staging complete overhauls. Forty-one percent said they are taking a hybrid approach to development with a mix of build or buy solutions, while 24% said they intend to build customer solutions with code.
“While there’s tremendous excitement around AI agents, many enterprises are currently missing essential building blocks to develop and deploy them safely and effectively,” Rich Waldron, CEO of Tray.ai, told ZDNET.
“They’re relying on numerous off-the-shelf SaaS apps that lead to fragmentation and integration challenges, leading CIOs to spend more time managing complexity than driving innovation.” He said greater process automation and attention to unstructured data handling are required.
Also: Why agentic AI is the new electricity, and nearly 80% of business leaders are afraid of the dark
Across the industry, there is agreement that more preparation is needed for agentic AI. Autonomous AI agents require an effective foundational infrastructure and data management practices, said Lan Guan, chief AI officer at Accenture.
“Organizations are at varying stages of readiness. A robust enterprise platform architecture can ensure seamless accessibility to foundation models, including technical infrastructure considerations such as cloud versus on-premises hosting, but also network capabilities, security measures, and the ability to efficiently scale the system as the demand for AI agent-driven capabilities grows.”
Taylor Bird, vice president at Excella, agreed most enterprises are underprepared for truly autonomous AI agents: “While companies have made strides in implementing traditional AI systems, agentic AI represents an additional challenge that requires new approaches to infrastructure, governance, and skill development.”
Infrastructure readiness is a must, Bird said: “Organizations need robust API ecosystems that allow AI agents to safely interact with their software systems. Companies that have silos between their systems will have limited capabilities with autonomous agents.”
Also: Agentic AI is the top strategic technology trend for 2025
Weak security and control frameworks will also hinder AI agent development and deployment. “Agents will bring more branching paths to what a company’s software can accomplish,” Bird said. “Traditional monitoring and safety mechanisms likely only provide coverage for deterministic scenarios. Growing from current state to an agentic system requires a learning curve.”
Keith Pijanowski, AI/ML subject matter expert for MinIO, said the success of AI agents will come down to models. “If businesses want to build effective agents, they need effective models,” he said. “Therefore, everything they need for training models and running them in production is required to be successful with agents. The better your models, the better your agents.”
Ultimately, AI agents will be the key to delivering AI to specific, narrow tasks. “AI agents represent a different way to implement control logic,” Pijanowski said. “Up to now, control logic was coded by developers and does not change at runtime. Using agents, an LLM is first asked to come up with a plan and then act on each step of the plan.”
As AI agents become more mainstream, “the models will become more specialized, which will also drive further adoption,” Pijanowski predicted. “There will be fewer all-purpose models and more smaller models trained to do one thing very well.”
Also: Enterprises are struggling with what to do with Gen AI, say venture capitalists
AI agents “represent a new way to implement control logic,” he continued. “Instead of being hardcoded by an engineer at design time, an agent figures out the course of action when the problem is presented to it. This is scary and exciting at the same time. Scary because without internal verification and guardrails, an agent could create a lot of problems. However, this is a new way to automate. Existing processes could be improved and problems that were too complex in the past may now be solvable.”
The best way to address the challenges of AI agents is awareness and training as AI agents proliferate. “As to how companies can build preparedness, we’ve had success implementing use cases with agents through internal learning events such as hackathons, and we leveraged our lessons learned building other AI/ML solutions,” said Bird.
Data is the fuel that will power AI agents, and much work still needs to be done here. Agentic AI systems must be ready to “leverage multi-modal models capable of processing diverse data types, including unstructured data like images, text, and video,” said Guan. “Effective data integration is essential to ensure that data from disparate systems is readily accessible to agents in real-time for informed decision-making and actions.”
Guan also said curated enterprise knowledge is another key component of AI agents. He urged organizations to pursue a centralized enterprise store that “should incorporate mechanisms for knowledge curation, a semantic layer to define relationships between data elements, and standardized definitions to ensure consistency.”
By capturing and organizing enterprise-specific knowledge, “agents can continuously learn and improve their performance over time,” said Guan. Key prerequisites include “agent API controls, observability and performance tracking, feedback and continuous learning, and model fine-tuning and training.”
However, the increased adoption of AI agents – with more models trained on specialized datasets – will place pressure on data storage, said Pijanowski.
Also: Sticker shock: Are enterprises growing disillusioned with AI?
“Not only do the models themselves need to be versioned and saved somewhere, so do the results of the experiments used to create them – a process known as machine learning operations (MLOps). Models used within agents must be instrumented and the telemetry saved. All this represents a lot more data than a traditional application that uses simple Boolean decision-making.”
What’s clear, concluded Pijanowski, is that the ever-increasing information stored requires more expansive data storage strategies.
