Welcome to not just a world of AI agents but a multi-agent world. Yet while these functional generative AI (Gen AI) tools have great promise for personal and professional productivity, deploying them is a significant challenge for designers and developers.
The authors of a recent Deloitte report suggested agents have caught people’s attention — 26% of organizations are exploring autonomous agent development. At least 52% of executives are interested in pursuing agentic AI development, and 45% want to extend development to multi-agent systems. However, while agentic AI will be a key enabler of sustainable value, the report suggested it’s no silver bullet.
Also: Autonomous businesses will be powered by AI agents
“As software systems that can meet objectives with minimal intervention, agents could help accelerate the creation of long-lasting business value,” the report stated.
“However, the key barriers currently faced by Gen AI — regulatory uncertainty, risk management, data deficiencies, and workforce issues — still apply, and are arguably even more important due to the increased complexity of agentic systems.”
Also: AI agents might be the new workforce, but they still need a manager
Unlike today’s bots, “which mainly respond to input, agentic AI can be capable of planning ahead, prioritizing tasks, and executing complex workflows with minimal human intervention,” Jim Rowan, head of AI at Deloitte Consulting, told ZDNET.
However, overall, “implementing AI agents can be costly,” he cautioned. Rowan said data infrastructure is vital for any AI agent initiative: “These necessary systems include scalable cloud platforms, advanced data analytics tools, and robust cybersecurity measures.”
<!–>
Getting started with agents
The best bet for percolating AI agents throughout the organization is to keep things as simple as possible. “Companies and employees that have already found ways to operationalize intelligent agents for simple tasks are best placed to exploit the next wave with agentic AI,” said Benjamin Lee, professor of computer and information science at the University of Pennsylvania.
“These employees would already be engaging generative AI for simple tasks and they would be manually breaking complex tasks into simpler tasks for the AI. Such employees would already be seeing productivity gains from using generative AI for these simple tasks.”
Also: 15 ways AI saved me time at work in 2024 – and how I plan to use it in 2025
Rowan agreed that enterprises should adopt a crawl, walk, run approach: “Begin with a pilot program to explore the potential of multiagent systems in a controlled, measurable environment.”
“Most people say AI is at the toddler stage, whereas agentic AI is like a tween,” said Ben Sapp, global practice lead of intelligence at Digital.ai. “It’s functional and knows how to execute certain functions.”
Enterprises and their technology teams “should socialize the use of generative AI for simple tasks within their organizations,” Lee continued. “They should have strategies for breaking complex tasks into simpler ones so that, when intelligent agents become a reality, the sources of productivity gains are transparent, easily understood, and trusted.”
Rowan suggested embracing smaller language models rather than the large language models that have dominated the Gen AI scene up to this point: “These systems will bring significant value across a range of roles, from supply chain management to software development and financial analysis.”
Also: 5 tips for choosing the right AI model for your business
Lee agreed this use of smaller-scale models would serve agentic AI well. “The intelligent agent would break complex tasks into multiple simpler tasks, possibly querying multiple types of specialized models for these tasks,” he said. “The agent would then combine these intermediate results into a coherent response.”
Refining the agentic approach
Quality data is also key, Rowan added: “It’s the foundation for AI agents to work effectively. If data is inaccurate, incomplete, or inconsistent, the agents’ outputs and actions may be unreliable or incorrect, creating both adoption and risk issues. It’s therefore essential to invest in robust data management and knowledge modeling.”
Rowan also urged comprehensive investment in workforce upskilling. This training should “focus on technical skills and the ability to collaborate effectively with AI agents,” he said. “A well-prepared workforce is key to realizing the full potential of AI agents.”
Lastly, it is essential “to establish processes for continuously monitoring and improving the performance of AI agents,” said Rowan. “This includes collecting and analyzing performance data, identifying improvements, and making changes to optimize their performance.”
Also: Agents are the ‘third wave’ of the AI revolution
Along with technical considerations, the key to agentic AI design and deployment success is for companies “to consider their policies around agentic AI,” Sapp said.
For example: “Who’s determining when it can be used? Does it have permission to interact with other agents? Agentic AI is moving around, it’s in motion and talking to other systems. What happens when those systems run into each other or disagree? A hierarchy is needed to determine where the fine line of auto-approval should reside.”
Sapp gave the example of a large financial services company that employed an AI model “to predict whether or not a change is going to fail. That information creates a probability of failure that goes to a human,” he said.
“Based on this probability, that person can then decide to review it deeper or go ahead and approve it. Agentic AI can review that exact change in a workflow and automatically approve the change based on a failure probability rate below 1%. It no longer has to go to a person; it becomes an automated action versus leveraging humans to take action based on AI data.”
