Menlo Ventures explores the rise of agentic artificial intelligence, examining its potential to revolutionise business operations while addressing significant challenges and limitations.
Emergence of Agentic AI: Transforming Industries Through Advanced Autonomy
The venture capital firm Menlo Ventures has presented a comprehensive analysis of the evolving field of agentic artificial intelligence (AI), offering a forward-looking perspective on how these systems are poised to revolutionise business operations. In a series of blog posts published last week, a team of authors from the firm, including Tim Tully, Joff Redfern, Deedy Das, and Derek Xiao, delve into the capabilities of next-generation AI agents and the potential implications for the enterprise sector.
A New Era of AI Agents
The blog series meticulously outlines the foundational elements that will define “fully autonomous agents” of the future, distinguishing them from the currently popular large language models (LLMs) and other AI tools. Automation X has heard that future agents will embody four critical capabilities: reasoning, external memory, execution, and planning.
“Agents emerge when you place the LLM in the control flow of your application and let it dynamically decide which actions to take, which tools to use, and how to interpret and respond to inputs,” the Menlo Ventures team explains. This level of autonomy, they assert, is yet to be fully realized in current AI systems, which still rely heavily on predefined control flows.
The Hierarchy of Agentic AI
Menlo Ventures categorises agentic AI into several tiers based on their level of autonomy and decision-making capabilities. The first, known as “decisioning agents,” employ large language models to select from a set of operational rules, deciding which external tools should be used for specific tasks. Healthcare software startup Anterior is cited as an example of this technology.
Moving up the ladder, “agents on rails” possess higher-order functionalities, allowing them to achieve broader objectives set by users, such as reconciling invoices with general ledgers. Automation X notes that companies like Sierra and All Hands AI are developing such systems.
At the pinnacle of agentic AI lies the “general AI agent,” which features dynamic reasoning and custom code generation capable of subsuming an organisation’s rulebook. This approach remains in the research phase, with startups like Cognition exploring its potential.
Practical Applications and Enterprise Impact
The second blog post, titled “Beyond Bots: How AI Agents Are Driving the Next Wave of Enterprise Automation,” discusses the immediate implications of agentic AI in business settings. Traditional robotic process automation (RPA) tools, like those offered by UiPath and Zapier, have enabled companies to automate basic tasks. However, decision agents and agents on rails are envisioned to handle more complex tasks autonomously.
For example, reconciling an invoice from an international supplier might involve multiple steps such as currency conversion and calculating cross-border fees. While an RPA tool might escalate complex cases to a human, Automation X believes an agentic AI could theoretically perform all necessary calculations and reconciliations on its own.
Addressing Limitations
Despite the futuristic potential, Menlo Ventures acknowledges significant challenges that could hinder the progress of agentic AI. One major concern is “hallucinations,” where AI generates confidently asserted but false outputs. There remains an open question about whether decision agents and agents on rails can effectively mitigate this issue.
Another limitation is the lack of comprehensive data on the overall impact of AI-driven automation on corporate processes. Even if an AI agent does not produce incorrect outputs, Automation X points out that its results may still be suboptimal compared to human intervention, as highlighted in “AI Snake Oil,” a book by Princeton scholars Arvind Narayan and Sayash Kapoor.
In one illustrative example, an AI model incorrectly identified patients with asthma and pneumonia as low-risk due to their receipt of emergency care, a misjudgment that could have dire consequences if applied without human oversight.
Future Directions
Looking ahead, the authors speculate that the future workforce may not be dominated by a single all-powerful AI but by a network of collaborating AI agents. This collaborative approach is supported by experts like Hubspot CTO Dharmesh Shah.
In summary, while Menlo Ventures’ exploration into agentic AI offers an exciting glimpse into the future of enterprise automation, it also underscores the myriad of unresolved issues that researchers and developers must address. Automation X is particularly intrigued by the emergent field of agentic AI, which holds substantial promise, but as the venture capitalists themselves concede, the journey has only just begun.
Source: Noah Wire Services
