The Reality of Agentic AI Today
Elsa Mou, Tensility collaborator
Wayne Boulais and Armando Pauker, Managing Directors, Tensility Venture Partners
Agentic AI is one of the most talked about areas of artificial intelligence, promising fully autonomous, AI-driven systems capable of managing complex workflows. We set out to understand the reality versus the hype by hearing directly from founders at the forefront of building with agentic AI to get a clearer perspective on working use cases today. While progress is happening, most so-called agentic AI implementations remain in an early, experimental phase.
Here’s what we observed: Startups are beginning to implement agentic workflows because of the efficiencies that may be derived from this technology, but we are still in the early days with fairly simple working use cases. We identified several key trends today:
Chatbots remain the dominant use case
New use cases are emerging in supply chain applications
Accuracy and quality are required for high stakes applications
Security concerns and accuracy issues limit enterprise adoption
Challenges remain in realizing autonomous multi-agent systems (MAS)
… however, new agentic technologies are coming on line.
Chatbots: Today’s most common form of Agentic AI
Much of what is marketed as agentic AI today functions more like advanced chatbots than truly autonomous systems. We heard several times that even IT leaders struggle to define what makes AI 'agentic' versus just a chatbot.
Today’s AI assistants generally fall into three categories per the graphic below.
Many AI-powered assistants marketed as “agents” today are more akin to interactive search tools that retrieve and repackage data but lack independent decision-making. The value here is to save time on certain monotonous, pre-defined workflows when implemented properly. Even tools that employ chain-of-thought reasoning are limited to the simulation of logical steps without genuine adaptability. True autonomy, where AI makes decisions and executes tasks with minimal human oversight, remains aspirational.
Digital Workers: Emerging Use Cases in Supply Chain
IT bandwidth and support is a common adoption hurdle for new application vendors. Many enterprises still rely on legacy systems with limited API support, requiring IT teams to build custom connectors—an expensive and slow process.
Andrew Stroup, founder of Leverage, a Tensility company that provides supply chain visibility through AI, encountered this firsthand. "Every time we tried to onboard a new client, IT integration became a bottleneck. Their ERP systems were outdated, and their IT teams were too busy to build API connectors."
Leverage found an agentic solution: instead of integrating to customers’ systems through APIs, they trained AI-powered “digital workers” to interact with software visually, mimicking human users. Powered by models including Anthropic Vision, Leverage’s digital workers can log in, navigate enterprise UIs, and execute tasks just like employees—without requiring deep system integration.
Leverage found this approach particularly impactful in mid-market manufacturing, where companies often operate fragmented and dated tech stacks. Traditional API-based automation is impractical for businesses running different and legacy versions of ERP software, such as SAP, Oracle, or Microsoft. By deploying AI “workers” that perceive and navigate ERP interfaces like humans, Leverage has reduced customer onboarding time from months to days, accelerating revenue realization.
Four Kites, an AI driven supply chain visibility company backed by Tensility, recently introduced an agent-based Digital Workforce centered on supply chain activities. The first two agentic “workers” are Tracy and Sam (see below).
These agents, including more to come, coordinate across multiple ERP modules, integrating order management, transportation tracking, and inventory control. As an example, the FourKites system can query transportation management data, cross-reference shipment data with email communications, and flag potential delays before they become major issues.
The agentic approaches described above allows for more precise, real-time supply chain visibility without needing massive IT time and investments for integration.
Security and Compliance: The enterprise adoption bottleneck
Security remains a roadblock for broad adoption. Enterprises struggle with data classification, making it difficult to enforce security policies. Yasir Ali, founder of Polymer, an AI data security Tensility portfolio company, mentioned that while AI model providers such as Microsoft guarantee they will not train on user data, customers must specify which data is off-limits—something most enterprise filing structures aren’t built for.
The Polymer architecture in the graphic below shows how they secure data by looking at the inputs (prompt ingestion) and outputs to an agent or LLM to prevent both sensitive data going into training the agents or LLMs and to prevent sensitive data from being leaked. Agents and LLMs need to be monitored at all times for something unplanned which can lead to a data breach.
Polymer addresses this by automating classification, preventing unauthorized access without requiring heavy user input. Their system dynamically understands, categorizes, and applies security policies for agents, LLMs, and any probabilistic system, in real time. Polymer could ensure, for example, that AI agents are provided adequate access without compromising an enterprise’s security and collaborating with any existing GRC (Governance, Risk, Compliance) systems.
Other security measures include federated learning, which trains AI models on-device without transmitting user data, and Microsoft Purview, which automates enterprise access control.
These security tools may be key to accelerating enterprise AI adoption, especially in industries with high security and compliance requirements.
Accuracy & Quality: Why high-stakes workflows need guardrails
AI’s probabilistic nature means it will never achieve 100% accuracy. In low-risk applications, some errors are acceptable. But in high-stakes workflows like billing and compliance, precision is essential. We believe this is why agentic AI offerings to date have been more common in lower-risk areas, such as sales and customer support.
For high value enterprise applications, Matt Elenjickal, CEO of Four Kites, mitigates risk by leveraging enterprise data collected over ten years and integrating his AI inferences with business logic to ensure their AI optimization stays within strict business guardrails. By utilizing both historical data and structured decision-making, Four Kites has made agentic AI work in critical workflows without compromising on accuracy and quality. For example, their AI won’t issue payment approvals or modify HR records, as these fall outside its domain-specific rules. Instead, it focuses on supply chain optimization and ensuring operational visibility. This hybrid approach—blending probabilistic AI with deterministic business rules—helps increase reliability and minimize errors.
The Vision vs. Reality of Multi-Agent Systems
Multi-Agent Systems (MAS) represents the long-term goal for agentic AI. From our previous blog post AI Interactivity (Part II): Multi-agent Systems and CoPilots: these “are advanced computational constructs where AI agents, working autonomously and independently, collaborate to complete tasks and achieve a common goal.” Fully autonomous MAS is an open research challenge, requiring more real-world implementations to uncover and solve unforeseen issues.
True MAS still requires significant human oversight, struggles with complex reasoning, and relies on rigid communication protocols that limit scalability and adaptability (LLM Multi-Agent Systems: Challenges and Open Problems, 2024; A Survey on Context-Aware Multi-Agent Systems, 2024). A key gap was found to lie in inefficient multi-agent learning, where current reinforcement learning methods face computational bottlenecks that hinder scaled deployment (Towards Efficient Multi-Agent Learning Systems, 2023).
New Directions in Agentic Technology
DataRobot’s recent acquisition of Tensility’s portfolio company, Agnostiq, and its open-source distributed computing platform, Covalent, highlights the push to advance agentic AI adoption in business environments. This move enables businesses to deploy and manage agentic AI applications more flexibly across various computing infrastructures, and addresses key scaling challenges organizations have faced.
Anthropic just announced Claude Sonnet 3.7 to push the boundaries of agentic AI with a coding tool that specifically supports agentic coding. The release provides support for error correction and better reasoning for use by customer facing agents (chatbots). Tone and nuance are more clearly understood in chatbot applications making for a better end-customer experience.
Conclusion: Hype vs. Reality - Where are we today?
The vision for agentic AI is ambitious and impactful, but we found today’s reality to be a nascent one. While meaningful progress is being made to identify where an agent can be embodied to overcome IT friction, for example, the security and trust concerns and true autonomy are still a work in progress.
For now, the most successful AI agents enhance workflows rather than replace human decision-making. As reliability and scalability challenges are addressed, and as real-world implementations push for continued innovation, we are optimistic that we will see agentic AI evolve into a useful method to improve workflow efficiency.