Agentic AI represents the next frontier in automated intelligence. Unlike traditional RPA (robotic process automation) or traditional GenAI, designed to answer a single prompt, Agentic AI can perform complex, multi-step tasks and automate entire workflows, making decisions across different systems. The potential is immense: Gartner forecasts that by 2028, a colossal 15% of all day-to-day decisions will be made autonomously by AI agents. This technology promises to drive efficiency, cut costs, and free up IT teams for strategic work.
But before VCs can confidently invest in this future, we must address why so many AI initiatives stall, and which structural challenges remain.
The GenAI Reality Check: The 95% Failure Rate
The most immediate challenge facing agent deployment is the painful reality of pilot failure. According to a recently published MIT study, the State of AI in Business 2025, 95% of GenAI pilots fail to cross the crucial “pilot-to-production cliff”. MIT calls this the “GenAI Divide”. The research suggests the core reason for failure is that companies try to eliminate all “friction”.
Friction, in this context, refers to the resistance: whether human, organisational, or technical, that systems encounter in the real world. Pilots that “glide frictionless from demo to deployment” collapse the moment they hit real-world organisational texture, compliance issues, politics, and data quality problems.
The 5% of winners, conversely, succeed because they design for friction, viewing resistance as the input necessary for learning and adaptation. They build systems that incorporate learning loops, retain context, and integrate deeply into high-value workflows.
The Main Barriers Still Standing for Agentic AI
It’s important to start by saying that a lot of challenges have been solved or are in the process of being solved. Whether it is Agent to Agent communication, the rise of agentic computer use (or autonomous browsing), some aspects of security, etc. But several critical areas remain friction points that slow mainstream enterprise adoption:
1. Reliability and Predictability
Agentic AI relies on autonomous decision-making. While autonomy offers efficiency and scalability, it also introduces unpredictability, as the agent might take a less predictable path to the solution. The technology cannot be useful unless it can be trusted to take reliable action. Similar to the early stages of LLM-based generative AI, achieving reliability requires significant investment, including quality control initiatives like human feedback loops, to make systems more consistent.
2. Data Quality and Relevance
For agents to succeed and avoid errors, they must draw on accurate, relevant, and timely data, ideally in real-time. Many AI models struggle with this pipeline. Solutions require specialised infrastructure, such as Data Streaming Platforms (DSPs), Apache Kafka, and Apache Flink, to collect, process, and transmit data from various sources in real-time.
This is a tricky one, but as an example of the steps being taken to address this, just this week Snowflake, Salesforce, dbt Labs and more than a dozen other technology companies announced they will create a universal standard for how business data is defined and shared across platforms, solving what executives call AI’s most fundamental bottleneck.
3. Enhancing Model Reasoning and Insight
Effective Agentic AI systems use multiple interacting agents—such as a “planner” to set a course of action and “critical thinkers” to assess and adjust—to create a continuous feedback loop. This sophisticated decision-making requires the underlying models to be trained on realistic, high-quality data that reflects real-world complexities. This process demands continuous iterations, sometimes involving thousands of scenarios, before the model can reliably make critical decisions.
4. Security, Privacy, and Legacy Integration
AI agents interact with multiple systems and databases, potentially accessing sensitive data and introducing vulnerabilities like data leaks or malicious injections. Companies must address this by isolating data, implementing robust segmentation protocols, and anonymizing sensitive information before it reaches the model.
Furthermore, integrating agents is challenging because many enterprises rely on legacy infrastructure that is rigid, making it difficult for autonomous agents to orchestrate processes. Overcoming this requires platform modernization, API-driven integration, and process re-engineering.
5. Governance, Risk, and Regulatory Gaps
The delegation of decision-making to AI occurs at a time when regulatory frameworks specific to agentic AI do not yet exist. Gaps remain concerning whether an AI can legally initiate a payment or give financial advice. Addressing these issues requires internal governance models and safeguards for human-AI collaboration. Integration with legacy systems and addressing risk and compliance concerns were the primary challenges cited by surveyed AI leaders concerning agentic AI adoption.
The Barrier That Is Rapidly Falling: Agentic Payments
While complex technical and governance barriers persist, one structural challenge necessary for agent adoption: the ability for agents to securely transact commerce, is rapidly being solved through industry collaboration and open protocols.
The emerging field of Agentic Commerce is slated to hit a total addressable market of $1.7 trillion by 2030. This shift requires a transaction layer that allows agents to pay one another, enforce contracts, and settle value.
This week Google announced the Agent Payments Protocol (AP2), an open protocol developed in collaboration with over 60 organisations, including major financial players like Adyen, American Express, Mastercard, PayPal, and Worldpay.
AP2 was created because existing payment systems assume a human is directly clicking “buy,” an assumption broken by autonomous agents. The protocol addresses three critical questions raised by agent-initiated payments:
- Authorization: Proving the user specifically permitted the agent to make a purchase.
- Authenticity: Ensuring the merchant can trust that the agent’s request reflects the user’s true intent.
- Accountability: Determining who is responsible if a fraudulent or incorrect transaction occurs.
AP2 establishes trust using Mandates – tamper-proof, cryptographically-signed digital contracts. The process typically involves an Intent Mandate (the initial request, creating auditable context) followed by a Cart Mandate (the final approval for specific items and price). This sequence creates a non-repudiable audit trail, providing the clear foundation financial institutions need to manage risk.
The commitment from major incumbents (Mastercard, American Express, PayPal, Stripe, Visa) to build tools and support protocols like AP2 demonstrates that the necessary infrastructure for secure, authenticated, and accountable agent-led payments is solidifying, significantly lowering this critical barrier to agentic commerce.
That being said, fintech is still a tough nut to crack when it comes to AI Agents autonomously doing transactions, and as Laura Salesse from Eight Roads mentioned in her post, several aspect of fintech will remain structurally hard.
Conclusion for VCs: Invest in Durable AI
The lesson from the 95% of GenAI pilots that fail is clear: success requires embracing friction as a design input, not a flaw. For Agentic AI to move from demos to durable, scalable enterprise systems, VCs must prioritize companies tackling the remaining hard problems: improving reliability, ensuring high-quality real-time data flow, and navigating complex governance issues.
While the foundation for agentic payments is rapidly being standardized by open protocols like AP2, the investment focus must shift to enabling agents to operate reliably and securely within high-value, complex, and often restrictive enterprise workflows. The firms that build systems designed for learning, context retention, and accountability will be the ones that successfully cross the GenAI Divide and capture real ROI.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
