新兴技术研究自主AI：向自主系统的演进（下）： 从输出到结果

EMERGING TECH RESEARCHAgentic AI: The Evolution to Institutional Research Group Autonomous Systems: Part II Dimitri ZabelinSenior Research Analyst,AI and Cybersecuritydimitri.zabelin@pitchbook.com pbinstitutionalresearch@pitchbook.com From outputs to outcomes Published on April 17, 2026 PitchBook is a Morningstar company providing the most comprehensive, mostaccurate, and hard-to-find data for professionals doing business in the private markets. Contents Key takeaways •Agentic AI adoption is concentrated in structured, high-volume workflowswith measurable outcomes, particularly across IT, customer support, andgo-to-market functions. •Governance, integration, and organizational readiness are the primary constraints,with trust and policy frameworks limiting broader enterprise autonomy. •Durable value is shifting toward workflow ownership, proprietary context, andplatform-level control as models commoditize and systems become the primarysource of differentiation. Introduction As detailed inPart Iof this series, agentic AI has entered a deployment phase, markedby a structural shift from assistive tools to systems that execute complete workflows.This transition is driving a concentration of value around orchestration and workflowownership, while reinforcing a divide between platform-oriented companies andapplication-layer solutions. Differentiation is increasingly defined by integration acrossenterprise systems and the ability to deliver consistent, outcome-based performance. Industry insights To bring an industry perspective into this analysis, PitchBook conducted interviewswith startups across the agentic AI landscape to assess how these dynamics areshaping product development, deployment, and competitive positioning in practice. HiddenLayer Chris Sestito, Co-founder & CEO of HiddenLayer Within cybersecurity and machine learning operations broadly, which functionsare closest to meaningful agentic AI adoption, and which remain structurallyconstrained? What is driving that divergence? The areas getting closest to real agentic AI adoption are security operations,especially SOC workflows. A lot of effort is going into bringing agentic systems intoalert triage and investigation. The core reason is scale. Even with more people entering security operations, thevolume of alerts is already beyond what humans can realistically handle. It’s anattention problem. Big cloud providers are already using automated agents to triagealerts because the alternative simply doesn’t scale. What makes this a strong fit for agentic AI is that triage is largely a mapping problem.You’re taking signals from an alert and routing them based on context. In many casestoday, agentic systems are already matching or even outperforming human analysts inaccuracy, so you’re not sacrificing efficacy by automating. By contrast, areas that require more judgment, coordination, or accountability still faceconstraints, which is where adoption slows down. Where can agentic systems move beyond monitoring and threat detection intoproactive multi-step mitigation and automated security enforcement workflows, andwhat technical, trust, or governance barriers still limit broader autonomy in practice? Detection and response is the clear starting point. Whether it’s network, endpoint, orAI systems, agentic workflows can compress what used to take days into seconds.They can detect an attack, evaluate it, and respond based on predefined thresholdsand behaviors. The key is customization. These systems need to be configured to behave differentlydepending on context, risk tolerance, and the type of threat. That flexibility is whatenables more proactive, multi-step responses. Where it gets interesting is the distinction between using AI for security versussecuring AI itself. Companies like Vectra AI or 7AI are using AI to secure data andinfrastructure. HiddenLayer is focused on securing the AI models themselves. That’s adifferent problem set entirely. What’s actually limiting broader deployment isn’t really technical. It’s organizationaland regulatory. Federal regulation hasn’t caught up to agentic AI, and there’s no cleargovernance framework yet. Some companies are moving aggressively, but many arestill cautious. So the bottleneck is less about capability and more about trust. Enterprises arehesitant to fully hand over control, even if the technology is ready. In your experience, what has proven hardest for competitors to replicate in buildingagentic model security systems at scale? The biggest differentiator is deep R&D investment. There’s a very limited global pool of true adversarial AI experts, probably in thehundreds, not thousands. HiddenLayer has concentrated heavily in that area and putsthe majority of its capital into R&D. That level of investment is what allows them to stay ahead in a space that’s evolvingextremely fast. New threat categories like prompt injection are emerging rapidly, anddefending against them requires constan