The conversation around artificial intelligence in the enterprise is rapidly evolving. While generative AI (GenAI) has captured imaginations with its ability to create text or summarize communications, many business leaders are realizing that these applications often deliver only localized productivity gains. They rarely impact an organization’s core cost structure or revenue streams in a truly transformative way.
Today, the focus is shifting. Enterprises are keenly interested in deploying systems capable of independent execution—what many are calling autonomous intelligence. This means applications that can navigate internal networks, carry out complex multi-step logic, and even finalize transactions without constant human intervention.
The Shift Towards Autonomous Intelligence
According to Prakul Sharma, Principal and AI & Insights Practice Leader at Deloitte Consulting LLP, this represents the third stage on an intelligence maturity curve. He describes a progression from ‘assisted intelligence,’ where AI helps people interpret information, through ‘artificial intelligence,’ where machine learning augments human decisions, to ‘autonomous intelligence,’ where AI decides and executes within defined boundaries.
Sharma notes that the GenAI abilities we see today, like chatbots and conversational AI, sit in the middle of this curve. “Agentic AI acts as the bridge into autonomy,” he explains, “and it is where the center of gravity is changing now.” The critical difference lies in agency: while GenAI produces an answer, autonomous intelligence actively pursues an outcome by reasoning over a goal, utilizing tools and data, and adapting as conditions change, all while humans set guardrails.
The true unlock for industries isn’t just the agent itself, but the robust governance architecture surrounding it. This includes strong identity management and crucial human-in-the-loop checkpoints, which are essential for safely scaling autonomy across the enterprise.
Unlocking Real Economic Value
To extract genuine economic value, these autonomous systems must be deeply integrated into workflows that directly generate revenue or carry significant costs. Consider an agentic application in enterprise procurement, for instance. Such a system could continuously cross-reference supply chain inventory against live vendor pricing within an enterprise resource planning (ERP) system.
It could then independently authorize purchase orders within predefined financial limits, pausing only for human approval when deviations occur. This kind of system demands a verifiable identity within the ERP, access to real-time pricing data that is legally binding, and operational thresholds formally endorsed by legal and compliance teams. Without addressing these critical dependencies, the case for autonomous execution quickly collapses.
Deloitte initiates this operational overhaul by advising clients to conduct a “decision audit.” Sharma recommends focusing on one or two value chains where outcomes are bottlenecked by decisions, not just tasks, and then meticulously mapping how those decisions are currently made.
This process involves asking key questions: “Who has the data, who has the authority, where the handoffs break, what actions are needed, and where judgment is being applied?” These insights reveal where autonomy can create significant economic value, while simultaneously exposing data and governance gaps that could otherwise derail a pilot project. From there, Deloitte helps leaders sequence the rewire, establishing foundational layers and using successful initial deployments as templates for wider scaling.
Navigating the Technical Landscape: Data is Key
While foundation models from major providers have rapidly advanced to handle complex reasoning tasks, the real friction often arises upstream. Sharma points out that the model itself is rarely the bottleneck, as frontier capabilities are fast becoming commoditized. Enterprises frequently stumble in the design phase by selecting a use case before thoroughly mapping the underlying workflow, which can result in automating an already broken or poorly instrumented process.
Another common pitfall is underestimating the need for decision-grade data, as opposed to mere reporting-grade data. Most enterprise data estates were built for human analysts, meaning data is often aggregated, structured for dashboards, and stripped of the lineage that explains how a value was derived. This is adequate when a person applies judgment before acting on it.
An autonomous agent, however, lacks this human backstop. When it retrieves a contract price or a stock level to execute a transaction, that figure must carry a current timestamp, traceable provenance, and access controls confirming the agent is authorized to read and act on it. Providing this decision-grade data involves integrating autonomous agents with appropriate event stores and databases designed for both structured and unstructured enterprise information, ensuring data freshness and integrity.
Scaling Autonomy: Beyond the Pilot Phase
Transitioning from controlled testing environments to live enterprise deployment presents distinct challenges. One significant hurdle is the financial model for scaling these systems. Agentic workflows often involve multiple interactions with large language models to achieve a single goal, leading to potentially unpredictable API costs. Mitigation strategies like retrieval-augmented generation (RAG) processes, while crucial for reducing hallucination risks, also increase compute overhead, necessitating strict financial controls before enterprise-wide deployment.
Another critical area is security and governance. A small-scale test might perform perfectly with carefully curated datasets, but deploying that capability across thousands of employees and interconnected software platforms exposes vulnerabilities. Integrating the agentic architecture deeply with existing identity providers and cloud-native security controls across hybrid cloud ecosystems is paramount.
Sharma highlights what he terms the “production gap.” He explains that a pilot can succeed with clever prompts and curated data, but enterprise deployment demands continuous evaluations, identity and authorization that function across all systems, effective change management for users, and a financial model that can absorb use-based costs at scale. Furthermore, “governance debt”—where controls and risk frameworks waived for a pilot become roadblocks during a production rollout—is a common issue.
Organizations that succeed treat pilots not as isolated experiments, but as the initial production instance of a reusable platform. This means incorporating identity models, continuous evaluations, and robust governance from the very beginning. By doing so, subsequent use cases can build upon a solid, secure foundation, avoiding the need to rebuild essential infrastructure repeatedly.
Source: AI News
