Joseph Enochs, Chief AI Officer at Enterprise Vision Technologies, joins Angelina Yang on TwoSetAI for a 66-minute workshop on what large companies are actually doing with AI — from board-level risk and growth decisions to the infrastructure, governance and operating loops required in production.
The boardroom sees both leverage and existential risk AI sits on both sides of a board’s mandate: it can create products and efficiency, but an internal agent or AI-accelerated attacker can also damage the business. Workforce conversations are real, though Enochs prefers “flattening the hiring curve” to simplistic replacement claims: equip the existing workforce to support more growth without hiring at the old rate. Cost reduction is not free. Microsoft’s reported $500 million customer-service saving came alongside billions in AI investment. He invokes Jevons paradox through radiology: cheaper, more available capability can expand demand rather than simply erase the profession. Measure the baton, not everyone’s legs A large enterprise is a relay race. Business owners, builders, platforms and operations each add value and pass work downstream. AI activity is easy to display — another agent, MCP server, prototype or internal tool — while end-to-end business movement remains unchanged. The useful metric is how quickly and reliably the baton reaches the business outcome, not how fast every department’s legs appear to move. Local optimization has a low ceiling. The larger gains require redesigning the full value stream across organizational boundaries. Backyard patios do not make a skyscraper Scattered pilots are “backyard patios”: useful for learning, but duplicated, disconnected and unable to support an enterprise-wide system. Trying to stack those experiments later produces a “leaning tower of AI” rather than a skyscraper. The shared foundation needs identity, access control, vulnerability management, code scanning, telemetry, policies and a repeatable path to production. Shadow AI makes the debt worse: shadow IT may have involved a minority of employees swiping cards for cloud tools; generative AI can reach 80–90% of a workforce. Bounded autonomy is the production model Enterprise agents need policy-constrained execution: real authority inside explicit permissions, data boundaries and approval rules. This is more than a generic human-in-the-loop checkbox. Keys, duties and data access must be separated so an agent cannot perform prohibited actions. In patch management, an agent can identify exposed systems, test a fix in sandbox, promote it through non-production environments and present a repeatedly validated production change to the human approval board. The goal is not unrestricted autonomy. It is progressively earned autonomy with evidence at every stage. Vibe coding belongs upstream of a governed pipeline Vibe coding is valuable for discovering whether an idea and its business value are real. Production still requires CI/CD, policy as code, security scans, code-quality checks, identity and observability. Enochs’s ratio is deliberately lopsided: roughly 80% of the effort should be the validation, security and governance harness; 20% the generated application. Trust shifts review away from inspecting every generated line toward validating the systems that constrain and test it. Build a digital nervous system Ordinary RAG supplies document context; an enterprise nervous system needs three connected layers: Knowledge graph: what the organization knows. Context graph: what is relevant and true now. Procedural or execution graph: what actions may happen next, under which policies. That stack turns AI from a question-answering surface into a system that can prepare and execute work before presenting a decision to a person. Enochs expects ontology and graph infrastructure to become more accessible to practitioners over the next 12–18 months, rather than remaining a bespoke black box. Start at the outcome and work backward Define the production outcome, break it into required sub-outcomes, then map the knowledge, context, systems and handoffs that support each one. Evaluate experiments by service improvement, business impact, achievable efficiency and difficulty of reaching production — not novelty. Brownfield systems carry the current revenue and regulatory pathway. Modernizing that 80% can fund the more exciting greenfield 20%. A practical project proves every step and handoff, rather than merely demonstrating an impressive final response. Cross-functional training can expose real value Enochs favors “vibathons” that teach teams safe AI development and then bring different departments together around a business problem. In one anonymized utility example, operations and billing teams combined datasets that a consultant had priced at $250,000 just to explore. They identified roughly $25 million per quarter in produced power that was not tied to customer accounts. Neither department could have found it by optimizing inside its own silo. The result came from joining the people who understood two halves of the same value stream. Progress means shorter, safer value streams Measure request-to-completion time, rework, customer-service quality and the time required to remediate a vulnerability across the enterprise. Do not mistake token use, prototype count, headcount reduction or visible busyness for business progress. Managers need to surface work across teams, help contributors find consequential problems and judge a portfolio of experiments more like investors. Leadership becomes more technical: people and financial skills remain necessary, but leaders must understand the tools that can amplify their teams. “Everyone’s legs are moving fast, but what they should really be measuring is how the baton is moving throughout the organization in alignment with the business goals.”
...