Enterprise delivery remains constrained by human-paced approvals, even as AI greatly increases the supply of deployable code.
2
AI projects need portfolio funding and hypothesis-driven delivery because their scope, value, and cost are often learned through experimentation.
3
Trust should grow through shadow mode, advisory mode, and controlled autonomy, while customer feedback becomes the product's living memory.
Summary
Jess Grogan-Avignon and Jack Wang argue that enterprise agentic projects usually slow down for organizational reasons rather than coding problems. Large companies built governance, security, data, and deployment processes for human-paced work. AI now produces code and possible applications faster than those approval systems can handle. They describe five tensions that determine whether an agentic project can succeed: speed, value, delivery, trust, and moat. Human approval processes need to become executable code. Finance should fund a portfolio of experiments instead of requiring certainty from every project. Teams should use short hypothesis-driven loops and measure growing statistical confidence. Agents should gain autonomy gradually, starting in shadow mode and moving through advisory mode to controlled autonomy. Finally, a company's lasting advantage comes from customer signals, corrections, edge cases, and behavior that accumulate after launch. The talk is direct about the organizational work required after an application has been built.
Enterprise governance was built for human speed, so AI exposes its limits
Large enterprises such as telecoms, utilities, government, healthcare, and consumer products have layered controls because mistakes can affect critical infrastructure or entire populations. Those structures worked when delivery moved at human pace. AI changes the pace without automatically changing data access, security reviews, deployment processes, or stakeholder approvals. The speakers describe an application that took about two weeks to build but another 12 months to reach production because infrastructure, security, the AI gateway, data governance, and application teams all had to align. Faster coding therefore creates a bottleneck in review and deployment.
Giving developers an AI coding agent can increase the amount of code waiting to ship, while approval and deployment infrastructure stays unchanged. The speakers cite GitHub's reported average of 275 million commits per week in 2025, with a projected 14 billion by year end. They argue that the deeper technical debt is years of underinvestment in engineering automation, including CI/CD. Each human process that controls delivery should become adaptable, executable code. AI can help build those capabilities faster and more cheaply, but organizations have to change their operating assumptions rather than add another sign-off meeting.
Finance should fund a portfolio of uncertain AI bets
Traditional business cases assume that scope, expected value, cost, and delivery time can be known before work begins. AI often reverses that order because teams discover the solution and the business case through building and experimentation. Lower prototyping costs also make new products, services, and customer experiences economically possible. The speakers recommend that CFOs think like venture capitalists: back a portfolio, expect that some bets will fail, and look for the projects that compound. The relevant question becomes the cost and impact of not trying, rather than whether one project can promise a fixed return.
Agentic delivery should build statistical confidence through short loops
The speakers say agentic systems cannot be scoped or milestone-managed like deterministic software features. Model outputs and agent behavior are non-deterministic and emergent, so teams need hypothesis-driven delivery. They recommend small loops of building, evaluating, iterating, and gathering evidence. Delivery teams should include people who can explain what they learned, handle ambiguity, and translate statistical results into stakeholder confidence. Organizations need to hire, train, and reward those skills instead of measuring progress mainly through requirements, features, status updates, or fixed program milestones.
Autonomy should increase only as outcome evidence builds trust
Trust in an agent includes confidence in its content, accuracy, responsible use, and privacy. The speakers describe deployment as an exposure ladder. In shadow mode, an agent runs alongside human work without affecting outcomes, and its recommendations are compared with human decisions. In advisory mode, it runs live but only recommends, while people approve or reject the result. Controlled autonomy allows the agent to trigger actions in narrow, low-risk situations with limits and kill switches. Each move upward should depend on evidence about target outcomes and behaviors, rather than project-plan completion or a simple pass-fail test.
Customer feedback becomes the living memory that competitors cannot copy
Existing CRM, ERP, standard operating procedures, and other enterprise knowledge give a company a starting point, which the speakers call transactional memory. Competitors usually have comparable information, so it is a floor rather than a fortress. The stronger advantage comes from what happens when customers use a product: edge cases, corrections, emotional intent, and actual behavior in a company's specific context. Every feature should either generate a useful feedback signal or deliver something learned from one. The product's launch begins a cycle of turning signals into improvements, and that accumulated living memory is harder to reproduce.
