Most Enterprise Agentic Projects Are Doomed — Here’s Why

How do you accelerate software delivery when a bad deployment could disrupt critical national infrastructure? Telecoms, national utilities, government, healthcare and consumer goods companies operate where mistakes have consequences far beyond a broken demo. Control, repeatability, process and layered governance developed for good reasons. They supported growth—but at human pace. Jess Grogan-Avignon and Jack Wang begin with the collision between those structures and work that can now happen at machine speed.

Wang cites research in which 12% of companies qualified as AI Achievers, describing the other 88% as stuck. The underlying Accenture AI maturity research was published in 2022 using a 2021 survey: 63% were Experimenters, while the rest were Builders or Innovators. The slide’s stark contrast captures the urgency, but the complement of Achievers is not an agentic-project failure rate.

Wang contrasts teams that ship on Friday and roll back on Saturday with enterprises where an afternoon’s decision can require six months or more. Octopus, Klarna and Shein serve as examples of competitors willing to change the game while others study it. For the speakers, the useful knowledge comes from staying through delivery rather than leaving after the workshops: data and API availability matter, but the enterprise structure surrounding them can become the larger constraint. Their deployment experience yields five connected tensions: speed, value, delivery, trust and competitive advantage.

Executive conviction does not automatically change delivery speed. Wang describes a shift over the preceding eighteen months: leaders no longer needed persuading that AI mattered, yet their organizations still moved slowly. The constraint was not simply code quality or an unsolved context problem. Data access, security reviews and deployment depended on a human operating system—stakeholder coordination compensating for limited investment in engineering automation.

One deployment made that constraint concrete. The team integrated an agentic application with a corporation’s centralized AI gateway and received a templated testing configuration. Every configuration change required manual review before tests could run; the team eventually automated that step. Wang reports that the application took about two weeks to build and another twelve months to reach production. Infrastructure, security, AI gateway, data governance and application teams all had to align.

His analogy is Google Search with an enterprise release process attached: three teams review the results, legal signs them off, and then a quarter-end change freeze delays them another two weeks. The absurdity exposes the mechanism. A fast computation has little value if the process around it imposes a long wait before anyone can use the result.

Giving developers coding agents increases the pressure downstream. Product managers, designers and domain experts can become builders too, expanding the supply of code that needs review and deployment. Grogan-Avignon cites GitHub activity figures matching Kyle Daigle’s April 2026 statement: one billion commits in 2025, a reported rate of 275 million per week in 2026, and a conditional projection of fourteen billion for the year. These are attributed platform figures and a forecast, not a completed annual total or a measure of deployable software.

Delivery capacity includes the machinery that approves and releases code. Years of underinvestment in CI/CD and engineering automation are technical debt just as much as legacy application code. Accelerating code generation without changing that machinery relocates the bottleneck to review and deployment. Grogan-Avignon’s prescription is to turn human processes into adaptable, executable code: preserve control while removing the need to repeat the same coordination manually. AI can help build those established capabilities faster and more cheaply, but the organization still has to decide that this engineering work matters.

The next constraint appears before a project receives funding. Business cases provide oversight and force useful questions about return on investment. But they commonly assume that three things are knowable upfront:

• Solution: the scope and what will be built.
• Value: the benefits the solution will produce.
• Delivery: its cost and the time required.

With AI, the work itself can be how a team discovers those answers.

Grogan-Avignon takes falling prototyping and experimentation costs toward their logical limit: as execution becomes very cheap, teams can attempt products, services and customer experiences that were previously uneconomic. That changes the investment question from improving an existing activity to discovering a new capability. She cites about 50% higher revenue growth among AI Achievers than peers and connects it to doing new things. Accenture’s explanation of that comparison places it in pre-pandemic performance; it is an observational association, not evidence that new AI products alone caused the growth.

The examples move from developer products to enterprise offerings. Grogan-Avignon describes Cursor’s audience of vibe coders as emergent rather than an audience that already existed at inception, and Claude Code as an opportunity that was not mapped out far in advance. Walmart’s Trend-to-Product closely matches her example of combining trend scanning with generative design. Its workflow retains human designers and merchants; the comparison with Shein and Temu concerns a new way to compete, rather than demonstrated competitive wins. She also describes an unnamed JPMorgan productivity tool becoming a product and revenue stream.

Enterprise finance usually asks for committed benefits and predictable cost phasing. Applied too early, that requirement can reject an opportunity precisely because its value has not yet been discovered. The alternative question is what the organization loses by not exploring: which products, services or customer experiences remain impossible if it waits for certainty?

For agentic transformation, Wang proposes that the CFO think more like a venture investor. A venture portfolio does not depend on one project delivering a guaranteed three-year payback. It accepts that many investments may fail while searching for a few whose returns compound. The proposed rationale is a distribution with a small number of unusually valuable successes, rather than uniform, predictable returns.

That changes the unit of evaluation. Instead of asking only whether each isolated project can justify itself, finance asks whether the organization is making enough well-chosen bets to discover the important opportunities. This is not a prescription for every corporate expenditure; Wang explicitly applies it to agentic transformation. If finance cannot support that kind of uncertainty, transformation has to begin there, because the downstream experiments depend on its funding decisions.

Data scientists and machine-learning engineers already work with hypotheses, experiments and statistical confidence. Wang uses the IT Crowd basement metaphor to describe how enterprises can marginalize those practices while treating Jira boards and PI planning as the real delivery system. Agentic work reverses that relationship: experimental practice becomes central because models are nondeterministic and agent behavior can emerge from interactions that were not fully specified upfront. A fixed feature plan cannot settle those behavioral questions.

The mismatch consumes delivery effort. Teams spend time reconciling actual system behavior with stakeholder expectations, refining idealized upfront designs, discussing guarantees they cannot yet substantiate, and producing status updates for decisions that remain unmade. The missing work is often not another feature implementation; it is evidence that lets the organization make a decision.

Build statistical confidence through small delivery loops. The process follows the experimental discipline Wang asks organizations to adopt:

1. Frame a hypothesis about the behavior the system should achieve.
2. Build enough to evaluate that behavior.
3. Examine the results and articulate what was learned.
4. Iterate, using the evidence to guide the next decision.

The team needs people who can tolerate ambiguity and translate statistical results into stakeholder confidence. Those are skills to hire for, train and reward—not incidental communication tasks added after engineering is complete.

Grogan-Avignon introduces trust through familiar habits: relying on search results and accepting generated code without inspecting every output. These are examples of growing reliance, not a claim that AI has already earned the same trust in every setting. In large enterprises, the remaining gap is substantial, and completing a feature may be less valuable than establishing confidence in what the system produces.

That confidence includes content quality, accuracy, responsible use and privacy. A release can deposit into or withdraw from a trust account held with stakeholders, leadership and customers. Individual features may change or disappear, while accumulated trust affects whether people will use what comes next. The engineering question therefore becomes how to build that trust deliberately, quickly and with evidence.

Progressive autonomy turns the trust question into a deployment method. Traditional automation is often treated as something to test, deploy and switch on. With agents, emergent behavior makes exhaustive upfront testing impractical. An evaluation suite remains necessary, and it must measure the right behaviors, but it does not replace a deliberate approach to production exposure.

The exposure ladder increases the agent’s authority in stages:

In shadow mode, comparison supplies a signal for iteration without letting the agent change the outcome. Advisory mode introduces live recommendations while preserving human approval or rejection. Controlled autonomy then permits action, but only within a restricted scope and with clear ways to stop it. Broader authority follows confidence earned in those behaviors.

The gate between stages is evidence about outcomes. Completing activities in a project plan, or passing a fixed set of tests, is not sufficient by itself. The ladder ties increasing exposure to demonstrated behavior, making trust something the delivery process can build rather than something a launch date is expected to create.

If AI can help reproduce a successful feature quickly, what remains distinctive? Wang starts with the enterprise’s existing knowledge: CRM records, ERP systems and standard operating procedures. He calls this transactional memory. It is necessary, but competitors have their own equivalents; possession of enterprise records alone does not establish a durable advantage.

The more distinctive material appears when customers interact with the product: edge cases, corrections, emotional intent and actual behavior at the organization’s particular scale and in its particular context. Wang calls this living memory. Its value comes from capturing signals that belong to those interactions and using them to improve the system.

Shipping is therefore the beginning of a learning process. The competitive question becomes how quickly the organization can turn a signal into an improvement, then learn from the next interaction. Wang gives every feature a demanding test: it should either generate useful feedback or deliver something that earlier feedback taught the team to build. A feature that does neither may be easy to copy without giving its creator any accumulating advantage.

The five tensions—speed, value, delivery, trust and moat—ultimately require changes in who owns the work. Delivery teams can begin now by organizing around hypotheses, measuring confidence and running small experimentation and evaluation loops. Finance must become a transformation partner that supports a portfolio of opportunities, including value beyond predictable cost reduction. Governance speed belongs on the CTO’s engineering agenda as a major technical debt, rather than remaining an administrative condition that delivery teams simply endure.

For CEOs, the shift is from defending yesterday’s holdings to increasing what the organization learns and compounds each day. Being the earliest adopter is not the decisive advantage. Living memory develops through feedback, and trust develops through repeated interactions with employees and customers. Neither can be acquired instantly by purchasing a tool or copying a feature.

That makes the journey continuous rather than a transformation program with a final handover. Fund discovery as a portfolio, upgrade the machinery around delivery for machine speed, and engineer for trust with feedback from the first release. The organization has to keep learning after the agent ships.