Amgen, Salesforce, and Thomson Reuters explain why AI pilots die after rollout

In corporate AI, the pattern is getting depressingly familiar: an AI pilot looks great, the business approves a broader rollout, and then reality shows up. Either the system stops working properly, or it fails to deliver the business results everyone expected. Fortune’s roundtable at Fortune Brainstorm Tech this month focused on a blunt truth for executives: the problem is not always the technology. Often, it is the planning, processes, and expectations companies put in place, or fail to put in place, before they scale.

The leaders leading that conversation made one point especially clear: scaling is not the next engineering step after a pilot. It is a governance and operating-model problem. Amgen Chief Technology Officer Sean Bruich warned that pilots invite experimentation, calling it “so easy with a pilot to let a thousand flowers bloom.” That creativity is not inherently bad, he said. But if organizations want pilots to scale, they need “a wide number of ideas, but a very tight governance on which pilots are actually greenlit.” In other words, the organization must be ruthless about which efforts earn runway, even if the early phase is messy and exploratory.

That governance question connects directly to another failure mode the executives highlighted: confusing feature success with business success. Salesforce Chief Customer and Commercial Officer Lashonda Anderson-Williams said too many companies are focused on successful implementation of AI features, “the technological bells of whistles,” rather than the intended outcome. The result is disappointment that looks like a technical failure but behaves like a strategy failure. The AI features may work as designed, yet the new technology does not drive meaningful business results, so the rollout runs into a wall during the phase where stakeholders are finally paying attention to metrics.

For agentic AI in particular, Anderson-Williams said the missing ingredient is often a detailed workflow map. Agentic systems are not just giving an answer; they are taking actions inside a process that involves people, groups, and specific touch points. If companies do not document the workflow, or document it poorly, then they set unrealistic expectations. Anderson-Williams described the common dynamic this way: “When you put AI on top of that, the expectation is you're going to see some magic, and there’s no magic there.” That is a second-order warning for leaders: if you skip workflow documentation, you do not merely risk an imperfect pilot. You risk misalignment, where the business believes the AI will cover gaps that the process itself has never properly defined.

Data access is the other recurring bottleneck in scaling from pilot to full deployment. Fortune’s discussion pointed out why this happens in practice: data is often scattered across different silos inside organizations, and those silos come with different access privileges plus varying privacy and security requirements. When you move from a pilot to broader use, the system usually touches more users, more cases, and more sensitive information. Without planning the data needs up front, this transition can get complicated fast.

Thomson Reuters Chief Data Officer Caitlin Halferty emphasized that executives should map the contours of the AI project and identify potential data requirements early, ideally in discovery. The earlier the organization uncovers those requirements, “the better we’ll be set up for success,” she said. Halferty also tied this to stakeholder alignment around privacy and security. She asked two practical questions that often determine whether an AI project can scale at all: “Is there some element of PII (personally identifiable information) or confidential data that’s going to trigger privacy?” and, if so, the right people must be part of the project. The other question is whether there is a cyber element, in which case security must be onboard. This is where AI programs meet the real-world constraints that regulators, internal policies, and enterprise risk teams enforce.

Amgen’s Bruich broadened the buy-in argument beyond data and security. He said AI initiatives that are transformational will naturally involve leaders across finance, technology, HR, and other groups. The company has to decide not just whether AI helps a narrow team, but whether it produces “an outcome that matters to the enterprise.” That framing matters because it forces leaders to evaluate AI as an organizational capability, not a tool. A rollout that is only “more efficient” for a small group may look good in a pilot and still fail when the enterprise asks whether the investment changed outcomes at scale.

For executives, the strategic stakes are straightforward: the cost of scaling a misplanned AI project is not only technical downtime, it is organizational credibility. A pilot that fails to translate into a measurable business outcome trains stakeholders to treat future AI approvals with skepticism. Meanwhile, an AI program that lacks governance risks endless experimentation without a path to decisions. If you run innovation, operations, or data strategy, these leaders’ shared message is a checklist for survival: tighten governance around greenlights, anchor decisions on business outcomes, document workflows for agentic systems, and map data plus privacy and security needs early so scaling does not become a surprise release of risk.