AI Doesn't Have a Pilot Problem. It Has a Decision Problem.
Almost every company is now doing something with AI. Far fewer can point to what it has done for the business.
McKinsey's 2025 State of AI survey puts numbers on that gap. 88% of organizations now use AI regularly in at least one function, yet nearly two-thirds have not started scaling it across the enterprise, and only 39% report any impact on EBIT at company level. Getting people to use AI is no longer the hard part. Getting value out of it is.
That gap was the subject of Elisa Sai's session at Bucharest Tech Week 2026. Speaking on the Future AI Summit stage, the VP Analytics & AI at Capgemini argued that turning AI capabilities into enterprise value is, above all, a leadership challenge. Her hypothesis was that it gets solved less by adding more use cases and more by making fewer, better leadership decisions.
She narrowed those decisions down to three: how value is defined and captured, how work is redesigned around AI, and how governance is built into the systems themselves. Get them right, she said, and AI value compounds. Get them wrong, and AI "will still remain a collection of very expensive experiments in pockets of the organization."
Value measured in pockets doesn't add up
Too often, Elisa argued, AI value is measured at the level of the task. Information arrives faster, code gets written faster. Those gains are real, but they are local: a team here, a department there, doing good work that never changes how the business performs.
Software development is the clearest example. Developers are coding faster and more accurately, but that alone doesn't show whether release cycles are shorter or whether customers get new features sooner.
She has watched this play out in boardrooms more than once. A head of department presents a use case delivering 30 to 40% higher productivity, everyone gets excited, and then a senior executive asks what it means for the business.
Her point is simple: dozens of pilots with nothing in production is a warning sign, not progress. The fix is to back fewer, stronger projects and put the money into the data they depend on.
The business metrics themselves don't change. Revenue, margin and customer satisfaction are here to stay. What changes is how AI moves them, and Elisa suggested three ways to measure that once AI agents start doing real work:
End-to-end ownership. Hours saved mean little on their own ("I could save many, many hours and then spend it to clean my mailbox"). The better question is how much of a messy workflow an agent can handle by itself, with the human orchestrating.
Velocity of action. The earlier wave of generative AI was about getting information faster. Agents are about acting faster, turning a three-week process into seconds, and the value lies in market share and new revenue rather than time saved.
System discipline. Agents can act, but that doesn't mean they should. What gets measured is how reliably they stay within boundaries, escalate to a person when needed and flag situations of low confidence and high risk.
The right model is the one the task can afford
The other half of the value equation is cost. Most teams have seen a pilot that works beautifully until it goes into production, where the business case stops adding up. Meanwhile, AI vendors keep changing their pricing, and organizations are still learning to think in tokens and consumption rather than licences.
Elisa pointed to Uber as a recent warning. The company had pushed its engineers to adopt AI coding tools, and its CTO disclosed in April 2026 that the entire 2026 budget for those tools was gone after four months. COO Andrew Macdonald later said the spending was getting harder to justify without a clear link to productivity.
Her point was not that organizations should hold back on new tools. It was about proportion. Using a frontier model to pull simple data out of an invoice isn't sophistication, she said, it is waste. The question leaders need to ask is what level of intelligence is proportional to the value of the task.
Redesign the work, not just the tools
Elisa's second priority is the one she called genuinely unprecedented: redesigning how work gets done. Earlier technologies changed the tools. Agentic AI changes the workflow itself, meaning the sequence of tasks, who makes each decision, who is accountable, when something escalates, what gets automated and what stays manual.
She brought it to life with a supply chain use case she had seen only days before. A monitoring agent picks up early warning signals, such as weather disruptions or inventory problems. Analyst agents assess the impact and set out the options, a planning agent turns those into plans and makes a recommendation, and an orchestration agent kicks off the interventions once a decision is made.
The human is still there, just doing different work. The person who used to gather the data and run the analysis now validates what the agents produce, and accountability stays with someone who has to approve the decision. "This is not automation," Elisa said. "This is not applying a layer of AI to something that exists." It is a new framework, with roles realigned around new capabilities.
Governance has to live inside the workflow
The third priority is the one people most often call boring: governance. Elisa argued it is becoming the most critical of the three, because AI is moving from assisting people to acting on its own.
If a chatbot gives a wrong answer, you can check the source or ignore it and move on. If an AI embedded in a workflow triggers a wrong action, the consequences are real, whether in compliance, cost, delays or, in the public sector, harm to citizens. Historically, governance has lived outside the technology, in a PDF, a policy, a committee meeting or a checklist. At scale, she said, it has to be codified into the workflow the same way the AI is. In the supply chain example, that means human checkpoints at each stage and what many organizations now call a control plane, rather than a monthly committee.
This is also where trust gets decided. Capgemini Research Institute's survey of 1,500 executives across 14 countries found that trust in fully autonomous AI agents fell from 43% to 27% in a single year, and only 2% of organizations have scaled their agentic AI deployments. Asked about that drop during the Q&A, Elisa said she doesn't see it as mainly a hallucination problem. Leaders are watching costs and effort rise without seeing results at scale.
The questions she put to leaders were practical:
Does the agent have an identity, and are its permissions set up properly?
Can we trace what it did, from data to decision to action?
Can we stop it or override it?
Is there a plan B if it doesn't work?
"These are fundamental leadership choices," she said. "They're not tech choices. They're not model choices." There is a bottom-line element too: poorly governed AI ends up costing more, so economic discipline and governance go hand in hand.
Buying a Ferrari without checking the engine
When the host asked whether AI today is more like the car (a question of which model fits your needs) or like electricity (a technology we don't yet know how to use fully), Elisa answered with an image that sums up the whole session. Most organizations, she said, are going out and buying a Ferrari when what they need is a small, practical car. And they rarely check that it has an engine, which in this case is their data.
Later, when asked about a scaled AI project that failed, she named the biggest obstacle as adoption rather than technology. Adoption is a mindset, she said. It takes more than training or literacy; it means changing behavior and culture, and that takes time.
Her closing line put the argument in one sentence: AI doesn't scale through speed and volume alone, but "through fewer, better leadership decisions."
So here is the question for the leaders reading this: which of those decisions is your organization still putting off? Deciding which pilots to stop, changing how AI value is measured, owning the redesign of a workflow, or deciding where the human checkpoints sit? Tell us in the comments; it helps shape the conversations we bring to the stage next.
Sources
McKinsey & Company, "The state of AI in 2025: Agents, innovation, and transformation" (November 2025).
Capgemini Research Institute, "Rise of agentic AI: How trust is the key to human-AI collaboration" (July 2025). Survey of 1,500 executives across 14 countries.
Yahoo Finance, "Uber burned through its entire 2026 AI budget in four months. Now its COO is questioning whether it's worth it" (2026).
Elisa Sai, session recording from Bucharest Tech Week 2026, Future AI Summit.




Couldn't agree more—running endless AI pilots without operational alignment or clear decision ownership rarely moves the needle. Real value happens when AI is embedded directly into daily team workflows alongside genuine subject-matter expertise. That is exactly why platforms like eminnt—an AI-native content marketing and operating platform for B2B teams that turns internal company knowledge and subject-matter expert input into search- and AI-optimized content—focus on operationalizing internal knowledge into structured, repeatable execution rather than isolated experiments. Insightful read!
Spot-on analysis! Most enterprise AI initiatives stall after the proof-of-concept stage not because the underlying models fail, but because leadership lacks clear decision frameworks, ownership, and workflow integration. Even the smartest AI tool goes unused if teams cannot easily interact with it—which is why partnering with a specialized SaaS UI UX Design Agency to turn complex AI outputs into intuitive, actionable product workflows is so critical for real adoption. Great perspective on moving from endless pilots to measurable business decisions!