Newsletter Article
A successful AI pilot proves one thing clearly: the use case creates value.
It rarely proves the second, harder thing, that the value survives contact with more users, more data and more of the organisation.
That gap between proving and scaling is not a fringe problem. It’s the norm.
~2/3
Of enterprises have experimented with AI agents (McKinsey)
Less than10
Have scaled them to deliver tangible value (McKinsey)
40%+
Of agentic AI projects predicted to be cancelled by end of 2027 (Gartner)
Sources: McKinsey, “Scaling agentic AI with data transformations,” 2026; Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 2025.
Neither figure is a verdict on the technology. Gartner is explicit about the cause: cancellations trace back to escalating costs, unclear business value and inadequate risk controls, not to models that don’t work. McKinsey points to the same root issue from the data side: eight in ten companies scaling agentic AI cite data limitations as the roadblock. In both cases, the pilot proved the idea. What was missing was the architecture to carry it further.
The pilot proves the idea. Scaling proves the architecture.
What changes between a pilot and a scaled capability
A pilot typically starts, sensibly, with the minimum needed to prove a use case:
One Use Case → One Dataset → One Team
Scaling doesn’t just mean doing more of that. It means the environment the AI operates in gets structurally more complex:
Multiple Use Cases → Multiple Systems → Multiple Personas → Multiple Decisions
That complexity is exactly what a pilot’s shortcuts weren’t built to handle, a one-off data pull, an informal definition of a metric, an integration wired up for a single use case. None of those decisions were wrong for a pilot. They just don’t survive being asked to do more.
Six foundations that carry a pilot into production
-
Build on trusted data
Every AI capability depends on the information available to it. As a pilot expands beyond the team that built it, the definitions that were understood informally need to become explicit and consistent. “Customer,” “interaction,” “revenue,” “case” and “service level” all need to mean the same thing to everyone relying on them, otherwise every new team that adopts the capability inherits a slightly different version of the truth.
Ask yourself: if two teams pulled the same metric today, would they get the same number, and could they explain why, if not?
-
Connect the wider business context
A pilot proves value with the minimum data required. Scaling is the point at which that context gets progressively richer: customer information connects to interaction history, interaction data connects to workforce information, workforce connects to operational demand, operational data connects to financial outcomes. Each connection isn’t just more data, it’s a new question the organisation can now answer.
Ask yourself: what’s the next system this use case would benefit from being connected to, and what question would that unlock?
-
Create reusable data and integration foundations
Wiring a new integration for every AI initiative creates duplicated effort and duplicated risk, every pipeline is another thing that can drift out of sync. A shared data and integration layer means the first use case contributes to the next, rather than every team starting from zero.
Ask yourself: if this pilot succeeds, does the next team building on it start from a foundation, or from scratch?
-
Design governance into the architecture
Governance becomes more valuable, not more optional, as usage grows. Who can access the data? Which information can each role see? What context can the AI use? What actions require human approval? How are decisions monitored? These questions are far cheaper to answer at the design stage than retrofitted after a rollout. Genesys makes a version of this same point in its own AI strategy, positioning centralised governance and guardrails as central to creating and scaling AI-powered experiences.
Ask yourself: if this capability were used by three times as many people tomorrow, would the current access controls still hold?
-
Make intelligence relevant to the person using it
Scaling doesn’t mean giving every employee an identical AI experience. A workforce manager and a CEO might rely on the same underlying data, but the decisions they make from it are entirely different, and the context, detail and format of what they see should reflect that. This is where persona-based intelligence becomes a scaling requirement rather than a nice-to-have.
Ask yourself: does this capability currently show the same view to every role, and would a new user know what to do with it?
-
Connect insight to action
The final measure of a scaled AI capability isn’t how many answers it produces, it’s what the organisation does differently as a result. That’s what separates an impressive demo from an operational asset.
Prove → Connect → Govern → Contextualise → Scale → Act
The pilot isn’t the finish line
A successful pilot doesn’t need to become a separate enterprise AI project competing for its own budget and attention. Treated well, it becomes the first building block of a broader intelligence capability, the proof that justifies investing in the six foundations above, rather than a one-off that quietly stalls once the initial excitement fades.
That’s also precisely where the gap between the ~2/3 who experiment and the <10% who scale tends to close: not with a bigger pilot, but with a sturdier foundation underneath the one that already worked.
IS YOUR PILOT READY TO SCALE?
We’ve turned these six foundations into a short, practical self-assessment you can run against any AI pilot in your organisation, the AI Pilot Scaling Checklist.
