Operating Design
How to Build an AI-Native Unit Without Creating an Innovation Island
A focused team can move faster than the core organization. The design challenge is giving successful capabilities and operating practices a path back in.
Many organizations begin their AI work with access. They buy licenses, offer training, and encourage employees to experiment.
That can reveal useful ideas, but access alone does not create an operating capability.
The experiments remain distributed among individuals, data access varies by team, and no one owns the path from a promising shortcut to a production system. The organization becomes busier with AI without becoming meaningfully more capable.
KPMG’s September 22 announcement of a new Client Technology & Innovation group points toward a different model. The firm says CT&I will build AI-native businesses at the edge, combining product development, alliances, venture relationships, and the resources of the larger organization. Some successful businesses may later return to the core.
KPMG’s expected outcomes are company claims, and the structure has not yet proved itself publicly. The announcement is still useful because it exposes the real design problem: how do you give a focused unit enough freedom to build differently without allowing it to become disconnected from the organization it is meant to strengthen?
Define the unit’s job and the route back
An AI-native unit needs more than talented people and newer tools. It needs a defined problem to solve, an accountable owner, controlled access to the right information, and a clear boundary around what it may change. It also needs to know what happens if the experiment works.
That last point is often missing.
Many innovation teams are designed for exploration but not transfer. They can produce prototypes, demonstrations, and early wins. When the work needs to enter normal operations, it collides with existing security requirements, data definitions, ownership disputes, budgets, and quality standards that were never included in the experiment.
When a promising pilot stalls at transfer, the missing bridge may be organizational rather than technical.
A stronger model begins by separating exploration from production without pretending they are unrelated. The edge can use faster cycles and narrower teams, but it should build against real operational constraints. The core should not impose every legacy process on day one, but it must identify the standards that cannot be postponed: data rights, security, auditability, financial authority, customer impact, and release ownership.
Task-specific agents fit naturally inside this design. A research agent, testing agent, documentation agent, or monitoring agent can be given the context and permissions required for one responsibility. That is more reliable than treating one general assistant as a universal employee. It also makes review easier because the organization can see which capability acted, what information it used, and who owns the result.
Decide what production readiness means
The unit’s mandate should include transfer criteria from the beginning. A capability may be ready to move toward production when the problem is stable, the outcome can be measured, exceptions are understood, access is limited to what is needed, and an operating owner is prepared to accept responsibility. If those conditions never arrive, the organization should know whether to continue learning, narrow the scope, or stop.
Smaller companies do not need a formal venture group to use this model. A team drawn from several functions can serve as the edge for one important process. What matters is that the team has room to build, access to the people who understand the work, and a defined route into normal operations.
AI-native thinking should eventually spread beyond the unit. That does not mean every employee receives every tool or that every workflow becomes autonomous. It means the organization learns to redesign work around better context, clearer decisions, and capabilities that were not previously practical.
The goal is an organization that can absorb what the unit discovers. That requires the route into production to be part of the original mandate, not a problem deferred until the pilot succeeds.
Source
- KPMG: Client Technology & Innovation group announcement, September 22, 2026.