Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Enterprise AI scales through the operating model¶
OpenAI's account of the next phase of enterprise artificial intelligence describes rapid growth in organizational use and increasing demand for agents that can operate across real workflows. The direction is clear: companies are moving beyond isolated conversations toward systems that research, update records, create artifacts, and complete multi-step work.
That movement makes the operating model—not access to the model—the limiting factor.
A useful demo can fit inside one team¶
An enterprise capability cannot. It touches data ownership, system integration, identity, security, records, quality, policy, training, support, and budget. It changes what people do and often redistributes work among roles.
Artificial intelligence (AI) pilots can avoid many of those dependencies by using a narrow dataset, an enthusiastic sponsor, and manual oversight. Scaling removes the protective bubble. Variations among business units become visible. Exceptions multiply. A workflow that looks standard from the center turns out to depend on local judgment and unofficial handoffs.
The temptation is to respond with a central platform and a list of approved tools. Those are useful, but they do not create adopted capability by themselves.
Scale needs both a center and an edge¶
Central teams are well positioned to provide shared infrastructure, security patterns, vendor management, evaluation methods, and reusable components. Local teams understand the task, the users, the exceptions, and the consequences of error.
The operating model has to connect them. Too much central control produces generic solutions and long queues. Too much local autonomy produces duplicated integrations, inconsistent assurance, and systems that cannot be sustained.
Wenger's work on communities of practice suggests one mechanism for connecting the two. Practitioners learn by sharing problems, artifacts, and experience across formal boundaries. An AI program can cultivate that network deliberately: reusable evaluation scenarios, office hours, incident reviews, reference architectures, and forums where product teams explain what has actually worked.
Product ownership must outlive the launch¶
Every scaled AI workflow needs an owner for the complete result. That owner does not have to build every component, but must be accountable for value, user experience, performance, exceptions, and change over time.
Without that role, responsibility fragments. The model team owns capability. Information technology owns integration. Security owns access. The business owns adoption. When the system creates hidden review work or a policy change invalidates its assumptions, no one owns the combined failure.
A durable product team needs a cross-functional core and an evidence loop:
- Define the consequential workflow outcome.
- Establish baseline quality, cost, delay, and user burden.
- Introduce AI at a bounded point in the work.
- Measure the complete system, including verification and recovery.
- Capture exceptions and local adaptations.
- Decide whether to deepen, redesign, scale, or stop.
Adoption is a design output¶
People adopt systems when those systems help them accomplish work, fit their constraints, and earn appropriate trust. Training and executive sponsorship matter, but neither can compensate indefinitely for poor workflow fit.
The Technology Acceptance Model has long connected adoption to perceived usefulness and ease of use. Contemporary AI adds another dimension: users must understand when the system deserves reliance and what to do when it does not.
Enterprise AI is entering a phase in which model capability is abundant and visible. Organizational capability remains uneven and mostly hidden. The companies that scale successfully will build connective tissue: shared foundations, local ownership, cross-functional learning, and evidence strong enough to guide investment.
The strategic question is no longer “How do we give everyone AI?” It is “How do we make the organization capable of changing its work responsibly as AI becomes part of it?”
Sources and research trail¶
- OpenAI, “The Next Phase of Enterprise AI” (April 8, 2026).
- Wenger, Communities of Practice: Learning, Meaning, and Identity (1998).
- Davis, “Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology” (1989).
- Edmondson, “The Local and Variegated Nature of Learning in Organizations” (2002).
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0) (2023).