Adopting more intelligent machines is not the only aspect of the next stage of corporate AI. It’s about knowing what intelligence a firm should own and what it should rent, and whether its people and its organization can change as it does so.” The subject of which model a corporation should utilize dominated the corporate AI discourse for two years. The answer was usually a familiar list of names from Silicon Valley, a cloud contract, an API key and a promise that ever more powerful general-purpose intelligence will change the business. Although that age is not going away, something more significant is starting to take place beneath it. Businesses are beginning to pose a different query: which aspects of our intelligence ought to be ours?
Salesforce’s most recent action demonstrates the shift. Salesforce unveiled Koa, its first CRM reasoning model, on September 15, 2026. It was created by post-training NVIDIA’s Nemotron 3 Super using a proprietary synthetic dataset that was derived from almost thirty years of Salesforce CRM expertise. Salesforce claims to perform post-training and inference within its own trust boundary and to control the model weights. The fact that Salesforce has chosen to take on the frontier labs at their own game is not what matters. It has taken a more calculated action. “It has taken a general capability and molded it into an intelligence that has been molded by the company’s own experience of how enterprise work is done.”
That distinction is important. Isolation and rewriting every layer of the AI stack are not necessary for sovereignty. Thomson Reuters made the case unusually bluntly in August when it established its own model for professional employment, Thomson. The company argues its edge comes from merging general frontier models with specialized intelligence on top of its unique information, professional knowledge and workflows. Sovereignty in its own description isn’t about not leveraging external intelligence, but owning the layers that matter to the organization.
It may be one of the defining principles of enterprise AI. A broad model is more and more accessible to all. What’s impossible to replicate, is what a company has gathered over decades: its decisions, methods, vocabulary, customer relationships, operational history, institutional knowledge and subtle judgments of its specialists. These assets once resided mostly in people, documents, databases and software. Increasingly they can also be in the intelligence layer of an organization .
SAP, for its part, is attacking the same challenge from a different angle. Its Joule Agents are meant to work inside the established bounds of authority of an organization. “An agent acting on behalf of a human can’t exceed the authorizations of that human, and agent activity is still controlled by role-based permissions, approval workflows, audit logging and other controls already in place,” SAP says. Its AI Agent Hub is being designed as a central system for discovering and governing agents, models and MCP servers, including identity and access control and links between agents, organizational structures and workforce capabilities.
That’s more than a security function. It is a paradigm shift in the concept of corporate intelligence. Intelligence must have an identity, an owner, a scope of action that is allowed, and a position in the organization that is accountable. The question is no longer whether an AI system can answer a query. The question is what that system can know, what it can do, who’s authority it functions under and how the company can prove what happened.
Siemens is applying a similar structural approach in the industrial sector. Its Intelligence Center X combines data, workflows and AI agents in a regulated environment where people and agents are in the same production processes. Siemens calls the idea a hybrid workforce, where AI is built directly into industrial workflows rather than being another separate tool on the side. Its more general sovereign AI study has also looked at control of data, infrastructure, model choice and the ability to avoid vendor lock-in.
The tendency is not limited to software businesses. JPMorganChase has developed its own LLM Suite, which it is deploying firmwide. Its recent annual report highlights an Employee Assistant that may assist workers in retrieving information and taking action across the enterprise. The bank states that more than 90% of its engineers are now using AI coding helpers and more than 65,000 colleagues in its commercial and investment bank are actively using LLM Suite. Crucially, JPMorgan points out how employees are moving beyond simple testing and incorporating generative AI via internal APIs into applications that are part of their business and daily routines.
UBS has a differing assessment of the transition. Its in-house assistant Red has been deployed to about 100,000 employees and has answered more than 25m questions. But UBS has done more than just provide employees with an assistance. It has built a variety of AI learning routes for employees, managers, leaders and technical specialists. Tens of thousands of employees have completed AI learning programs, and over 96% of managers have completed its initial AI Enabler training as of July 2026. “The structure of the program makes the message clear: The giving of AI to people is not the same as teaching an organization how to work with AI.
That second challenge can be tougher than the first.
“Companies can buy GPUs. It can subscribe to a frontier model. It can construct a private inference environment. It can build agents and connect them to enterprise systems. None of this guaranties the organization will change. An AI helper in a poorly designed process can just make the old process go quicker. An agent within a hierarchy structured on sluggish approvals can become just another layer of bureaucracy. Thus, a corporation can become technologically sophisticated without organizational change.
The recent research from Microsoft describes the problem in unusually blunt language. Its 2026 Work Trend Index examined 20,000 AI-using workers in ten countries and found only 19 percent were in what it calls the Frontier category, where individual AI capacity and organizational readiness reinforce each other. Microsoft showed that organizational characteristics (e.g., culture, manager support, and talent procedures) were more than twice as impactful as individual elements for claimed AI impact. The conclusion is that employees are typically ahead of the organizations surrounding them.
Another view McKinsey’s analysis for 2026 arrives at a similar result. In a survey of 701 CEOs and senior leaders, it identified a small segment of enterprises it labels reinventors, companies that are rethinking how work gets done, rather than just adding AI to current procedures. That was the case for just 13 percent of the organizations surveyed, but 48 percent of leaders in that category said they saw meaningful enterprise value from AI, compared with 24 percent of organizations focused on automation and 13 percent of those still primarily enabling employees with general AI tools. McKinsey’s research shows it’s not just technology, but changes in structure, governance, procedures, leadership, talent, purpose and rewards.
This is where the concept of human intelligence becomes more crucial, not less. As computers get better at doing, human value moves to what ought to be done in the first place. In its 2026 research, Microsoft portrays the emerging job in similar terms: the effective worker is increasingly the person who sets intent, defines the desired goal and quality bar, determines how humans and AI should work together, applies judgment and accepts responsibility for the result. The work shifts from doing each task to managing a system that can do multiple things.
The managerial implications are significant. Organizations have typically been designed around people in set positions. AI generates an on-demand workforce that can be formed dynamically. A financial analyst can be coupled with research agents, data agents and forecasting systems. A software engineer can work with coding agents, testing agents and documentation agents. Lawyers can supervise systems that pull up precedent, compare terms and generate first drafts. The human is not removed from the system. The unit of work is itself altered.
That’s why the evolving architecture of the organization may become less about departments and more about skills. McKinsey’s research indicated firms further along in the AI transformation journey are more likely to organize cross-functional teams around products, customer journeys and end-to-end processes, while pushing decisions closer to where value gets created. They also tend to emphasize skills rather than static job definitions, and to regularly re-engineer procedures as capabilities develop.
There is an even more interesting outcome. A really sovereign AI system shouldn’t just have the knowledge that the corporation already has. It should be learned from the company itself. The potential for organizational learning exists in every transaction, exception, successful choice, failed procedure, customer engagement, and engineering iteration. The intelligence system becomes a recollection of how the organization does things, but the people are still accountable for determining what that experience means and what things should change.
We can already see this among AI-native companies. Anthropic said this week that Claude now makes up 26 percent of the company’s AI research and development activities, up from just 1 percent in March. The company also announced that in August, over 90 percent of its research was human-AI collaborative, with over 30,000 AI agents engaged on its internal platform. It’s not just the scale of AI adoption that matters. The point is that the corporation generating intelligence is itself becoming a laboratory for new kinds of human-machine work.
The evolving corporate model is not “AI versus people.” It’s not that straightforward. Businesses will more and more have a number of types of intelligence working in concert: frontier intelligence leased from third-party suppliers; specialized intelligence trained or adapted from proprietary knowledge; small models running close to the data; deterministic software; agents able to take action and, critically, human intelligence that sets intent, context, accountability and consequence.
The competitive question will thus be taken to the next level. It won’t just be who has access to the brightest model. Most serious companies will have access to more powerful models. The one that has built the best relationship between its own knowledge and those models, that controls the intelligence that differentiates the business, can change providers without losing its institutional memory, and has re-organized its people and processes so that intelligence actually changes what the company is capable of doing. That is how Technology Transcendents helps firms get this right, harmonizing both human and machine intelligence and driving the organizational design necessary to drive their business from good to great.
In that sense, sovereign AI is essentially a human question. The machines may do more of the killing. The company still needs to figure out what it knows, what it values, what it is willing to hand off, what it has to keep and how it learns from experience. The organizations who get this aren’t only going to have AI. They’ll have something more like their own institutional intelligence.
And that may be the true change taking place now: from firms that deploy artificial intelligence to organizations that generate their own intelligence, robots and humans learning to function together as one system without becoming identical.