How Google could gain enterprise AI control through model choice

Model choice and the contest for enterprise AI control

Google’s offer of more model choice could make its platform more central to enterprise work. At Gemini at Work 2026, Google unveiled what it calls a “single, universal agent” and an application programming interface, or API, in Gemini. An API lets software systems communicate, while an agent uses tools to work toward an objective. For chief information officers, chief information security officers, and engineering and platform leads, the strategic question is who coordinates that work, grants access, and remains responsible as the underlying models change.

Choosing a model and choosing the platform that directs its work are separate decisions. Google’s proposal offers flexibility in model selection while seeking a central role in coordinating enterprise activity, a position that would benefit Google commercially. For buyers, this raises a practical test: whether the platform can support trusted autonomy, with agents pursuing objectives under limited step-by-step human direction. The answer depends on how the platform controls the authority enterprises delegate.

Model flexibility can strengthen the coordinating platform

Google’s approach separates the agent from its underlying model, allowing different workflows to use different models. The named options are the Gemini family and Claude, and Google says support for additional leading private and open models will arrive “soon.” Google’s stated objectives are better quality and lower cost. Selecting a model would become part of the platform’s responsibility for managing work.

Model selection may need to change as tasks and models change. Google Cloud CEO Thomas Kurian says the leading model changes “every few months,” while “the best model for the task is not always the largest one.” His argument supports Google’s commercial proposition: an enterprise could keep the same coordinating platform while changing the model used for a workload. The continuing relationship would therefore depend partly on how well that platform selects and coordinates models over time.

Independent tech analyst Carmi Levy sees this separation as central to Google’s strategy. “It separates the agent from the underlying model, and allows enterprises to use whatever model makes sense for a given workload,” he said. In his view, Google is pursuing a position that could outlast leadership in a particular model comparison. “Instead, it’s trying to establish itself as the gatekeeper of the new enterprise operating system, powered by whatever underlying model makes the most sense,” he said.

Levy places that ambition within a history of strategic control moving through operating systems, browsers, apps, cloud platforms, and large language models, or LLMs, which generate language-based responses. He now sees the infrastructure coordinating enterprise workflows as the contested area. “The agent layer could be the ultimate prize of the AI era, and it largely explains why every vendor announcement seems to essentially claim the same thing,” he said. In that reading, easier model switching could make the system coordinating their use more important.

But competition for that position is already substantial. Mahmoud Ramin, research director at Info-Tech Research Group, says Google has not created a new category, since OpenAI, Anthropic, Microsoft, and Meta already position multiagent capabilities, in which several agents work together, within their platforms. Levy also points to similar claims from OpenAI and Microsoft. Model flexibility therefore gives Google an offer to put before buyers, but the overlap means buyers still need to assess why its approach would suit their work.

Persistent agents expand the authority enterprises delegate

To assess that offer, buyers first need to understand how Google expects work to proceed. A user supplies an objective through a prompt box, and the Gemini agent uses connected skills, tools, and enterprise systems to pursue it. Google describes capabilities including answering questions, generating media, and writing and executing code, with completed work delivered to documents, inboxes, or developer environments. Kurian describes the interaction this way: “You give it objectives, not instructions,” and “Work now starts in the prompt window.”

Starting with an objective makes the agent responsible for deciding intermediate actions. Google says its agent can create temporary or job-specific sub-agents, supporting agents assigned parts of the work, and communicate with them to coordinate execution. These agents can work in parallel or carry out steps in sequence. Google describes workflows that may continue for “hours or days,” extending the delegation beyond a single exchange.

Work that continues across sessions also needs an identity and a place to retain information. Google says coworker agents can have dedicated identities, agent-specific email addresses in the illustrative form @agents.company.com, and persistent storage that remains available across sessions. They can work across days as their responsibilities change. An enterprise adopting this arrangement would give an identifiable participant authority that can remain active after the initiating conversation ends.

Google connects that continuity to defined workplace roles, describing agents as personal assistants or team members with project-manager or analyst responsibilities. Role-based agents would be optimized for particular duties, a feature Levy characterizes as “compellingly unique.” The proposal is to assign continuing work to an agent configured for a role. Its stored information, identity, and permitted responsibilities would therefore need ongoing management.

Those responsibilities could span several ways of accessing the agent. Kurian describes “omnipresent access” through cloud-based execution, with support for iOS and Android phones and Windows and Mac desktops. Users could reach the agent through command line interfaces, which accept typed commands, as well as Google Workspace, Microsoft 365, Slack, and third-party applications. Google also describes headless agents, which operate without a dedicated user interface and can therefore continue independently of a particular screen or application.

Across these channels, an agent could receive work in one place, act through connected systems, coordinate supporting agents, and deliver results elsewhere. Google says access is limited to specifically defined data and channels. To govern the delegated work, those limits must apply as it moves between systems and agents. Continuing authority thus becomes a central part of the purchasing decision.

Governance must follow the work across systems

Continuing authority expands the consequences of an agent’s decisions. A basic AI interaction may produce an answer for a person to review, while an autonomous agent can act through several business systems and communicate with other agents. Ramin describes the risk associated with basic AI as “much more contained.” With autonomous work, buyers must establish which permissions apply, which decisions an agent may make, and who is accountable for the resulting workflow.

These questions become more serious when the workflow supports a mission-critical business process. An agent might have enough access to act and enough continuity to keep acting before a person intervenes, so a serious mistake could disrupt operations and cause reputational damage. Successful task completion therefore addresses only part of the buying decision. Buyers also need to examine what happens when a decision goes wrong during continuing work.

Google says its offering includes enterprise-defined security, governance, and cost controls, alongside restrictions to specified data and channels. Because these controls are part of the platform Google is selling, buyers need to assess how its claims apply to the proposed activities. An agent might change tools, create supporting agents, or continue across sessions. A control must remain effective through those changes to constrain the work it was intended to govern.

Ramin sets out the responsibilities enterprises should address alongside those claims. “As organizations unleash the power of agent autonomy, it raises the pivotal and non-negotiable requirement of enforcing governance,” he said. He calls for permissions, identity, human oversight, and system audits: permissions constrain access, identity establishes which agent is acting, oversight enables human involvement, and audits allow examination of system activity. These safeguards cover different parts of the same workflow, so buyers need to understand how they work together.

Access controls also need explicit decision limits, because permission to reach information does not by itself establish authority to act on it. Ramin says enterprises should “clearly outline which decisions the agent is allowed to make, what data should be accessed, and how performance will be monitored.” For persistent coworker agents, those limits must remain tied to their current responsibilities as assignments change. Granular permissions, meaning access defined for specific activities and information, help keep a continuing role within its assigned scope.

Levy examines accountability as part of Google’s competitive positioning. He says Google is “arguably leading the conversation around identity, audit trails, and who, or what, is ultimately responsible for workflows.” Audit trails record actions taken and help establish responsibility when several agents contribute to a result. His assessment concerns Google’s emphasis in the enterprise discussion; buyers still have to examine how its controls work during autonomous activity.

That examination broadens as the platform connects to more external systems and workplace channels. Each connection extends the activity that permissions, oversight, and accountability must cover, including actions by supporting agents. Ramin’s recommendations identify governance responsibilities, while Levy’s assessment describes their place in platform competition. Buyers can use both perspectives to assess how governance follows a workflow across connected systems.

Levy’s trust test focuses on the duration and consequences of delegated work. He describes the industry’s direction as an “inevitably agentic future,” but asks “whether CIOs and CISOs can trust it to run mission critical business processes indefinitely without doing something stupid enough to generate damaging headlines is another story altogether.” His question sets a demanding threshold for confidence in sustained operation: whether the platform can preserve enterprise-defined boundaries as responsibility continues beyond the initial prompt.

Workplace fit can make the proposition easier to consider

While governance determines the scope of acceptable delegation, existing applications shape where an enterprise might first consider using it. Ramin highlights Google’s combination of direct operation in Gmail, Drive, Sheets, and Slides with connections to external systems. For organizations already using those products, agents would work inside familiar applications while reaching other parts of the business environment. The proposal therefore connects directly to existing work.

Within that environment, Ramin identifies repetitive work and information synthesis, gathering material and turning it into a usable result, as strong applications. He points to research by project managers, analysts, and marketing specialists. Related examples include writing an email, adjusting a spreadsheet, and building a presentation. These tasks show where an agent could reduce the work a person directs step by step, although competing offerings have similar capabilities.

Similar capabilities make the buyer’s existing environment more important to the commercial decision. “Will enterprises care? If they’re already heavily invested in Google’s workplace stack, this could be an easier sell in the C-suite,” Levy said. Familiar applications can make the proposed expansion easier to explain to executives, giving Google an advantage tied to the customer’s current investment. Performance, safety, governance, and value remain separate matters for assessment.

For organizations outside Google’s workplace environment, external connections keep the proposal relevant, but the advantage of an installed product base is smaller. Buyers would need to examine how agents fit the applications through which their own work moves. Familiarity can help a platform enter consideration; permission to act through those applications requires a decision about the specific work and authority being delegated.

Key takeaways for decision-makers

• Separate model choice from platform control: Google’s support for multiple models could strengthen its role in coordinating enterprise work. CIOs evaluating that flexibility need to assess the platform’s authority over workflows alongside model quality and cost.
• Define continuing agent authority: Persistent agents can retain identities, store information, and coordinate supporting agents across days. Business owners and security teams need explicit limits on their decisions, access, and ongoing responsibilities.
• Test governance across the workflow: Permissions, oversight, and audit trails need to remain effective as agents change tools, delegate tasks, and cross business systems. CISOs and platform leads can use those transitions to test Google’s governance claims before approving mission-critical work.
• Use workplace fit to scope adoption: Existing use of Gmail, Drive, Sheets, and Slides gives Google a practical entry point. Application owners can assess bounded research, email, spreadsheet, and presentation tasks against performance, safety, and value requirements.