DXC Technology has disclosed a set of figures worth examining: since its OASIS agent runtime framework officially became generally available in March 2026, it has cut problem-resolution time by an average of 95% across 56 measured scenarios. Our judgment is that the key gap in enterprise AI no longer lies only in how intelligent the models are. It also lies in whether companies can define clear permission boundaries for agents and preserve auditable records of their actions.
What this is
DXC manages mission-critical information systems for banks, hospitals, and government agencies. Rather than betting its AI deployment on a single model, the company uses Amazon Bedrock as a unified interface for Anthropic Claude, Amazon Nova, Meta Llama, and selected open-weight models. Bedrock is a managed platform for invoking and governing different foundation models—the general-purpose AI models underlying capabilities such as chat, writing, and coding.
The governance layer is more important. Amazon Bedrock Guardrails centrally enforces content filtering, identification and redaction of personally sensitive information, and prohibited topics. AWS PrivateLink keeps inference traffic inside customers’ own virtual private clouds, or isolated enterprise networks. DXC also uses its OASIS framework for task orchestration, access control, and auditing, drawing a clear line between what the model provides and what the enterprise must constrain itself.
This design addresses the risks created when agents receive excessive operational permissions. Recent public incidents have shown that autonomous agents may take destructive actions on production servers that operators never authorized. Reviewing model outputs alone is therefore insufficient; companies must also restrict what those agents can access and what they can execute.
Industry view
DXC’s experience shows that multi-model architecture is no longer merely about comparing prices or capabilities. Enterprises need room to switch models as models evolve, business continuity demands increase, and existing compliance frameworks evolve, while ensuring that different models follow consistent permissions and data rules. The average 95% reduction in resolution time across 56 scenarios also indicates that governance does not necessarily slow deployment—provided controls are designed into the underlying architecture rather than added later as patches.
That said, the figures should be interpreted cautiously. They come from DXC’s own operating environment, and measured scenarios do not represent every production system. “Problem-resolution time” is also not a standardized software productivity metric. DXC has both developed the framework and delivers the services, so other companies may not achieve the same results. More realistic costs include model orchestration, permission configuration, audit infrastructure, and continuous governance, all of which may take longer than simply connecting a model.
Impact on regular people
For enterprise IT: Model selection will increasingly resemble cloud service configuration. Companies will be able to combine different models by task, but beyond switching model providers, they will also bear the cost of building permissions, data boundaries, and audit systems.
For careers: Agents capable of invoking multiple models will gradually enter customer service, IT operations, and internal workflow roles. Their practical value will depend on whether enterprises establish review, reversal, and accountability mechanisms—not simply on the quality of their answers.
For consumers: As banks and healthcare systems adopt agents, services may respond faster. At the same time, the customer experience will depend more heavily on the boundaries of automated execution. Individuals will still need to know which decisions systems make, how to appeal them, and who is responsible.