Agentic AI systems, software that act across multiple business apps instead of only answering chat prompts, are replacing the human routing that enterprise software forced on workers for thirty years. Goldman Sachs is now testing the same model on accounting and onboarding.
An employee asks one question about inventory, IT, or HR and the answer lives behind three separate apps, each with its own login, vocabulary and approval chain. For three decades, that friction was simply the cost of doing business in a large company. Now Levi's, Goldman Sachs and a growing list of enterprise software buyers are routing around it with the same kind of fix: a single AI doorway that takes the question, fans it out to the right system and comes back with an answer or an action.
The construct has a vendor nickname, "super agent," and a more academic one, "multi-agent AI." In plain English, it is a layer of software that holds a conversation and, instead of stopping at a chat answer, hands the underlying task to the specialized system best suited for it. One bot stays at the surface for the user. Behind it, a handful of narrower agents handle the work in human resources, finance, IT, retail or accounting. The promise is that the employee or operator never has to know which system to open next.
Levi Strauss & Co. is the clearest named example of the model in a public deployment today. The retailer has built specialized AI agents across its HR, finance, IT and retail operations and is layering what it calls a "Super Agent" orchestration tier on top of them using Microsoft Foundry, Microsoft's platform for building and wiring AI agents. Jason Gowans, Levi's chief digital and technology officer, framed the project in company-published interview material as a "wholesale workplace transformation," citing the speed required of a direct-to-consumer retailer. Sheena Kunhiraman, the company's vice president of HR technology and analytics, said the human-and-agent collaboration inside Levi will center on augmentation and "giving time back," language the company has used throughout the rollout. Because that material is published by Microsoft rather than independently audited, the architecture is best read as Levi's self-reported blueprint for the model, not as an outside-verified outcome.
That is not the only place the pattern shows up. Goldman Sachs is testing AI agents built with Anthropic's Claude, a large language model from the AI lab Anthropic, to handle transaction reconciliation, trade accounting, client vetting and client onboarding. These are back-office functions historically resistant to automation because the volumes of regulatory data make mistakes expensive. PYMNTS and American Banker have reported the rollout as a pilot across several compliance and finance workflows, not as a fully productionized deployment. The cross-industry move is the point. The same doorway logic that runs a retailer's HR and retail operations is now being tested where the regulatory bar is higher.
How widespread is the pattern? Databricks's State of AI Agents 2026 report, cited by PYMNTS in February, found that multi-agent workflows grew more than 300% over "several months" as organizations moved from pilots to production. That is a vendor-research headline over an unspecified time window and should be read as a trend signal rather than a precise measurement. On the buyer side, PYMNTS Intelligence found that 43% of CFOs surveyed said agentic AI could have a high impact on dynamic budget planning, with nearly half already running AI in at least one finance workflow in production. The survey's sample size was not disclosed in the public reporting, so the figure is best treated as executive sentiment rather than hard budget reallocation.
The architectural shift underneath those numbers is the part worth naming. For thirty years, enterprise software has been sold and installed function by function. A customer-relationship system for sales. A human-capital-management platform for HR. An enterprise-resource-planning system for finance. A ticketing tool for IT. The integrations between them have been the most expensive, fragile part of every large IT estate, and the connectors that survived were always narrowly scoped. What an AI doorway changes is not the underlying systems. It changes who, or what, knows how to reach across them. If the layer holds up, the next several years of enterprise integration spending may be less about wiring systems to other systems and more about teaching one AI how to use the systems that are already there.
What to watch next. Whether the Levi's deployment and Goldman's pilots graduate from orchestrated demos to the harder work of audit trails, regulatory traceability and error cost in regulated processes. Goldman's accounting and onboarding workloads will show whether the doorway model survives where the failure modes are expensive. Whether the "augmentation" framing inside Levi's HR rollout holds as roles change, or whether the workforce-shape consequences of routing half the old handoffs through software become the real story over the next four quarters. And whether the multi-agent system line, as currently defined by vendor marketing, becomes a category or merges back into something more mundane, like a workflow engine with a chat surface.