The Problem No ERP Solved

Enterprise software is organized around functions. Finance has its system. HR has its own. IT has another. A request that touches all three — say, an employee onboarding a new vendor — still requires a human to open three separate tools, navigate each one, and hold context across the gap. That gap is where time and cost accumulate. It is also, increasingly, where enterprise AI is aimed.

The solution getting traction in mid-2026 is the super agent: an orchestration layer that sits above existing specialized agents and routes work to the right one through a single conversational interface. It does not replace the underlying systems. It eliminates the handoff between them.

Levi Strauss & Co. made the clearest public case for this architecture when Microsoft published a customer story on June 4, 2026 detailing how the company built specialized agents across HR, finance, IT, and retail operations — and is now connecting them into a unified super agent inside Microsoft Teams. It is one of the more concrete examples of what had been, until recently, a largely theoretical enterprise AI pattern.

How Super Agents Work

The architecture follows a consistent pattern across early deployments: build the specialists first, add orchestration second. Levi’s built agents for HR self-service, a custom SAP integration, and a retail agent spanning inventory and store operations before layering the super agent on top. The orchestration layer — built on Microsoft Foundry with Azure Functions handling event-driven execution — interprets employee intent and routes the request to the right underlying agent without requiring the employee to know which system is relevant.

The key distinction from a standard chatbot is what happens with the output. A chatbot produces an answer. A super agent produces an outcome: it takes the response from one specialized agent and passes it as input to another if the task requires it. Multi-step workflows — submit an IT ticket, update an HR record, trigger a finance approval — can be completed through one interface in sequence rather than three interfaces in parallel.

Jason Gowans, Levi’s chief digital and technology officer, described the stakes plainly: “This isn’t just about a tool — it’s a wholesale workplace transformation.” The framing matters because it signals where the real investment is. The model is not the bottleneck. The orchestration layer, the data access, and the governance around what agents can and cannot do — that is where implementation effort concentrates.

You can read more about how orchestration frameworks are evolving in our earlier coverage of agentic frameworks that actually ship to production.

What Early Deployments Actually Show

Goldman Sachs is applying the same logic in financial services. The bank is testing AI agents — built with Anthropic’s Claude — to automate transaction reconciliation, trade accounting, client vetting, and onboarding. These are workflows that have resisted automation for decades because they require processing large volumes of data against strict regulatory requirements. Multi-agent coordination makes it tractable: one agent handles data extraction, another checks regulatory thresholds, a third routes exceptions to humans when guardrails are triggered.

Ramp launched Applied AI Solutions in June 2026, specifically targeting workflows that span multiple finance systems and require judgment when exceptions occur. Ori Daniel, Ramp’s head of AI solutions, described the core challenge: “In finance, every decision depends on buried layers of context: the policy, the vendor, the contract, the approval chain, and the exception history.” The product captures that context and turns it into agents that complete work within finance-defined controls — rather than around them.

EY’s Canvas platform is the most mature at scale: it processes 1.4 trillion lines of audit data annually across 160,000 global engagements, with a multi-agent framework serving 130,000 assurance professionals. In April 2026 EY extended the platform on Microsoft Azure, moving from individual AI assistance to coordinated agent teams that handle sub-tasks of an audit in parallel and reconcile outputs against a master workflow. JPMorgan, meanwhile, reports 83% faster research cycles for portfolio managers using its LLM Suite and has automated over 360,000 manual hours per year across 450 daily production use cases.

The Numbers Behind the Shift

Multi-agent workflows grew more than 300% over several months as organizations moved from pilots to production, according to Databricks data reported by PYMNTS in February 2026. That number reflects the pattern: most enterprises spent 2025 deploying isolated agents within a single department. 2026 is when those departments started talking to each other.

The finance function is where the ROI argument is clearest. PYMNTS Intelligence found that 43% of CFOs believe agentic AI could have a high impact on dynamic budget planning — and nearly half already use AI to monitor working capital and cash flows. The gap is between monitoring and acting. Super agents, operating within defined guardrails, can update projections, flag variances, and initiate approvals without routing each step through a human. That is the function enterprises are now building toward, not just piloting.

Gartner’s earlier prediction — that 40% of enterprise applications would feature task-specific AI agents by 2026, up from under 5% in 2025 — appears to be tracking. What Gartner did not fully anticipate is that the agents would not stay task-specific for long. The orchestration layer is what turns a collection of task agents into something with cross-functional reach.

Where This Gets Hard

The constraint that has kept work fragmented across enterprise systems is not technology — it is that the systems were never designed to share state. Super agents sit on top of that infrastructure and coordinate across it rather than replace it. That is a practical advantage in the short term (you do not have to rip and replace anything) and a governance challenge in the medium term.

When a single agent makes a mistake, the blast radius is limited. When an orchestration layer routes a flawed decision across three downstream systems before a human sees the output, the blast radius is not. Defining what agents can authorize autonomously versus what requires human-in-the-loop is not a technical problem — it is a policy problem, and most enterprise legal and compliance teams are still writing those policies in 2026.

Microsoft Scout, covered previously on this site, introduced agent identity as a partial answer: giving each AI agent a persistent credential that logs what it accessed and what it did. That helps with auditability. It does not resolve the question of how to define the authorization envelope before deployment. The companies getting this right — Levi’s, Goldman, EY — are the ones that built governance frameworks before they built the super agent, not after.

Further Reading

• Levi Strauss & Co. on Microsoft Foundry — the primary source for the Levi’s architecture, including stack details and rollout timeline
• Super Agents Are Connecting What Enterprise Software Kept Separate — PYMNTS’ July 1 analysis covering Goldman, Ramp, and the broader CFO survey data
• AI Agent Orchestration Goes Enterprise: The April 2026 Playbook — deeper look at EY Canvas and JPMorgan’s production metrics