How multi-agent AI is redefining enterprise software

For decades, enterprise software has been designed primarily to record what businesses already know: customer information, financial transactions, employee activity, inventory, sales pipelines, and operational decisions. These platforms became the systems of record that organizations relied on to maintain a consistent version of the truth.

Artificial intelligence is beginning to challenge that model. The emergence of multi-agent AI systems could push enterprise software toward something more dynamic: systems of action that do not simply store information but interpret it, coordinate decisions, and initiate tasks.

The shift is significant because enterprise workflows rarely consist of a single question with a single answer. A customer-service issue, for example, may require reviewing account history, checking contractual terms, identifying a potential solution, updating a ticket, and communicating with another department. A single AI assistant can help with individual steps, but coordinating an entire workflow requires something closer to a team of specialized digital workers.

That is where multi-agent orchestration enters the picture. For Swaroop Borukar, a seasoned Product Manager building in the AI Infrastructure domain at Workday, this represents a fundamental change in how enterprise applications can be conceived.

Swaroop says, “The real opportunity with multi-agent AI is moving beyond systems that simply answer questions. By giving specialized agents defined responsibilities and the ability to coordinate, enterprises can begin building software that understands a business objective and helps execute the workflow required to achieve it.”

Why single-agent AI isn’t enough for enterprise workflows

The first generation of enterprise generative AI largely focused on the chatbot or copilot model. A user asks a question, the AI retrieves information or generates content, and the user decides what happens next.

That approach remains useful, but complex enterprise environments demand more.

Modern businesses operate across dozens or even hundreds of interconnected applications. A workflow might span CRM platforms, financial systems, project-management tools, internal databases, and communication applications. Each system has its own permissions, data structures, and business rules.

A multi-agent architecture can divide these responsibilities among specialized agents. One agent might retrieve relevant information, another could analyze it, a third could determine what action should be taken, while another validates the proposed action against organizational policies.

“A single general-purpose agent is not always the best architecture for complex enterprise environments,” Swaroop explains. “Specialized agents can be designed around specific functions, permissions, and business rules, making the overall system more controllable, observable, and reliable.”

The objective is not simply to have multiple AI models running simultaneously. It is to create an architecture in which agents have defined responsibilities and controlled ways of interacting with one another and with enterprise systems.

This distinction is important. Without effective orchestration, adding more agents can create more complexity rather than greater efficiency.

For technology leaders, the interesting question is therefore not whether enterprises can connect AI models to their software. It is how they can build an architecture in which autonomous components collaborate reliably enough to participate in real business processes.

Inside the architecture of a multi-agent enterprise system

A well-designed multi-agent system resembles a coordinated organization more than a conventional chatbot.

At its foundation is an orchestration layer responsible for determining which agent should handle a particular task, what information it should receive, and what should happen after its work is completed. Individual agents can then be optimized for narrower functions rather than being expected to solve every problem.

This specialization can have practical advantages.

An agent designed for financial reconciliation, for instance, can operate under different rules from an agent responsible for customer communications. A compliance-focused agent might verify whether a proposed action is permitted before another agent executes it.

The architecture can also incorporate tools and APIs, allowing agents to interact with existing enterprise systems rather than operating in isolation.

For Swaroop, the engineering challenge is making those interactions predictable and observable.

“The intelligence of an agent is only one part of the equation,” she says. “Enterprise systems also need strong orchestration, well-defined interfaces, secure access to data and tools, and observability across the entire workflow. Without those foundations, adding more agents can simply add more complexity.”

However, orchestration introduces its own engineering challenges. Developers must account for communication between agents, conflicting recommendations, incomplete information and situations in which an agent makes an incorrect assumption. The system needs mechanisms for detecting errors and determining when a task should be escalated to a human.

This makes multi-agent AI as much a systems-engineering problem as an AI problem.

The quality of the underlying language model matters, but so do the surrounding architecture, data access, observability, authentication, and workflow design.

Designing guardrails for autonomous AI

Giving an AI system the ability to generate an answer is fundamentally different from giving it the ability to change something in an enterprise environment.

An incorrect summary may be inconvenient. An incorrect payment, customer-account modification, or regulatory decision can have substantially greater consequences.

Consequently, enterprise multi-agent systems require carefully designed guardrails.

Permissions are one important layer. Agents should only be able to access the information and tools necessary for their assigned responsibilities. A customer-support agent, for example, may need access to account information but not unrestricted access to financial systems.

Human oversight is another. Not every action needs manual approval, but organizations can establish thresholds where higher-risk decisions are automatically routed to a person.

Auditability is equally important. Enterprises need to understand what an agent did, which information influenced its decision, and which other agents or systems were involved. Without that visibility, diagnosing failures becomes difficult.

For Swaroop, these safeguards need to be incorporated into the architecture from the outset rather than added after deployment.

“Autonomy in an enterprise environment has to be engineered, not assumed,” says Swaroop, Product Manager at Workday

According to her, agents need clearly defined permissions, auditability and escalation mechanisms so that organizations can benefit from automation while retaining human oversight over decisions where the consequences are significant.

The architecture therefore needs to treat autonomy as a spectrum rather than an on/off switch.

Low-risk, repetitive tasks can potentially be automated end-to-end. More consequential activities can require validation, additional agent review, or explicit human approval.

This approach creates a more realistic path toward enterprise adoption: increase autonomy where the risk is manageable while maintaining control where the consequences are significant.

The infrastructure challenge behind agentic AI

There is another challenge that receives less attention outside technical circles: economics.

Running one AI interaction is relatively straightforward. Running a complex agentic workflow involving multiple agents, repeated model calls, retrieval systems, tool executions, and validation processes can require considerably more computational resources.

At enterprise scale, this creates questions around latency, capacity, and cost.

Should every task use the most powerful available model? Probably not. Some operations may be handled by smaller, faster models, while more complex reasoning tasks are routed to larger systems.

Workload management also becomes important. An enterprise platform may need to handle thousands of simultaneous agentic workflows, each competing for compute resources. Systems must be designed to prevent a sudden increase in activity from creating unacceptable delays or costs.

This is where traditional distributed-systems expertise becomes increasingly relevant to AI architecture.

Load balancing, queue management, caching, observability, and fault tolerance are not new concepts. What is changing is the nature of the workloads they must support.

“Multi-agent systems introduce a different infrastructure challenge because one task can trigger a chain of model calls, tool executions and downstream actions,” Swaroop notes. “The architecture therefore has to account for latency, capacity and cost, while using the appropriate model for each part of the agentic workflow rather than treating every task as equally computationally demanding.”

Multi-agent systems can generate unpredictable chains of activity. One agent’s output may trigger another agent, which calls an external system and produces information that initiates a third task.

Designing infrastructure around these patterns requires organizations to think beyond model performance and consider the economics of the entire agentic workflow.

From systems of record to systems of action

The larger opportunity lies in how these technologies could change the role of enterprise software itself.

Traditional SaaS applications primarily help employees find, enter, organize and analyze information. AI introduces the possibility of software that can interpret that information and participate in the resulting agentic workflow.

Imagine a sales platform that does more than display a pipeline. It could identify stalled opportunities, investigate the reasons behind delays, prepare recommended actions, and coordinate follow-ups across multiple systems.

Or consider an IT environment where an incident triggers a sequence of specialized agents: one investigates the logs, another assesses potential causes, another checks whether a proposed fix complies with operational policies, and another prepares the required change for human approval.

In both cases, the software is moving from recording activity to helping execute it.

That is the fundamental distinction between a system of record and a system of action.

The transition does not mean humans disappear from the agentic workflow. Instead, their role can shift toward supervision, judgment, and exception handling while software manages increasingly complex sequences of routine activity.

For enterprise technology leaders, this could eventually change how SaaS products are designed. Rather than building applications around screens and forms alone, developers may increasingly design around goals, agents, tools, permissions and outcomes.

What enterprise AI architecture looks like next

The next stage of enterprise AI is unlikely to be defined solely by which organization has access to the largest language model.

Increasingly, competitive advantage may come from what companies build around those models.

The difficult engineering questions involve orchestration, reliability, security, data access, infrastructure economics, and the ability to integrate AI into existing business processes without sacrificing control.

Swaroop’s perspective on multi-agent architecture reflects this broader transition. The important development is not simply that enterprises can deploy multiple AI agents. It is that these agents can potentially become components of a new software architecture in which information flows directly into decisions and, under appropriate controls, decisions flow into actions.

That creates a different vision for enterprise SaaS.

The systems of the past were built to remember what happened. The systems emerging now could increasingly help determine what happens next.

The organizations that successfully make that transition will need more than sophisticated AI models. They will need disciplined architecture that balances autonomy with oversight, intelligence with reliability, and innovation with infrastructure economics.

Multi-agent AI may ultimately prove valuable not because it replaces the enterprise software stack, but because it gives that stack a new operating layer, one capable of turning static information into coordinated action.