Gartner predicts that by 2028, an average Fortune 500 enterprise could have more than 150,000 AI agents, up from fewer than 15 in 2025. Yet only 13% of organizations say they have the right AI agent governance in place. [1] Most of those agents will be capable. The harder question is whether they can operate inside a real enterprise environment.
An agent may be able to make decisions and complete tasks, but without the right enterprise foundations, it can end up operating outside IT governance, accessing data without proper controls, or making decisions without an audit trail. It may also operate separately from the approval workflows that protect the business from unauthorized actions.
This matters more as organizations move from experimenting with AI agents to putting them into production.
According to McKinsey’s 2026 State of AI survey, 40% of respondents at organizations with more than $1 billion in annual revenue say they are scaling AI agents, up from 27% the previous year. [2]
The question is no longer simply whether an AI agent can perform a task.
It is whether the agent can perform that task within the systems, processes, permissions, and controls that an enterprise already depends on.
1. AI agents can act, but they don’t automatically know what they’re allowed to touch
An AI agent can be given access to data, connected to tools, and instructed to take action. But giving an agent the ability to act does not automatically give it the boundaries needed to operate safely in an enterprise.
Consider an expense approval agent.
It might be able to review an expense claim, check the amount against company policy, and recommend whether it should be approved. But which records should it be able to read? Should it see the payroll or banking details attached to a claim? Can it approve a claim, or only recommend? And who is accountable if it gets one wrong?
These are governance questions.
The agent needs to know what it is allowed to access, what actions it can take, and how those actions are recorded.
The payoff for getting this right appears to be real. IBM’s 2026 Tech Leader Study found that organizations that engineer governance into their AI systems deploy 16 times more AI agents and report 18% higher operating margins. [3]
The problem: AI agents deployed without enterprise foundations
An agent built independently from the systems that govern the business can create a gap between what the agent can do and what it should be allowed to do.
Without enterprise controls, organizations may struggle to enforce:
- Role-based access and permissions
- Authentication and identity
- Approval workflows
- Audit trails
- Data access controls
- Monitoring and incident response
- Clear ownership of agent actions
The more autonomy an agent has, the more important these controls become. [1]
2. AI agents can make decisions, but they don’t automatically know when a human should step in
One of the biggest questions surrounding enterprise AI agents is not whether they can make decisions. It is which decisions they should make on their own.
A customer service agent deciding how to categorize a routine request is very different from an agent approving a large financial transaction. The risk changes with the decision.
For example, an inventory agent could automatically reorder commonly used supplies when stock falls below a certain threshold. That may be a reasonable use of autonomy.
But an agent deciding to make a multimillion-dollar procurement commitment is a different matter.
This is where human-in-the-loop workflows become important.
The human does not necessarily need to review every action. Instead, the system can define where human approval is required based on factors such as transaction value, risk, sensitivity or business rules.
Gartner describes different levels of agent autonomy, ranging from agents that observe information to agents that advise, act with approval, or act autonomously within defined guardrails. [1] It recommends matching governance to each agent’s level of autonomy and trust boundary, rather than applying the same restrictions to every agent. [1]
For higher levels of autonomy, Gartner recommends controls such as continuous monitoring, enforced guardrails, rapid rollback, and clear ownership. [1]
The important point is that human oversight should be designed into the workflow, rather than added as an afterthought.
3. AI agents can use tools, but they don’t automatically fit into your existing business processes
An AI agent may be able to call an API, retrieve information from a database, or interact with an external service.
But connecting to a system is not the same as being integrated into a business process.
Imagine a loan origination agent.
It could collect information from an applicant, retrieve financial data, check eligibility criteria, and recommend whether the application should proceed.
But a real loan process may also require:
- Identity verification
- Document collection
- Credit checks
- Risk assessment
- Approval by the appropriate authority
- Notifications
- Audit records
- Updates to multiple enterprise systems
The agent needs to operate within that process rather than alongside it.
This is one of the challenges highlighted in PwC’s 2025 AI Agent Survey. Respondents cited connecting AI agents across applications and workflows as an obstacle to scaling them. [4]
The survey also found that trust becomes more important for higher-stakes use cases, including financial transactions.
Research from BCG and MIT Sloan Management Review reached a similar conclusion: as agents become more capable of acting independently, organizations need to rethink workflows, governance, roles, and investment approaches. [5]
An enterprise agent should not simply be connected to more systems. It needs to operate within the organization’s processes, permissions, and decision-making structure.
How Joget provides the three missing capabilities
This is where an enterprise application platform can make a practical difference.
With Joget AI Agent Builder, AI agents can be created within the same platform used to build enterprise applications, workflows and integrations. This gives agents access to the enterprise capabilities they need to operate within real business processes.
1. Boundaries and guardrails
An agent can only reach the data and systems its configured tools allow, and guards added to a task can check what goes in and out.
Guards can be added to any agent task to check what goes to the language model and what comes back. Rule-based guards can redact sensitive patterns, an LLM-based guard can apply a safety policy you define, and a webhook guard can hand the decision to your own system. When a check fails, Joget can block the task, sanitize the content, or redirect execution.
This gives organizations control over what an agent can see, what it can do, and what it can pass on.
2. Workflows and human approval
Agents run as a step in Joget workflows, so human approval and escalation can be designed around them.
In Process Builder, an agent is added as a process tool, alongside the human activities, routing rules, and escalations that already make up the workflow.
For example, an agent could review an expense claim, make a recommendation, and route it to the appropriate person when approval is required.
The agent handles the work. The workflow determines where human judgment remains part of the process.
3. Enterprise integrations and auditability
Joget connects agents to the databases, APIs, and systems they need while keeping their actions within the application workflow.
Built-in tools cover database queries, JSON APIs, MCP servers, and email, and enhancers can write results back to forms or workflow variables. In the loan origination example, the agent could gather applicant data and check eligibility as one step, while identity verification, credit checks, and final approval remain their own steps in the process.
Every agent run is recorded in the Agent Execution Audit Trail: who ran it, what ran, and when, with a node-by-node trace of each request and response. Administrators can abort a run that is still in progress, and a usage dashboard tracks token consumption by model, so organizations can see what happened, why, and how many tokens it used.
This is important because enterprise AI is rarely about one agent working in isolation.
It is about connecting AI capabilities to the applications, data, and processes that already run the business.
The path forward: what it takes to scale agents in production
The next stage of enterprise AI adoption will not simply be about creating more agents.
It will be about creating agents that can operate reliably within the organization.
That means thinking about:
1. The level of autonomy
Determine what an agent can observe, recommend, or execute on its own.
2. The boundaries
Define the data, systems, and actions the agent can access.
3. The workflow
Build approval, escalation, and exception handling into the process.
4. The oversight
Monitor agent activity and make it possible to intervene when something goes wrong.
5. The enterprise foundation
Connect agents to the applications, integrations, and business processes they need to actually get work done.
AI agents are becoming increasingly capable. The challenge for enterprises is making that capability usable within the realities of the business.
Organizations moving from experimentation to production will need more than capable AI models. They will need a way to put those agents to work within the workflows, controls, and systems that keep the business running.
That is the role of an enterprise agentic AI application platform.
With Joget, AI agents can be built, connected, and governed within the same environment used to build and run enterprise applications.
The result is a more practical path from AI agent to enterprise application.
Frequently Asked Questions
Can AI agents work without an enterprise application platform?
Yes. AI agents can be built as standalone systems and connected to individual tools and services.
The challenge comes when agents need to operate across multiple enterprise systems while following business rules, permissions, approval processes and compliance requirements.
Why do AI agents need human approval?
Not every agent action requires human approval. The level of oversight should depend on the risk and autonomy involved.
For high-impact actions, human approval can provide an additional control before an agent changes data, makes a financial commitment or takes another consequential action.
What is the difference between an AI agent and an enterprise AI agent?
An AI agent can perform tasks autonomously. An enterprise AI agent also needs to operate within organizational rules and systems.
That includes permissions, authentication, workflows, integrations, auditability, monitoring and human oversight where appropriate.
How does Joget support AI agents?
Joget AI Agent Builder allows organizations to create AI agents within the same platform used to build applications, workflows, and integrations.
This allows agents to work with existing enterprise processes and capabilities such as runtime guardrails, workflow approvals, APIs, databases, and execution audit trails.
What should organizations consider before deploying AI agents?
Organizations should consider the agent’s level of autonomy, the data and systems it can access, the actions it can take, where human approval is required, how its activity will be monitored, and how it fits into existing business processes.
Technology is only one part of the equation. The surrounding application and governance foundation matters just as much.
Sources
1. Gartner
1a. Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure, May 26, 2026.
https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure
1b. Managing AI Agent Sprawl, April 28, 2026
https://www.gartner.com/en/newsroom/press-releases/2026-04-28-gartner-identifies-six-steps-to-manage-artificial-intelligence-agent-sprawl
2. McKinsey & Company, “The state of AI in 2026: On the road to ROI,” August 25, 2026.
https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
3. IBM, 2026 Tech Leader Study: Building the IT Foundation for Agentic AI at Scale
https://www.ibm.com/downloads/documents/us-en/16ddce7b4954875b
4. PwC, “AI Agent Survey,” 2025.
https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html
5. BCG + MIT Sloan Management Review
Leading in the Age of AI Agents — 2025 research specifically focused on agentic AI, workflows, and governance.
https://www.bcg.com/publications/2025/machines-that-manage-themselves
Related
Last modified: November 4, 2026