The conversation around AI agents has moved remarkably fast.
Just two years ago, organizations were experimenting with generative AI through chat interfaces and copilots. Today, many are exploring autonomous and semi-autonomous agents capable of reasoning, making decisions, interacting with systems and executing business processes.
As organizations look beyond that, they’re discovering that deploying enterprise AI requires more than selecting the right large language model (LLM).
While LLMs provide reasoning and language capabilities, enterprise agents require an entirely different architectural layer to operate securely, reliably and at scale.
New research shows why supporting architecture matters. According to the Data and AI Impact Report: The New Economics of Trust, 89% of respondents say their AI agents already play some role in decision-making. Yet only 66% trust agentic AI, compared with 76% who trust generative AI. As AI moves from generating answers to taking action, trust declines while the potential consequences increase.
The next phase of enterprise AI isn’t only about building bigger models. It’s about building the infrastructure that allows intelligent agents to operate securely, reliably and at scale.
From language generation to enterprise execution
Consumer AI applications often operate in relatively simple environments. A user asks a question, the model generates a response and the interaction ends.
Enterprise environments are fundamentally different.
Agents must:
- Access business systems.
- Retrieve organizational knowledge.
- Interact with applications.
- Execute actions.
- Follow policies and regulations.
- Coordinate with other agents.
- Produce auditable outcomes.
This transforms the challenge from language generation into enterprise execution.
It also creates a compounding reliability challenge. Even when an agent performs each step with relatively high accuracy, errors can accumulate across a multistep workflow. An agent operating at 85% accuracy per step may successfully complete the entire workflow only a fraction of the time.
That makes validation, monitoring and appropriate human oversight essential, particularly when an agent is acting on consequential business decisions.
An agent helping a customer write an email is one thing.
An agent approving a loan, triggering a payment, initiating a fraud investigation, updating a patient record, or modifying a supply chain workflow is something entirely different.
The latter requires trust, governance, security and operational controls that go far beyond prompting an LLM.
Why connectivity matters
As organizations deploy more AI agents, connecting them to enterprise data and systems becomes increasingly important.
Rather than building custom integrations for every application, organizations are moving toward standardized approaches that allow agents to securely discover data, access tools and exchange context across systems.
The quality and completeness of that context directly affect whether people are willing to trust the resulting decisions. In the Data and AI Impact Report, 25.6% of respondents said they override AI when its output does not adequately reflect the full context of a situation. Connecting agents to governed enterprise data, metadata and business rules can help reduce those contextual blind spots.
This interoperability enables AI capabilities to expand without requiring organizations to rebuild their technology stack whenever new capabilities emerge.
From custom integrations to common protocols
As those deployed agents connect to more systems, a new challenge emerges: integration complexity.
Historically, every application required custom APIs, custom connectors and custom integrations, creating a growing web of dependencies that is difficult to maintain.
That’s one reason protocols like model context protocol (MCP) are attracting significant attention.
Rather than building custom integrations for every use, MCP introduces a standardized way for AI models and agents to discover and interact with external tools and data sources.
The significance of MCP is not merely technical. It represents a shift toward a more interoperable AI ecosystem where agents can securely access enterprise capabilities without requiring custom engineering for every use case.
Just as APIs transformed software development, protocols like MCP may become foundational infrastructure for the agent economy.
Of course, connectivity is only part of the equation. An agent’s ability to connect to a system does not automatically mean an agent should be allowed to use it.
Security must evolve with AI agents
Traditional cybersecurity was built around users, applications and devices. AI agents introduce a new participant into that environment – one that can retrieve information, interact with systems and, in some cases, take action on its own.
That shift changes how organizations think about security. It’s no longer enough to control who has access to enterprise systems. Organizations also need to determine what AI agents can access, what actions they’re allowed to perform and under what circumstances they can operate autonomously.
Autonomy should therefore be proportional to risk. A low-risk agent retrieving public product information may require minimal intervention. An agent approving credit, moving money or modifying a patient record may require tighter permissions, continuous monitoring and mandatory human approval at designated points in the workflow.
The challenge becomes even greater as agents begin working together. An individual agent may have limited permissions, but multiple agents collaborating across workflows can create unintended access paths and new security risks if they aren’t properly governed.
As enterprise AI matures, security can no longer be treated as a layer added after deployment. Identity, access controls, human intervention points and operational boundaries must be built into the architecture from the beginning.
Governance is moving closer to execution
Many organizations have invested heavily in AI governance programs focused on model development and deployment.
Agentic AI expands that responsibility. Organizations now need to govern not only models, but also the decisions agents make, the actions they take and the workflows they participate in.
Leaders increasingly need answers to questions such as:
- Why did the agent take this action?
- What information influenced its decision?
- Which tools were used?
- What level of autonomy was granted?
- Could a human intervene?
- Can the outcome be reproduced and audited?
These questions become especially important in regulated industries such as banking, health care, insurance, government and life sciences.
These aren’t theoretical concerns. The Data and AI Impact Report found that 79% of organizations override AI in more than 10% of cases. Insufficient explanation was the most commonly cited reason for intervention, named by 34.5% of respondents, followed by failure to reflect the full context of the situation at 25.6%. An outcome can be technically correct yet go unused if people cannot understand how the agent arrived at it.
The governance challenge is shifting from governing models and predictions to governing decisions, actions and entire workflows.
Trust begins with the data foundation
Agents cannot make reliable decisions without reliable context. Enterprise data must be accurate, governed, traceable and accessible in ways that allow organizations to understand how it influenced an agent’s behavior.
Yet the Data and AI Impact Report found that only 17.5% of organizations have reached the highest level of data infrastructure maturity. Many are deploying AI on foundations that cannot fully support the lineage, explainability and validation that agentic systems require.
This matters because an organization cannot reliably reproduce or defend an agent’s decision if it cannot trace the data used, the tools called and the transformations applied throughout the workflow. As agents gain autonomy, data lineage becomes part of operational accountability.
Enterprise AI requires orchestration
One of the biggest misconceptions in the market is the belief that enterprise AI will consist of a few highly capable agents. But the reality is likely to be very different.
Organizations may eventually deploy hundreds or even thousands of specialized agents.
Some agents will retrieve information. Others will analyze data. Others will make recommendations. Others will execute actions.
This creates a new requirement: orchestration.
Without orchestration, agent ecosystems risk becoming fragmented, unpredictable and difficult to govern.
Organizations will need mechanisms to coordinate agents, manage handoffs, resolve conflicts, monitor performance, and maintain accountability across increasingly complex workflows.
The future enterprise will likely resemble an ecosystem of collaborating agents rather than a single super-agent.
The organizations best positioned for this future will not necessarily be those that deploy the most agents first. They will be those who build the data, governance, security, validation and orchestration capabilities needed to deploy them responsibly.
That investment does more than reduce risk. Organizations with highly trustworthy AI practices are 15 times more likely to report strong or high returns from AI than organizations with low trustworthiness. In the agentic era, trust is not simply a safeguard for innovation. It is part of the infrastructure that allows innovation to scale.
