Beyond the Pilot: How Enterprises Build AI They Can Actually Trust

AI is moving beyond experimentation. The next challenge for enterprise leaders is no longer proving that AI can generate an answer. It is determining whether an AI agent can make a decision and take action that the organization can trust.

Constellation analyst Mike Ni speaks with Waqas Ahmed, VP of AI Development at OpenText, about what it takes to operationalize agentic AI. The conversation moves beyond the familiar discussion of AI-ready data to explore the architecture, governance, context, evaluation, and provenance required for AI to operate across the enterprise.

AI readiness is more than data preparation

Enterprises are evolving in how they use AI. It starts with AI embedded in existing applications and chatbots. From there, organizations are introducing discrete agents that users can invoke for specific functions. The next step is for agents to work across application boundaries and interact with other agents. Eventually, applications themselves could become collections of agents working together as an agentic system.

The challenge is that the models themselves are not necessarily what is holding these initiatives back. The bigger gaps are context, grounding, trust, and measurement.

As Ahmed explains, AI solutions can produce confident but inaccurate answers. Organizations may also hesitate to give an agent access to sensitive information or allow it to act on behalf of employees. Even when an AI system performs well initially, small changes to prompts, context, or models can cause performance to drift.

That creates a fundamental problem for enterprises: How do you know an AI system will continue to behave reliably once it is operating in production?

Agents need more than access to data

One of the biggest misconceptions about agentic AI is that giving an agent access to enterprise data is enough. It isn't.

Agents operate on processes, decisions, and actions. They need to understand not only what information exists, but how that information relates to a business process and what decisions have already been made.

That means building a contextual history around business activity.

Ahmed describes this as bringing together the system of record, the decisions being made, and the execution of processes across their lifecycle. That information can then be captured in a live, governed context layer that gives the right agent the right information at the right time.

Security is equally important. An agent with unrestricted access to enterprise information can create risk at scale. Context therefore needs to include permissions, policies, and constraints so agents operate within clearly defined boundaries.

Four foundations for trusted agentic AI

Ahmed outlines four capabilities that enterprises need to establish around their AI solutions.

1. Continuous evaluation

OpenText calls this EvalOps, drawing on DevOps principles of DevOps.

The idea is straightforward: Enterprises need a consistent way to establish baselines, evaluate AI outcomes, and continue measuring performance in production.

That becomes especially important because AI is non-deterministic. The same question can produce different answers, and seemingly small changes can affect performance. Continuous evaluation helps organizations a way to identify drift rather than discovering it after an AI system has already lost credibility.

2. Identity and security

Agents need identities and clearly defined boundaries. As organizations deploy multiple agents, they need to know which agents are operating, what those agents are allowed to access, and what actions they are permitted to take. That means extending traditional identity and security practices into an agentic environment.

3. Governed context

AI needs access to information that is both relevant and trustworthy. A governed context layer helps ensure agents receive the information they need while preserving the appropriate permissions and controls. The objective is not simply to provide more information. It is to provide the right information within the right context.

4. Cross-application interaction

Enterprise processes rarely exist within a single application. Agents need to be able to interact with other agents and systems across the enterprise if they are going to support end-to-end processes rather than simply automate isolated tasks.

That is where agentic AI begins to move from application-level assistance toward enterprise-level automation.

From assistance to action

The distinction between AI assistance and AI automation is becoming increasingly important. An AI assistant might summarize information or recommend a next step. An AI agent can potentially execute the next step.

That shift changes the requirements for trust. Once an AI system can act, organizations need to understand not only whether its answer is accurate, but why it made a decision, what information it used, what action it took, and whether that action was within its authority.

This is where runtime trust becomes critical. Traditional governance often focuses on establishing rules before a system is deployed. Agentic AI requires those controls to remain active as the system operates.

Enterprises need visibility into agentic identities, executions, outcomes, audit trails, and provenance.

A practical example: AI-powered incident management

The conversation brings these ideas to life through an incident management example.

In a typical operational environment, teams may receive hundreds or thousands of alerts. Humans have to correlate those signals, understand recent system changes, review previous incidents, and determine the likely root cause.

AI can help because the problem requires connecting multiple pieces of context rather than following a simple deterministic rule.

An agent can examine recent changes, related incidents, and previous resolutions to build a more complete picture of what is happening. From there, it can identify a likely root cause and recommend actions.

OpenText has applied this approach internally, resulting in a significant reduction in the time required to identify and resolve incidents.

This is an important distinction for enterprise AI. The value doesn't come simply from generating an answer. It comes from combining context to support a business decision or action.

How agents earn autonomy

A useful analogy from the conversation is to think of an AI agent like a new employee. On day one, that employee doesn't receive unrestricted access to every system and every business process.

They receive defined responsibilities, access to the information they need, corporate policies, and clear procedures. Their work is monitored closely. As they demonstrate reliable performance, they earn greater responsibility. Agentic AI can follow a similar progression.

Organizations can start with tightly defined guardrails and approval points. They can measure outcomes, evaluate performance, and gradually expand what an agent is allowed to do as confidence increases.

That creates a reinforcement cycle:

Guardrails → measurement → evaluation → greater autonomy

The foundation underneath that cycle is governed by access, continuous measurement, and evaluation.

From productivity to process transformation

Once organizations establish enough trust to let AI act, the conversation changes.

Instead of asking:

"Can AI help someone do their job?"

Leaders can start asking:

"What work should AI actually own?"

That's a much more consequential question. It shifts AI from a productivity tool toward a mechanism for redesigning business processes.

Waqas cites an example involving a large organization with hundreds of thousands of employee records. A previous HR process required teams to work through those records to generate compensation letters.

Using OpenText's Content Aviator, the organization was able to apply contextual decision-making across those records and generate the compensation letters in hours instead of the weeks the process previously required.

The example illustrates the larger opportunity. AI can create meaningful value when it is connected to trusted enterprise information, context, and the processes where work actually happens.

The capability enterprise AI platforms cannot afford to overlook

As organizations move from pilots to production, one capability stands out: grounding and provenance. Enterprises need to be able to prove that their agents operate consistently. A successful demo isn't enough.

A model update, prompt change, or system change can introduce drift. Without continuous evaluation, organizations may not know that performance has degraded until users lose confidence in the system.

Grounding connects AI decisions to the information that supports them. Provenance provides visibility into how those decisions were made. Together, they create the foundation for operational trust.

The next phase of enterprise AI

Enterprise AI is entering a phase where intelligence alone isn't enough. The organizations that move successfully from experimentation to production will need to build the infrastructure around AI that makes action trustworthy.

That means:

• Trusted information.
• Governed context.
• Agentic identity and security.
• Continuous evaluation.
• Provenance.
• Human oversight where it matters.

The question for CIOs and IT leaders is no longer simply whether AI works. It's whether the organization has built the foundation that allows AI to act reliably, securely, and at scale.